Add h8l and rocm to release workflow (#14648)

* Add h8l to release workflow

* Add rocm to release workflow

* Variants

.cspell/frigate-dictionary.txt

@@ -42,6 +42,7 @@ codeproject
 colormap
 colorspace
 comms
+coro
 ctypeslib
 CUDA
 Cuvid
@@ -59,6 +60,7 @@ dsize
 dtype
 ECONNRESET
 edgetpu
+fastapi
 faststart
 fflags
 ffprobe
@@ -212,6 +214,7 @@ rcond
 RDONLY
 rebranded
 referer
+reindex
 Reolink
 restream
 restreamed
@@ -236,6 +239,7 @@ sleeptime
 SNDMORE
 socs
 sqliteq
+sqlitevecq
 ssdlite
 statm
 stimeout
@@ -270,6 +274,7 @@ unraid
 unreviewed
 userdata
 usermod
+uvicorn
 vaapi
 vainfo
 variations

.github/DISCUSSION_TEMPLATE/detector-support.yml

@@ -74,19 +74,6 @@ body:
         - CPU (no coral)
     validations:
       required: true
-  - type: dropdown
-    id: object-detector
-    attributes:
-      label: Object Detector
-      options:
-        - Coral
-        - OpenVino
-        - TensorRT
-        - RKNN
-        - Other
-        - CPU (no coral)
-    validations:
-      required: true
   - type: textarea
     id: screenshots
     attributes:

.github/DISCUSSION_TEMPLATE/general-support.yml

@@ -102,19 +102,6 @@ body:
         - CPU (no coral)
     validations:
       required: true
-  - type: dropdown
-    id: object-detector
-    attributes:
-      label: Object Detector
-      options:
-        - Coral
-        - OpenVino
-        - TensorRT
-        - RKNN
-        - Other
-        - CPU (no coral)
-    validations:
-      required: true
   - type: dropdown
     id: network
     attributes:

.github/workflows/ci.yml

@@ -6,6 +6,8 @@ on:
     branches:
       - dev
       - master
+    paths-ignore:
+      - 'docs/**'      
 
 # only run the latest commit to avoid cache overwrites
 concurrency:
@@ -155,6 +157,28 @@ jobs:
             tensorrt.tags=${{ steps.setup.outputs.image-name }}-tensorrt
             *.cache-from=type=registry,ref=${{ steps.setup.outputs.cache-name }}-amd64
             *.cache-to=type=registry,ref=${{ steps.setup.outputs.cache-name }}-amd64,mode=max
+  arm64_extra_builds:
+    runs-on: ubuntu-latest
+    name: ARM Extra Build
+    needs:
+      - arm64_build
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Set up QEMU and Buildx
+        id: setup
+        uses: ./.github/actions/setup
+        with:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      - name: Build and push Rockchip build
+        uses: docker/bake-action@v3
+        with:
+          push: true
+          targets: rk
+          files: docker/rockchip/rk.hcl
+          set: |
+            rk.tags=${{ steps.setup.outputs.image-name }}-rk
+            *.cache-from=type=gha
   combined_extra_builds:
     runs-on: ubuntu-latest
     name: Combined Extra Builds

.github/workflows/pull_request.yml

@@ -1,6 +1,9 @@
 name: On pull request
 
-on: pull_request
+on:
+  pull_request:
+    paths-ignore:
+      - 'docs/**'
 
 env:
   DEFAULT_PYTHON: 3.9

.github/workflows/release.yml

@@ -34,14 +34,14 @@ jobs:
           STABLE_TAG=${BASE}:stable
           PULL_TAG=${BASE}:${BUILD_TAG}
           docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG} docker://${VERSION_TAG}
-          for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk; do
+          for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk h8l rocm; do
             docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG}-${variant} docker://${VERSION_TAG}-${variant}
           done
 
           # stable tag
           if [[ "${BUILD_TYPE}" == "stable" ]]; then
             docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG} docker://${STABLE_TAG}
-            for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk; do
+            for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk h8l rocm; do
               docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG}-${variant} docker://${STABLE_TAG}-${variant}
             done
           fi

docker/hailo8l/user_installation.sh

@@ -38,7 +38,7 @@ cd ../../
 if [ ! -d /lib/firmware/hailo ]; then
   sudo mkdir /lib/firmware/hailo
 fi
-sudo mv hailo8_fw.4.17.0.bin /lib/firmware/hailo/hailo8_fw.bin
+sudo mv hailo8_fw.4.18.0.bin /lib/firmware/hailo/hailo8_fw.bin
 
 # Install udev rules
 sudo cp ./linux/pcie/51-hailo-udev.rules /etc/udev/rules.d/

docker/main/Dockerfile

@@ -180,9 +180,6 @@ RUN /build_pysqlite3.sh
 COPY docker/main/requirements-wheels.txt /requirements-wheels.txt
 RUN pip3 wheel --wheel-dir=/wheels -r /requirements-wheels.txt
 
-COPY docker/main/requirements-wheels-post.txt /requirements-wheels-post.txt
-RUN pip3 wheel --no-deps --wheel-dir=/wheels-post -r /requirements-wheels-post.txt
-
 
 # Collect deps in a single layer
 FROM scratch AS deps-rootfs
@@ -225,14 +222,6 @@ RUN --mount=type=bind,from=wheels,source=/wheels,target=/deps/wheels \
     python3 -m pip install --upgrade pip && \
     pip3 install -U /deps/wheels/*.whl
 
-# We have to uninstall this dependency specifically
-# as it will break onnxruntime-openvino
-RUN pip3 uninstall -y onnxruntime
-
-RUN --mount=type=bind,from=wheels,source=/wheels-post,target=/deps/wheels \
-    python3 -m pip install --upgrade pip && \
-    pip3 install -U /deps/wheels/*.whl
-
 COPY --from=deps-rootfs / /
 
 RUN ldconfig

docker/main/install_deps.sh

@@ -8,6 +8,7 @@ apt-get -qq install --no-install-recommends -y \
     apt-transport-https \
     gnupg \
     wget \
+    lbzip2 \
     procps vainfo \
     unzip locales tzdata libxml2 xz-utils \
     python3.9 \
@@ -45,7 +46,7 @@ if [[ "${TARGETARCH}" == "amd64" ]]; then
     wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2022-07-31-12-37/ffmpeg-n5.1-2-g915ef932a3-linux64-gpl-5.1.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/5.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/5.0/doc /usr/lib/ffmpeg/5.0/bin/ffplay
-    wget -qO btbn-ffmpeg.tar.xz "https://github.com/BtbN/FFmpeg-Builds/releases/download/autobuild-2024-09-30-15-36/ffmpeg-n7.1-linux64-gpl-7.1.tar.xz"
+    wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2024-09-19-12-51/ffmpeg-n7.0.2-18-g3e6cec1286-linux64-gpl-7.0.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/7.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/7.0/doc /usr/lib/ffmpeg/7.0/bin/ffplay
 fi
@@ -57,7 +58,7 @@ if [[ "${TARGETARCH}" == "arm64" ]]; then
     wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2022-07-31-12-37/ffmpeg-n5.1-2-g915ef932a3-linuxarm64-gpl-5.1.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/5.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/5.0/doc /usr/lib/ffmpeg/5.0/bin/ffplay
-    wget -qO btbn-ffmpeg.tar.xz "https://github.com/BtbN/FFmpeg-Builds/releases/download/autobuild-2024-09-30-15-36/ffmpeg-n7.1-linuxarm64-gpl-7.1.tar.xz"
+    wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2024-09-19-12-51/ffmpeg-n7.0.2-18-g3e6cec1286-linuxarm64-gpl-7.0.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/7.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/7.0/doc /usr/lib/ffmpeg/7.0/bin/ffplay
 fi
@@ -76,6 +77,9 @@ if [[ "${TARGETARCH}" == "amd64" ]]; then
     apt-get -qq install --no-install-recommends --no-install-suggests -y \
         i965-va-driver-shaders
 
+    # intel packages use zst compression so we need to update dpkg
+    apt-get install -y dpkg
+
     rm -f /etc/apt/sources.list.d/debian-bookworm.list
 
     # use intel apt intel packages

docker/main/requirements-wheels-post.txt

@@ -1,3 +0,0 @@
-# ONNX
-onnxruntime-openvino == 1.19.* ; platform_machine == 'x86_64'
-onnxruntime == 1.19.* ; platform_machine == 'aarch64'

docker/main/requirements-wheels.txt

@@ -1,13 +1,13 @@
 click == 8.1.*
 # FastAPI
 starlette-context == 0.3.6
-fastapi == 0.115.0
+fastapi == 0.115.*
 uvicorn == 0.30.*
-slowapi == 0.1.9
+slowapi == 0.1.*
 imutils == 0.5.*
 joserfc == 1.0.*
 pathvalidate == 3.2.*
-markupsafe == 3.0.*
+markupsafe == 2.1.*
 mypy == 1.6.1
 numpy == 1.26.*
 onvif_zeep == 0.2.12
@@ -16,10 +16,10 @@ paho-mqtt == 2.1.*
 pandas == 2.2.*
 peewee == 3.17.*
 peewee_migrate == 1.13.*
-psutil == 5.9.*
+psutil == 6.1.*
 pydantic == 2.8.*
 git+https://github.com/fbcotter/py3nvml#egg=py3nvml
-pytz == 2024.1
+pytz == 2024.*
 pyzmq == 26.2.*
 ruamel.yaml == 0.18.*
 tzlocal == 5.2
@@ -30,11 +30,12 @@ norfair == 2.2.*
 setproctitle == 1.3.*
 ws4py == 0.5.*
 unidecode == 1.3.*
-# OpenVino (ONNX installed in wheels-post)
+# OpenVino & ONNX
 openvino == 2024.3.*
+onnxruntime-openvino == 1.19.* ; platform_machine == 'x86_64'
+onnxruntime == 1.19.* ; platform_machine == 'aarch64'
 # Embeddings
 transformers == 4.45.*
-onnx_clip == 4.0.*
 # Generative AI
 google-generativeai == 0.8.*
 ollama == 0.3.*

docker/tensorrt/requirements-amd64.txt

@@ -9,6 +9,6 @@ nvidia-cuda-runtime-cu11 == 11.8.*; platform_machine == 'x86_64'
 nvidia-cublas-cu11 == 11.11.3.6; platform_machine == 'x86_64'
 nvidia-cudnn-cu11 == 8.6.0.*; platform_machine == 'x86_64'
 nvidia-cufft-cu11==10.*; platform_machine == 'x86_64'
-onnx==1.14.0; platform_machine == 'x86_64'
-onnxruntime-gpu==1.17.*; platform_machine == 'x86_64'
+onnx==1.16.*; platform_machine == 'x86_64'
+onnxruntime-gpu==1.18.*; platform_machine == 'x86_64'
 protobuf==3.20.3; platform_machine == 'x86_64'

docs/docs/configuration/genai.md

@@ -3,7 +3,7 @@ id: genai
 title: Generative AI
 ---
 
-Generative AI can be used to automatically generate descriptions based on the thumbnails of your tracked objects. This helps with [Semantic Search](/configuration/semantic_search) in Frigate by providing detailed text descriptions as a basis of the search query.
+Generative AI can be used to automatically generate descriptive text based on the thumbnails of your tracked objects. This helps with [Semantic Search](/configuration/semantic_search) in Frigate to provide more context about your tracked objects.
 
 Semantic Search must be enabled to use Generative AI. Descriptions are accessed via the _Explore_ view in the Frigate UI by clicking on a tracked object's thumbnail.
 
@@ -29,11 +29,21 @@ cameras:
 
 ## Ollama
 
-[Ollama](https://ollama.com/) allows you to self-host large language models and keep everything running locally. It provides a nice API over [llama.cpp](https://github.com/ggerganov/llama.cpp). It is highly recommended to host this server on a machine with an Nvidia graphics card, or on a Apple silicon Mac for best performance. Most of the 7b parameter 4-bit vision models will fit inside 8GB of VRAM. There is also a [docker container](https://hub.docker.com/r/ollama/ollama) available.
+:::warning
+
+Using Ollama on CPU is not recommended, high inference times make using generative AI impractical.
+
+:::
+
+[Ollama](https://ollama.com/) allows you to self-host large language models and keep everything running locally. It provides a nice API over [llama.cpp](https://github.com/ggerganov/llama.cpp). It is highly recommended to host this server on a machine with an Nvidia graphics card, or on a Apple silicon Mac for best performance.
+
+Most of the 7b parameter 4-bit vision models will fit inside 8GB of VRAM. There is also a [docker container](https://hub.docker.com/r/ollama/ollama) available.
+
+Parallel requests also come with some caveats. See the [Ollama documentation](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-does-ollama-handle-concurrent-requests).
 
 ### Supported Models
 
-You must use a vision capable model with Frigate. Current model variants can be found [in their model library](https://ollama.com/library). At the time of writing, this includes `llava`, `llava-llama3`, `llava-phi3`, and `moondream`.
+You must use a vision capable model with Frigate. Current model variants can be found [in their model library](https://ollama.com/library). At the time of writing, this includes `llava`, `llava-llama3`, `llava-phi3`, and `moondream`. Note that Frigate will not automatically download the model you specify in your config, you must download the model to your local instance of Ollama first i.e. by running `ollama pull llava:7b` on your Ollama server/Docker container. Note that the model specified in Frigate's config must match the downloaded model tag.
 
 :::note
 
@@ -48,7 +58,7 @@ genai:
   enabled: True
   provider: ollama
   base_url: http://localhost:11434
-  model: llava
+  model: llava:7b
 ```
 
 ## Google Gemini
@@ -122,12 +132,18 @@ genai:
   api_key: "{FRIGATE_OPENAI_API_KEY}"
 ```
 
+## Usage and Best Practices
+
+Frigate's thumbnail search excels at identifying specific details about tracked objects – for example, using an "image caption" approach to find a "person wearing a yellow vest," "a white dog running across the lawn," or "a red car on a residential street." To enhance this further, Frigate’s default prompts are designed to ask your AI provider about the intent behind the object's actions, rather than just describing its appearance.
+
+While generating simple descriptions of detected objects is useful, understanding intent provides a deeper layer of insight. Instead of just recognizing "what" is in a scene, Frigate’s default prompts aim to infer "why" it might be there or "what" it could do next. Descriptions tell you what’s happening, but intent gives context. For instance, a person walking toward a door might seem like a visitor, but if they’re moving quickly after hours, you can infer a potential break-in attempt. Detecting a person loitering near a door at night can trigger an alert sooner than simply noting "a person standing by the door," helping you respond based on the situation’s context.
+
 ## Custom Prompts
 
 Frigate sends multiple frames from the tracked object along with a prompt to your Generative AI provider asking it to generate a description. The default prompt is as follows:
 
 ```
-Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.
+Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.
 ```
 
 :::tip
@@ -144,10 +160,10 @@ genai:
   provider: ollama
   base_url: http://localhost:11434
   model: llava
-  prompt: "Describe the {label} in these images from the {camera} security camera."
+  prompt: "Analyze the {label} in these images from the {camera} security camera. Focus on the actions, behavior, and potential intent of the {label}, rather than just describing its appearance."
   object_prompts:
-    person: "Describe the main person in these images (gender, age, clothing, activity, etc). Do not include where the activity is occurring (sidewalk, concrete, driveway, etc)."
-    car: "Label the primary vehicle in these images with just the name of the company if it is a delivery vehicle, or the color make and model."
+    person: "Examine the main person in these images. What are they doing and what might their actions suggest about their intent (e.g., approaching a door, leaving an area, standing still)? Do not describe the surroundings or static details."
+    car: "Observe the primary vehicle in these images. Focus on its movement, direction, or purpose (e.g., parking, approaching, circling). If it's a delivery vehicle, mention the company."
 ```
 
 Prompts can also be overriden at the camera level to provide a more detailed prompt to the model about your specific camera, if you desire. By default, descriptions will be generated for all tracked objects and all zones. But you can also optionally specify `objects` and `required_zones` to only generate descriptions for certain tracked objects or zones.
@@ -159,10 +175,10 @@ cameras:
   front_door:
     genai:
       use_snapshot: True
-      prompt: "Describe the {label} in these images from the {camera} security camera at the front door of a house, aimed outward toward the street."
+      prompt: "Analyze the {label} in these images from the {camera} security camera at the front door. Focus on the actions and potential intent of the {label}."
       object_prompts:
-        person: "Describe the main person in these images (gender, age, clothing, activity, etc). Do not include where the activity is occurring (sidewalk, concrete, driveway, etc). If delivering a package, include the company the package is from."
-        cat: "Describe the cat in these images (color, size, tail). Indicate whether or not the cat is by the flower pots. If the cat is chasing a mouse, make up a name for the mouse."
+        person: "Examine the person in these images. What are they doing, and how might their actions suggest their purpose (e.g., delivering something, approaching, leaving)? If they are carrying or interacting with a package, include details about its source or destination."
+        cat: "Observe the cat in these images. Focus on its movement and intent (e.g., wandering, hunting, interacting with objects). If the cat is near the flower pots or engaging in any specific actions, mention it."
       objects:
         - person
         - cat

docs/docs/configuration/motion_detection.md

@@ -92,10 +92,16 @@ motion:
   lightning_threshold: 0.8
 ```
 
-:::tip
+:::warning
 
 Some cameras like doorbell cameras may have missed detections when someone walks directly in front of the camera and the lightning_threshold causes motion detection to be re-calibrated. In this case, it may be desirable to increase the `lightning_threshold` to ensure these objects are not missed.
 
 :::
 
+:::note
+
+Lightning threshold does not stop motion based recordings from being saved.
+
+:::
+
 Large changes in motion like PTZ moves and camera switches between Color and IR mode should result in no motion detection. This is done via the `lightning_threshold` configuration. It is defined as the percentage of the image used to detect lightning or other substantial changes where motion detection needs to recalibrate. Increasing this value will make motion detection more likely to consider lightning or IR mode changes as valid motion. Decreasing this value will make motion detection more likely to ignore large amounts of motion such as a person approaching a doorbell camera.

docs/docs/configuration/reference.md

@@ -518,6 +518,9 @@ semantic_search:
   enabled: False
   # Optional: Re-index embeddings database from historical tracked objects (default: shown below)
   reindex: False
+  # Optional: Set the model size used for embeddings. (default: shown below)
+  # NOTE: small model runs on CPU and large model runs on GPU
+  model_size: "small"
 
 # Optional: Configuration for AI generated tracked object descriptions
 # NOTE: Semantic Search must be enabled for this to do anything.

docs/docs/configuration/semantic_search.md

@@ -5,10 +5,18 @@ title: Using Semantic Search
 
 Semantic Search in Frigate allows you to find tracked objects within your review items using either the image itself, a user-defined text description, or an automatically generated one. This feature works by creating _embeddings_ — numerical vector representations — for both the images and text descriptions of your tracked objects. By comparing these embeddings, Frigate assesses their similarities to deliver relevant search results.
 
-Frigate has support for two models to create embeddings, both of which run locally: [OpenAI CLIP](https://openai.com/research/clip) and [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2). Embeddings are then saved to Frigate's database.
+Frigate has support for [Jina AI's CLIP model](https://huggingface.co/jinaai/jina-clip-v1) to create embeddings, which runs locally. Embeddings are then saved to Frigate's database.
 
 Semantic Search is accessed via the _Explore_ view in the Frigate UI.
 
+## Minimum System Requirements
+
+Semantic Search works by running a large AI model locally on your system. Small or underpowered systems like a Raspberry Pi will not run Semantic Search reliably or at all.
+
+A minimum of 8GB of RAM is required to use Semantic Search. A GPU is not strictly required but will provide a significant performance increase over CPU-only systems.
+
+For best performance, 16GB or more of RAM and a dedicated GPU are recommended.
+
 ## Configuration
 
 Semantic search is disabled by default, and must be enabled in your config file before it can be used. Semantic Search is a global configuration setting.
@@ -27,18 +35,34 @@ If you are enabling the Search feature for the first time, be advised that Friga
 
 :::
 
-### OpenAI CLIP
+### Jina AI CLIP
 
-This model is able to embed both images and text into the same vector space, which allows `image -> image` and `text -> image` similarity searches. Frigate uses this model on tracked objects to encode the thumbnail image and store it in the database. When searching for tracked objects via text in the search box, Frigate will perform a `text -> image` similarity search against this embedding. When clicking "Find Similar" in the tracked object detail pane, Frigate will perform an `image -> image` similarity search to retrieve the closest matching thumbnails.
+The vision model is able to embed both images and text into the same vector space, which allows `image -> image` and `text -> image` similarity searches. Frigate uses this model on tracked objects to encode the thumbnail image and store it in the database. When searching for tracked objects via text in the search box, Frigate will perform a `text -> image` similarity search against this embedding. When clicking "Find Similar" in the tracked object detail pane, Frigate will perform an `image -> image` similarity search to retrieve the closest matching thumbnails.
 
-### all-MiniLM-L6-v2
+The text model is used to embed tracked object descriptions and perform searches against them. Descriptions can be created, viewed, and modified on the Search page when clicking on the gray tracked object chip at the top left of each review item. See [the Generative AI docs](/configuration/genai.md) for more information on how to automatically generate tracked object descriptions.
 
-This is a sentence embedding model that has been fine tuned on over 1 billion sentence pairs. This model is used to embed tracked object descriptions and perform searches against them. Descriptions can be created, viewed, and modified on the Search page when clicking on the gray tracked object chip at the top left of each review item. See [the Generative AI docs](/configuration/genai.md) for more information on how to automatically generate tracked object descriptions.
+Differently weighted CLIP models are available and can be selected by setting the `model_size` config option:
 
-## Usage
+:::tip
+
+The CLIP models are downloaded in ONNX format, which means they will be accelerated using GPU hardware when available. This depends on the Docker build that is used. See [the object detector docs](../configuration/object_detectors.md) for more information.
+
+:::
+
+```yaml
+semantic_search:
+  enabled: True
+  model_size: small
+```
+
+- Configuring the `large` model employs the full Jina model and will automatically run on the GPU if applicable.
+- Configuring the `small` model employs a quantized version of the model that uses much less RAM and runs faster on CPU with a very negligible difference in embedding quality.
+
+## Usage and Best Practices
 
 1. Semantic search is used in conjunction with the other filters available on the Search page. Use a combination of traditional filtering and semantic search for the best results.
-2. The comparison between text and image embedding distances generally means that results matching `description` will appear first, even if a `thumbnail` embedding may be a better match. Play with the "Search Type" filter to help find what you are looking for.
-3. Make your search language and tone closely match your descriptions. If you are using thumbnail search, phrase your query as an image caption.
-4. Semantic search on thumbnails tends to return better results when matching large subjects that take up most of the frame. Small things like "cat" tend to not work well.
-5. Experiment! Find a tracked object you want to test and start typing keywords to see what works for you.
+2. Use the thumbnail search type when searching for particular objects in the scene. Use the description search type when attempting to discern the intent of your object.
+3. Because of how the AI models Frigate uses have been trained, the comparison between text and image embedding distances generally means that with multi-modal (`thumbnail` and `description`) searches, results matching `description` will appear first, even if a `thumbnail` embedding may be a better match. Play with the "Search Type" setting to help find what you are looking for. Note that if you are generating descriptions for specific objects or zones only, this may cause search results to prioritize the objects with descriptions even if the the ones without them are more relevant.
+4. Make your search language and tone closely match exactly what you're looking for. If you are using thumbnail search, **phrase your query as an image caption**. Searching for "red car" may not work as well as "red sedan driving down a residential street on a sunny day".
+5. Semantic search on thumbnails tends to return better results when matching large subjects that take up most of the frame. Small things like "cat" tend to not work well.
+6. Experiment! Find a tracked object you want to test and start typing keywords and phrases to see what works for you.

docs/docs/frigate/installation.md

@@ -250,10 +250,7 @@ The community supported docker image tags for the current stable version are:
 - `stable-tensorrt-jp5` - Frigate build optimized for nvidia Jetson devices running Jetpack 5
 - `stable-tensorrt-jp4` - Frigate build optimized for nvidia Jetson devices running Jetpack 4.6
 - `stable-rk` - Frigate build for SBCs with Rockchip SoC
-- `stable-rocm` - Frigate build for [AMD GPUs and iGPUs](../configuration/object_detectors.md#amdrocm-gpu-detector), all drivers
-  - `stable-rocm-gfx900` - AMD gfx900 driver only
-  - `stable-rocm-gfx1030` - AMD gfx1030 driver only
-  - `stable-rocm-gfx1100` - AMD gfx1100 driver only
+- `stable-rocm` - Frigate build for [AMD GPUs](../configuration/object_detectors.md#amdrocm-gpu-detector)
   - `stable-h8l` - Frigate build for the Hailo-8L M.2 PICe Raspberry Pi 5 hat
 
 ## Home Assistant Addon

docs/static/frigate-api.yaml

@@ -172,76 +172,65 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: labels
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Labels
         - name: zones
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Zones
         - name: reviewed
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             default: 0
             title: Reviewed
         - name: limit
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             title: Limit
         - name: severity
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            allOf:
+              - $ref: '#/components/schemas/SeverityEnum'
             title: Severity
         - name: before
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: Before
         - name: after
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: After
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                type: array
+                items:
+                  $ref: '#/components/schemas/ReviewSegmentResponse'
+                title: Response Review Review Get
         '422':
           description: Validation Error
           content:
@@ -259,36 +248,28 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: labels
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Labels
         - name: zones
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Zones
         - name: timezone
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: utc
             title: Timezone
       responses:
@@ -296,7 +277,8 @@ paths:
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSummaryResponse'
         '422':
           description: Validation Error
           content:
@@ -310,17 +292,18 @@ paths:
       summary: Set Multiple Reviewed
       operationId: set_multiple_reviewed_reviews_viewed_post
       requestBody:
+        required: true
         content:
           application/json:
             schema:
-              type: object
-              title: Body
+              $ref: '#/components/schemas/ReviewSetMultipleReviewedBody'
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -334,17 +317,18 @@ paths:
       summary: Delete Reviews
       operationId: delete_reviews_reviews_delete_post
       requestBody:
+        required: true
         content:
           application/json:
             schema:
-              type: object
-              title: Body
+              $ref: '#/components/schemas/ReviewDeleteMultipleReviewsBody'
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -363,96 +347,38 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: before
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: Before
         - name: after
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: After
         - name: scale
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             default: 30
             title: Scale
       responses:
         '200':
           description: Successful Response
-          content:
-            application/json:
-              schema: { }
-        '422':
-          description: Validation Error
           content:
             application/json:
               schema:
-                $ref: '#/components/schemas/HTTPValidationError'
-  /review/activity/audio:
-    get:
-      tags:
-        - Review
-      summary: Audio Activity
-      description: Get motion and audio activity.
-      operationId: audio_activity_review_activity_audio_get
-      parameters:
-        - name: cameras
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: string
-              - type: 'null'
-            default: all
-            title: Cameras
-        - name: before
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: number
-              - type: 'null'
-            title: Before
-        - name: after
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: number
-              - type: 'null'
-            title: After
-        - name: scale
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
-            default: 30
-            title: Scale
-      responses:
-        '200':
-          description: Successful Response
-          content:
-            application/json:
-              schema: { }
+                type: array
+                items:
+                  $ref: '#/components/schemas/ReviewActivityMotionResponse'
+                title: Response Motion Activity Review Activity Motion Get
         '422':
           description: Validation Error
           content:
@@ -477,57 +403,60 @@ paths:
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSegmentResponse'
         '422':
           description: Validation Error
           content:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  /review/{event_id}:
+  /review/{review_id}:
     get:
       tags:
         - Review
       summary: Get Review
-      operationId: get_review_review__event_id__get
+      operationId: get_review_review__review_id__get
       parameters:
-        - name: event_id
+        - name: review_id
           in: path
           required: true
           schema:
             type: string
-            title: Event Id
+            title: Review Id
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSegmentResponse'
         '422':
           description: Validation Error
           content:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  /review/{event_id}/viewed:
+  /review/{review_id}/viewed:
     delete:
       tags:
         - Review
       summary: Set Not Reviewed
-      operationId: set_not_reviewed_review__event_id__viewed_delete
+      operationId: set_not_reviewed_review__review_id__viewed_delete
       parameters:
-        - name: event_id
+        - name: review_id
           in: path
           required: true
           schema:
             type: string
-            title: Event Id
+            title: Review Id
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -763,13 +692,25 @@ paths:
           content:
             application/json:
               schema: { }
+  /nvinfo:
+    get:
+      tags:
+        - App
+      summary: Nvinfo
+      operationId: nvinfo_nvinfo_get
+      responses:
+        '200':
+          description: Successful Response
+          content:
+            application/json:
+              schema: { }
   /logs/{service}:
     get:
       tags:
         - App
         - Logs
       summary: Logs
-      description: Get logs for the requested service (frigate/nginx/go2rtc/chroma)
+      description: Get logs for the requested service (frigate/nginx/go2rtc)
       operationId: logs_logs__service__get
       parameters:
         - name: service
@@ -781,7 +722,6 @@ paths:
               - frigate
               - nginx
               - go2rtc
-              - chroma
             title: Service
         - name: download
           in: query
@@ -1042,7 +982,8 @@ paths:
         - Preview
       summary: Preview Hour
       description: Get all mp4 previews relevant for time period given the timezone
-      operationId: preview_hour_preview__year_month___day___hour___camera_name___tz_name__get
+      operationId: >-
+        preview_hour_preview__year_month___day___hour___camera_name___tz_name__get
       parameters:
         - name: year_month
           in: path
@@ -1092,7 +1033,8 @@ paths:
         - Preview
       summary: Get Preview Frames From Cache
       description: Get list of cached preview frames
-      operationId: get_preview_frames_from_cache_preview__camera_name__start__start_ts__end__end_ts__frames_get
+      operationId: >-
+        get_preview_frames_from_cache_preview__camera_name__start__start_ts__end__end_ts__frames_get
       parameters:
         - name: camera_name
           in: path
@@ -1177,7 +1119,8 @@ paths:
       tags:
         - Export
       summary: Export Recording
-      operationId: export_recording_export__camera_name__start__start_time__end__end_time__post
+      operationId: >-
+        export_recording_export__camera_name__start__start_time__end__end_time__post
       parameters:
         - name: camera_name
           in: path
@@ -1656,6 +1599,30 @@ paths:
               - type: 'null'
             default: utc
             title: Timezone
+        - name: min_score
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: number
+              - type: 'null'
+            title: Min Score
+        - name: max_score
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: number
+              - type: 'null'
+            title: Max Score
+        - name: sort
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: string
+              - type: 'null'
+            title: Sort
       responses:
         '200':
           description: Successful Response
@@ -1942,6 +1909,15 @@ paths:
           schema:
             type: string
             title: Event Id
+        - name: source
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - $ref: '#/components/schemas/RegenerateDescriptionEnum'
+              - type: 'null'
+            default: thumbnails
+            title: Source
       responses:
         '200':
           description: Successful Response
@@ -2029,12 +2005,12 @@ paths:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  '{camera_name}':
+  /{camera_name}:
     get:
       tags:
         - Media
       summary: Mjpeg Feed
-      operationId: mjpeg_feed_camera_name__get
+      operationId: mjpeg_feed__camera_name__get
       parameters:
         - name: camera_name
           in: path
@@ -2241,7 +2217,8 @@ paths:
       tags:
         - Media
       summary: Get Snapshot From Recording
-      operationId: get_snapshot_from_recording__camera_name__recordings__frame_time__snapshot__format__get
+      operationId: >-
+        get_snapshot_from_recording__camera_name__recordings__frame_time__snapshot__format__get
       parameters:
         - name: camera_name
           in: path
@@ -2363,7 +2340,9 @@ paths:
       tags:
         - Media
       summary: Recordings
-      description: Return specific camera recordings between the given 'after'/'end' times. If not provided the last hour will be used
+      description: >-
+        Return specific camera recordings between the given 'after'/'end' times.
+        If not provided the last hour will be used
       operationId: recordings__camera_name__recordings_get
       parameters:
         - name: camera_name
@@ -2377,14 +2356,14 @@ paths:
           required: false
           schema:
             type: number
-            default: 1727542549.303557
+            default: 1729274204.653048
             title: After
         - name: before
           in: query
           required: false
           schema:
             type: number
-            default: 1727546149.303926
+            default: 1729277804.653095
             title: Before
       responses:
         '200':
@@ -2423,13 +2402,6 @@ paths:
           schema:
             type: number
             title: End Ts
-        - name: download
-          in: query
-          required: false
-          schema:
-            type: boolean
-            default: false
-            title: Download
       responses:
         '200':
           description: Successful Response
@@ -2800,13 +2772,6 @@ paths:
           schema:
             type: string
             title: Event Id
-        - name: download
-          in: query
-          required: false
-          schema:
-            type: boolean
-            default: false
-            title: Download
       responses:
         '200':
           description: Successful Response
@@ -3121,7 +3086,9 @@ paths:
       tags:
         - Media
       summary: Label Snapshot
-      description: Returns the snapshot image from the latest event for the given camera and label combo
+      description: >-
+        Returns the snapshot image from the latest event for the given camera
+        and label combo
       operationId: label_snapshot__camera_name___label__snapshot_jpg_get
       parameters:
         - name: camera_name
@@ -3193,6 +3160,32 @@ components:
       required:
         - password
       title: AppPutPasswordBody
+    DayReview:
+      properties:
+        day:
+          type: string
+          format: date-time
+          title: Day
+        reviewed_alert:
+          type: integer
+          title: Reviewed Alert
+        reviewed_detection:
+          type: integer
+          title: Reviewed Detection
+        total_alert:
+          type: integer
+          title: Total Alert
+        total_detection:
+          type: integer
+          title: Total Detection
+      type: object
+      required:
+        - day
+        - reviewed_alert
+        - reviewed_detection
+        - total_alert
+        - total_detection
+      title: DayReview
     EventsCreateBody:
       properties:
         source_type:
@@ -3237,7 +3230,6 @@ components:
         description:
           anyOf:
             - type: string
-              minLength: 1
             - type: 'null'
           title: The description of the event
       type: object
@@ -3278,6 +3270,19 @@ components:
         - jpg
         - jpeg
       title: Extension
+    GenericResponse:
+      properties:
+        success:
+          type: boolean
+          title: Success
+        message:
+          type: string
+          title: Message
+      type: object
+      required:
+        - success
+        - message
+      title: GenericResponse
     HTTPValidationError:
       properties:
         detail:
@@ -3287,6 +3292,133 @@ components:
           title: Detail
       type: object
       title: HTTPValidationError
+    Last24HoursReview:
+      properties:
+        reviewed_alert:
+          type: integer
+          title: Reviewed Alert
+        reviewed_detection:
+          type: integer
+          title: Reviewed Detection
+        total_alert:
+          type: integer
+          title: Total Alert
+        total_detection:
+          type: integer
+          title: Total Detection
+      type: object
+      required:
+        - reviewed_alert
+        - reviewed_detection
+        - total_alert
+        - total_detection
+      title: Last24HoursReview
+    RegenerateDescriptionEnum:
+      type: string
+      enum:
+        - thumbnails
+        - snapshot
+      title: RegenerateDescriptionEnum
+    ReviewActivityMotionResponse:
+      properties:
+        start_time:
+          type: integer
+          title: Start Time
+        motion:
+          type: number
+          title: Motion
+        camera:
+          type: string
+          title: Camera
+      type: object
+      required:
+        - start_time
+        - motion
+        - camera
+      title: ReviewActivityMotionResponse
+    ReviewDeleteMultipleReviewsBody:
+      properties:
+        ids:
+          items:
+            type: string
+            minLength: 1
+          type: array
+          minItems: 1
+          title: Ids
+      type: object
+      required:
+        - ids
+      title: ReviewDeleteMultipleReviewsBody
+    ReviewSegmentResponse:
+      properties:
+        id:
+          type: string
+          title: Id
+        camera:
+          type: string
+          title: Camera
+        start_time:
+          type: string
+          format: date-time
+          title: Start Time
+        end_time:
+          type: string
+          format: date-time
+          title: End Time
+        has_been_reviewed:
+          type: boolean
+          title: Has Been Reviewed
+        severity:
+          $ref: '#/components/schemas/SeverityEnum'
+        thumb_path:
+          type: string
+          title: Thumb Path
+        data:
+          title: Data
+      type: object
+      required:
+        - id
+        - camera
+        - start_time
+        - end_time
+        - has_been_reviewed
+        - severity
+        - thumb_path
+        - data
+      title: ReviewSegmentResponse
+    ReviewSetMultipleReviewedBody:
+      properties:
+        ids:
+          items:
+            type: string
+            minLength: 1
+          type: array
+          minItems: 1
+          title: Ids
+      type: object
+      required:
+        - ids
+      title: ReviewSetMultipleReviewedBody
+    ReviewSummaryResponse:
+      properties:
+        last24Hours:
+          $ref: '#/components/schemas/Last24HoursReview'
+        root:
+          additionalProperties:
+            $ref: '#/components/schemas/DayReview'
+          type: object
+          title: Root
+      type: object
+      required:
+        - last24Hours
+        - root
+      title: ReviewSummaryResponse
+    SeverityEnum:
+      type: string
+      enum:
+        - alert
+        - detection
+      title: SeverityEnum
     SubmitPlusBody:
       properties:
         include_annotation:

frigate/api/auth.py

@@ -357,6 +357,7 @@ def create_user(request: Request, body: AppPostUsersBody):
         {
             User.username: body.username,
             User.password_hash: password_hash,
+            User.notification_tokens: [],
         }
     ).execute()
     return JSONResponse(content={"username": body.username})

frigate/api/defs/events_body.py

@@ -11,9 +11,7 @@ class EventsSubLabelBody(BaseModel):
 
 
 class EventsDescriptionBody(BaseModel):
-    description: Union[str, None] = Field(
-        title="The description of the event", min_length=1
-    )
+    description: Union[str, None] = Field(title="The description of the event")
 
 
 class EventsCreateBody(BaseModel):

frigate/api/defs/events_query_parameters.py

@@ -35,7 +35,7 @@ class EventsQueryParams(BaseModel):
 class EventsSearchQueryParams(BaseModel):
     query: Optional[str] = None
     event_id: Optional[str] = None
-    search_type: Optional[str] = "thumbnail,description"
+    search_type: Optional[str] = "thumbnail"
     include_thumbnails: Optional[int] = 1
     limit: Optional[int] = 50
     cameras: Optional[str] = "all"
@@ -44,7 +44,12 @@ class EventsSearchQueryParams(BaseModel):
     after: Optional[float] = None
     before: Optional[float] = None
     time_range: Optional[str] = DEFAULT_TIME_RANGE
+    has_clip: Optional[bool] = None
+    has_snapshot: Optional[bool] = None
     timezone: Optional[str] = "utc"
+    min_score: Optional[float] = None
+    max_score: Optional[float] = None
+    sort: Optional[str] = None
 
 
 class EventsSummaryQueryParams(BaseModel):

frigate/api/defs/generic_response.py

@@ -0,0 +1,6 @@
+from pydantic import BaseModel
+
+
+class GenericResponse(BaseModel):
+    success: bool
+    message: str

frigate/api/defs/review_body.py

@@ -0,0 +1,6 @@
+from pydantic import BaseModel, conlist, constr
+
+
+class ReviewModifyMultipleBody(BaseModel):
+    # List of string with at least one element and each element with at least one char
+    ids: conlist(constr(min_length=1), min_length=1)

frigate/api/defs/review_query_parameters.py

@@ -1,28 +1,31 @@
-from typing import Optional
+from typing import Union
 
 from pydantic import BaseModel
+from pydantic.json_schema import SkipJsonSchema
+
+from frigate.review.maintainer import SeverityEnum
 
 
 class ReviewQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    labels: Optional[str] = "all"
-    zones: Optional[str] = "all"
-    reviewed: Optional[int] = 0
-    limit: Optional[int] = None
-    severity: Optional[str] = None
-    before: Optional[float] = None
-    after: Optional[float] = None
+    cameras: str = "all"
+    labels: str = "all"
+    zones: str = "all"
+    reviewed: int = 0
+    limit: Union[int, SkipJsonSchema[None]] = None
+    severity: Union[SeverityEnum, SkipJsonSchema[None]] = None
+    before: Union[float, SkipJsonSchema[None]] = None
+    after: Union[float, SkipJsonSchema[None]] = None
 
 
 class ReviewSummaryQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    labels: Optional[str] = "all"
-    zones: Optional[str] = "all"
-    timezone: Optional[str] = "utc"
+    cameras: str = "all"
+    labels: str = "all"
+    zones: str = "all"
+    timezone: str = "utc"
 
 
 class ReviewActivityMotionQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    before: Optional[float] = None
-    after: Optional[float] = None
-    scale: Optional[int] = 30
+    cameras: str = "all"
+    before: Union[float, SkipJsonSchema[None]] = None
+    after: Union[float, SkipJsonSchema[None]] = None
+    scale: int = 30

frigate/api/defs/review_responses.py

@@ -0,0 +1,43 @@
+from datetime import datetime
+from typing import Dict
+
+from pydantic import BaseModel, Json
+
+from frigate.review.maintainer import SeverityEnum
+
+
+class ReviewSegmentResponse(BaseModel):
+    id: str
+    camera: str
+    start_time: datetime
+    end_time: datetime
+    has_been_reviewed: bool
+    severity: SeverityEnum
+    thumb_path: str
+    data: Json
+
+
+class Last24HoursReview(BaseModel):
+    reviewed_alert: int
+    reviewed_detection: int
+    total_alert: int
+    total_detection: int
+
+
+class DayReview(BaseModel):
+    day: datetime
+    reviewed_alert: int
+    reviewed_detection: int
+    total_alert: int
+    total_detection: int
+
+
+class ReviewSummaryResponse(BaseModel):
+    last24Hours: Last24HoursReview
+    root: Dict[str, DayReview]
+
+
+class ReviewActivityMotionResponse(BaseModel):
+    start_time: int
+    motion: float
+    camera: str

frigate/api/event.py

@@ -259,66 +259,61 @@ def events(params: EventsQueryParams = Depends()):
 
 @router.get("/events/explore")
 def events_explore(limit: int = 10):
-    subquery = Event.select(
-        Event.id,
-        Event.camera,
-        Event.label,
-        Event.zones,
-        Event.start_time,
-        Event.end_time,
-        Event.has_clip,
-        Event.has_snapshot,
-        Event.plus_id,
-        Event.retain_indefinitely,
-        Event.sub_label,
-        Event.top_score,
-        Event.false_positive,
-        Event.box,
-        Event.data,
-        fn.rank()
-        .over(partition_by=[Event.label], order_by=[Event.start_time.desc()])
-        .alias("rank"),
-        fn.COUNT(Event.id).over(partition_by=[Event.label]).alias("event_count"),
-    ).alias("subquery")
+    # get distinct labels for all events
+    distinct_labels = Event.select(Event.label).distinct().order_by(Event.label)
 
-    query = (
-        Event.select(
-            subquery.c.id,
-            subquery.c.camera,
-            subquery.c.label,
-            subquery.c.zones,
-            subquery.c.start_time,
-            subquery.c.end_time,
-            subquery.c.has_clip,
-            subquery.c.has_snapshot,
-            subquery.c.plus_id,
-            subquery.c.retain_indefinitely,
-            subquery.c.sub_label,
-            subquery.c.top_score,
-            subquery.c.false_positive,
-            subquery.c.box,
-            subquery.c.data,
-            subquery.c.event_count,
-        )
-        .from_(subquery)
-        .where(subquery.c.rank <= limit)
-        .order_by(subquery.c.event_count.desc(), subquery.c.start_time.desc())
-        .dicts()
-    )
+    label_counts = {}
 
-    events = list(query.iterator())
+    def event_generator():
+        for label_obj in distinct_labels.iterator():
+            label = label_obj.label
 
-    processed_events = [
-        {k: v for k, v in event.items() if k != "data"}
-        | {
-            "data": {
-                k: v
-                for k, v in event["data"].items()
-                if k in ["type", "score", "top_score", "description"]
+            # get most recent events for this label
+            label_events = (
+                Event.select()
+                .where(Event.label == label)
+                .order_by(Event.start_time.desc())
+                .limit(limit)
+                .iterator()
+            )
+
+            # count total events for this label
+            label_counts[label] = Event.select().where(Event.label == label).count()
+
+            yield from label_events
+
+    def process_events():
+        for event in event_generator():
+            processed_event = {
+                "id": event.id,
+                "camera": event.camera,
+                "label": event.label,
+                "zones": event.zones,
+                "start_time": event.start_time,
+                "end_time": event.end_time,
+                "has_clip": event.has_clip,
+                "has_snapshot": event.has_snapshot,
+                "plus_id": event.plus_id,
+                "retain_indefinitely": event.retain_indefinitely,
+                "sub_label": event.sub_label,
+                "top_score": event.top_score,
+                "false_positive": event.false_positive,
+                "box": event.box,
+                "data": {
+                    k: v
+                    for k, v in event.data.items()
+                    if k in ["type", "score", "top_score", "description"]
+                },
+                "event_count": label_counts[event.label],
             }
-        }
-        for event in events
-    ]
+            yield processed_event
+
+    # convert iterator to list and sort
+    processed_events = sorted(
+        process_events(),
+        key=lambda x: (x["event_count"], x["start_time"]),
+        reverse=True,
+    )
 
     return JSONResponse(content=processed_events)
 
@@ -348,6 +343,7 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     search_type = params.search_type
     include_thumbnails = params.include_thumbnails
     limit = params.limit
+    sort = params.sort
 
     # Filters
     cameras = params.cameras
@@ -355,7 +351,11 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     zones = params.zones
     after = params.after
     before = params.before
+    min_score = params.min_score
+    max_score = params.max_score
     time_range = params.time_range
+    has_clip = params.has_clip
+    has_snapshot = params.has_snapshot
 
     # for similarity search
     event_id = params.event_id
@@ -394,6 +394,7 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
         Event.end_time,
         Event.has_clip,
         Event.has_snapshot,
+        Event.top_score,
         Event.data,
         Event.plus_id,
         ReviewSegment.thumb_path,
@@ -430,6 +431,20 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     if before:
         event_filters.append((Event.start_time < before))
 
+    if has_clip is not None:
+        event_filters.append((Event.has_clip == has_clip))
+
+    if has_snapshot is not None:
+        event_filters.append((Event.has_snapshot == has_snapshot))
+
+    if min_score is not None and max_score is not None:
+        event_filters.append((Event.data["score"].between(min_score, max_score)))
+    else:
+        if min_score is not None:
+            event_filters.append((Event.data["score"] >= min_score))
+        if max_score is not None:
+            event_filters.append((Event.data["score"] <= max_score))
+
     if time_range != DEFAULT_TIME_RANGE:
         tz_name = params.timezone
         hour_modifier, minute_modifier, _ = get_tz_modifiers(tz_name)
@@ -472,13 +487,8 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
                 status_code=404,
             )
 
-        thumb_result = context.embeddings.search_thumbnail(search_event)
-        thumb_ids = dict(
-            zip(
-                [result[0] for result in thumb_result],
-                context.thumb_stats.normalize([result[1] for result in thumb_result]),
-            )
-        )
+        thumb_result = context.search_thumbnail(search_event)
+        thumb_ids = {result[0]: result[1] for result in thumb_result}
         search_results = {
             event_id: {"distance": distance, "source": "thumbnail"}
             for event_id, distance in thumb_ids.items()
@@ -486,15 +496,18 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     else:
         search_types = search_type.split(",")
 
+        # only save stats for multi-modal searches
+        save_stats = "thumbnail" in search_types and "description" in search_types
+
         if "thumbnail" in search_types:
-            thumb_result = context.embeddings.search_thumbnail(query)
+            thumb_result = context.search_thumbnail(query)
+
+            thumb_distances = context.thumb_stats.normalize(
+                [result[1] for result in thumb_result], save_stats
+            )
+
             thumb_ids = dict(
-                zip(
-                    [result[0] for result in thumb_result],
-                    context.thumb_stats.normalize(
-                        [result[1] for result in thumb_result]
-                    ),
-                )
+                zip([result[0] for result in thumb_result], thumb_distances)
             )
             search_results.update(
                 {
@@ -504,13 +517,14 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
             )
 
         if "description" in search_types:
-            desc_result = context.embeddings.search_description(query)
-            desc_ids = dict(
-                zip(
-                    [result[0] for result in desc_result],
-                    context.desc_stats.normalize([result[1] for result in desc_result]),
-                )
+            desc_result = context.search_description(query)
+
+            desc_distances = context.desc_stats.normalize(
+                [result[1] for result in desc_result], save_stats
             )
+
+            desc_ids = dict(zip([result[0] for result in desc_result], desc_distances))
+
             for event_id, distance in desc_ids.items():
                 if (
                     event_id not in search_results
@@ -555,11 +569,19 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
 
         processed_events.append(processed_event)
 
-    # Sort by search distance if search_results are available, otherwise by start_time
+    # Sort by search distance if search_results are available, otherwise by start_time as default
     if search_results:
         processed_events.sort(key=lambda x: x.get("search_distance", float("inf")))
     else:
-        processed_events.sort(key=lambda x: x["start_time"], reverse=True)
+        if sort == "score_asc":
+            processed_events.sort(key=lambda x: x["score"])
+        elif sort == "score_desc":
+            processed_events.sort(key=lambda x: x["score"], reverse=True)
+        elif sort == "date_asc":
+            processed_events.sort(key=lambda x: x["start_time"])
+        else:
+            # "date_desc" default
+            processed_events.sort(key=lambda x: x["start_time"], reverse=True)
 
     # Limit the number of events returned
     processed_events = processed_events[:limit]
@@ -927,27 +949,19 @@ def set_description(
 
     new_description = body.description
 
-    if new_description is None or len(new_description) == 0:
-        return JSONResponse(
-            content=(
-                {
-                    "success": False,
-                    "message": "description cannot be empty",
-                }
-            ),
-            status_code=400,
-        )
-
     event.data["description"] = new_description
     event.save()
 
     # If semantic search is enabled, update the index
     if request.app.frigate_config.semantic_search.enabled:
         context: EmbeddingsContext = request.app.embeddings
-        context.embeddings.upsert_description(
-            event_id=event_id,
-            description=new_description,
-        )
+        if len(new_description) > 0:
+            context.update_description(
+                event_id,
+                new_description,
+            )
+        else:
+            context.db.delete_embeddings_description(event_ids=[event_id])
 
     response_message = (
         f"Event {event_id} description is now blank"
@@ -1001,7 +1015,7 @@ def regenerate_description(
         content=(
             {
                 "success": False,
-                "message": "Semantic search and generative AI are not enabled",
+                "message": "Semantic Search and Generative AI must be enabled to regenerate a description",
             }
         ),
         status_code=400,
@@ -1033,8 +1047,8 @@ def delete_event(request: Request, event_id: str):
     # If semantic search is enabled, update the index
     if request.app.frigate_config.semantic_search.enabled:
         context: EmbeddingsContext = request.app.embeddings
-        context.embeddings.delete_thumbnail(id=[event_id])
-        context.embeddings.delete_description(id=[event_id])
+        context.db.delete_embeddings_thumbnail(event_ids=[event_id])
+        context.db.delete_embeddings_description(event_ids=[event_id])
     return JSONResponse(
         content=({"success": True, "message": "Event " + event_id + " deleted"}),
         status_code=200,

frigate/api/export.py

@@ -13,8 +13,12 @@ from peewee import DoesNotExist
 
 from frigate.api.defs.tags import Tags
 from frigate.const import EXPORT_DIR
-from frigate.models import Export, Recordings
-from frigate.record.export import PlaybackFactorEnum, RecordingExporter
+from frigate.models import Export, Previews, Recordings
+from frigate.record.export import (
+    PlaybackFactorEnum,
+    PlaybackSourceEnum,
+    RecordingExporter,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -45,6 +49,7 @@ def export_recording(
 
     json: dict[str, any] = body or {}
     playback_factor = json.get("playback", "realtime")
+    playback_source = json.get("source", "recordings")
     friendly_name: Optional[str] = json.get("name")
 
     if len(friendly_name or "") > 256:
@@ -55,25 +60,48 @@ def export_recording(
 
     existing_image = json.get("image_path")
 
-    recordings_count = (
-        Recordings.select()
-        .where(
-            Recordings.start_time.between(start_time, end_time)
-            | Recordings.end_time.between(start_time, end_time)
-            | ((start_time > Recordings.start_time) & (end_time < Recordings.end_time))
+    if playback_source == "recordings":
+        recordings_count = (
+            Recordings.select()
+            .where(
+                Recordings.start_time.between(start_time, end_time)
+                | Recordings.end_time.between(start_time, end_time)
+                | (
+                    (start_time > Recordings.start_time)
+                    & (end_time < Recordings.end_time)
+                )
+            )
+            .where(Recordings.camera == camera_name)
+            .count()
         )
-        .where(Recordings.camera == camera_name)
-        .count()
-    )
 
-    if recordings_count <= 0:
-        return JSONResponse(
-            content=(
-                {"success": False, "message": "No recordings found for time range"}
-            ),
-            status_code=400,
+        if recordings_count <= 0:
+            return JSONResponse(
+                content=(
+                    {"success": False, "message": "No recordings found for time range"}
+                ),
+                status_code=400,
+            )
+    else:
+        previews_count = (
+            Previews.select()
+            .where(
+                Previews.start_time.between(start_time, end_time)
+                | Previews.end_time.between(start_time, end_time)
+                | ((start_time > Previews.start_time) & (end_time < Previews.end_time))
+            )
+            .where(Previews.camera == camera_name)
+            .count()
         )
 
+        if previews_count <= 0:
+            return JSONResponse(
+                content=(
+                    {"success": False, "message": "No previews found for time range"}
+                ),
+                status_code=400,
+            )
+
     export_id = f"{camera_name}_{''.join(random.choices(string.ascii_lowercase + string.digits, k=6))}"
     exporter = RecordingExporter(
         request.app.frigate_config,
@@ -88,6 +116,11 @@ def export_recording(
             if playback_factor in PlaybackFactorEnum.__members__.values()
             else PlaybackFactorEnum.realtime
         ),
+        (
+            PlaybackSourceEnum[playback_source]
+            if playback_source in PlaybackSourceEnum.__members__.values()
+            else PlaybackSourceEnum.recordings
+        ),
     )
     exporter.start()
     return JSONResponse(

frigate/api/fastapi_app.py

@@ -82,6 +82,10 @@ def create_fastapi_app(
             database.close()
         return response
 
+    @app.on_event("startup")
+    async def startup():
+        logger.info("FastAPI started")
+
     # Rate limiter (used for login endpoint)
     auth.rateLimiter.set_limit(frigate_config.auth.failed_login_rate_limit or "")
     app.state.limiter = limiter

frigate/api/media.py

@@ -7,6 +7,7 @@ import os
 import subprocess as sp
 import time
 from datetime import datetime, timedelta, timezone
+from pathlib import Path as FilePath
 from urllib.parse import unquote
 
 import cv2
@@ -450,8 +451,27 @@ def recording_clip(
     camera_name: str,
     start_ts: float,
     end_ts: float,
-    download: bool = False,
 ):
+    def run_download(ffmpeg_cmd: list[str], file_path: str):
+        with sp.Popen(
+            ffmpeg_cmd,
+            stderr=sp.PIPE,
+            stdout=sp.PIPE,
+            text=False,
+        ) as ffmpeg:
+            while True:
+                data = ffmpeg.stdout.read(8192)
+                if data is not None and len(data) > 0:
+                    yield data
+                else:
+                    if ffmpeg.returncode and ffmpeg.returncode != 0:
+                        logger.error(
+                            f"Failed to generate clip, ffmpeg logs: {ffmpeg.stderr.read()}"
+                        )
+                    else:
+                        FilePath(file_path).unlink(missing_ok=True)
+                    break
+
     recordings = (
         Recordings.select(
             Recordings.path,
@@ -467,18 +487,18 @@ def recording_clip(
         .order_by(Recordings.start_time.asc())
     )
 
-    playlist_lines = []
-    clip: Recordings
-    for clip in recordings:
-        playlist_lines.append(f"file '{clip.path}'")
-        # if this is the starting clip, add an inpoint
-        if clip.start_time < start_ts:
-            playlist_lines.append(f"inpoint {int(start_ts - clip.start_time)}")
-        # if this is the ending clip, add an outpoint
-        if clip.end_time > end_ts:
-            playlist_lines.append(f"outpoint {int(end_ts - clip.start_time)}")
-
-    file_name = sanitize_filename(f"clip_{camera_name}_{start_ts}-{end_ts}.mp4")
+    file_name = sanitize_filename(f"playlist_{camera_name}_{start_ts}-{end_ts}.txt")
+    file_path = f"/tmp/cache/{file_name}"
+    with open(file_path, "w") as file:
+        clip: Recordings
+        for clip in recordings:
+            file.write(f"file '{clip.path}'\n")
+            # if this is the starting clip, add an inpoint
+            if clip.start_time < start_ts:
+                file.write(f"inpoint {int(start_ts - clip.start_time)}\n")
+            # if this is the ending clip, add an outpoint
+            if clip.end_time > end_ts:
+                file.write(f"outpoint {int(end_ts - clip.start_time)}\n")
 
     if len(file_name) > 1000:
         return JSONResponse(
@@ -489,67 +509,32 @@ def recording_clip(
             status_code=403,
         )
 
-    path = os.path.join(CLIPS_DIR, f"cache/{file_name}")
-
     config: FrigateConfig = request.app.frigate_config
 
-    if not os.path.exists(path):
-        ffmpeg_cmd = [
-            config.ffmpeg.ffmpeg_path,
-            "-hide_banner",
-            "-y",
-            "-protocol_whitelist",
-            "pipe,file",
-            "-f",
-            "concat",
-            "-safe",
-            "0",
-            "-i",
-            "/dev/stdin",
-            "-c",
-            "copy",
-            "-movflags",
-            "+faststart",
-            path,
-        ]
-        p = sp.run(
-            ffmpeg_cmd,
-            input="\n".join(playlist_lines),
-            encoding="ascii",
-            capture_output=True,
-        )
+    ffmpeg_cmd = [
+        config.ffmpeg.ffmpeg_path,
+        "-hide_banner",
+        "-y",
+        "-protocol_whitelist",
+        "pipe,file",
+        "-f",
+        "concat",
+        "-safe",
+        "0",
+        "-i",
+        file_path,
+        "-c",
+        "copy",
+        "-movflags",
+        "frag_keyframe+empty_moov",
+        "-f",
+        "mp4",
+        "pipe:",
+    ]
 
-        if p.returncode != 0:
-            logger.error(p.stderr)
-            return JSONResponse(
-                content={
-                    "success": False,
-                    "message": "Could not create clip from recordings",
-                },
-                status_code=500,
-            )
-    else:
-        logger.debug(
-            f"Ignoring subsequent request for {path} as it already exists in the cache."
-        )
-
-    headers = {
-        "Content-Description": "File Transfer",
-        "Cache-Control": "no-cache",
-        "Content-Type": "video/mp4",
-        "Content-Length": str(os.path.getsize(path)),
-        # nginx: https://nginx.org/en/docs/http/ngx_http_proxy_module.html#proxy_ignore_headers
-        "X-Accel-Redirect": f"/clips/cache/{file_name}",
-    }
-
-    if download:
-        headers["Content-Disposition"] = "attachment; filename=%s" % file_name
-
-    return FileResponse(
-        path,
+    return StreamingResponse(
+        run_download(ffmpeg_cmd, file_path),
         media_type="video/mp4",
-        filename=file_name,
-        headers=headers,
     )
 
 
@@ -1028,7 +1013,7 @@ def event_snapshot_clean(request: Request, event_id: str, download: bool = False
 
 
 @router.get("/events/{event_id}/clip.mp4")
-def event_clip(request: Request, event_id: str, download: bool = False):
+def event_clip(request: Request, event_id: str):
     try:
         event: Event = Event.get(Event.id == event_id)
     except DoesNotExist:
@@ -1048,7 +1033,7 @@ def event_clip(request: Request, event_id: str, download: bool = False):
         end_ts = (
             datetime.now().timestamp() if event.end_time is None else event.end_time
         )
-        return recording_clip(request, event.camera, event.start_time, end_ts, download)
+        return recording_clip(request, event.camera, event.start_time, end_ts)
 
     headers = {
         "Content-Description": "File Transfer",
@@ -1059,9 +1044,6 @@ def event_clip(request: Request, event_id: str, download: bool = False):
         "X-Accel-Redirect": f"/clips/{file_name}",
     }
 
-    if download:
-        headers["Content-Disposition"] = "attachment; filename=%s" % file_name
-
     return FileResponse(
         clip_path,
         media_type="video/mp4",
@@ -1546,11 +1528,11 @@ def label_snapshot(request: Request, camera_name: str, label: str):
         )
 
     try:
-        event = event_query.get()
-        return event_snapshot(request, event.id)
+        event: Event = event_query.get()
+        return event_snapshot(request, event.id, MediaEventsSnapshotQueryParams())
     except DoesNotExist:
         frame = np.zeros((720, 1280, 3), np.uint8)
-        ret, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
+        _, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
 
         return Response(
             jpg.tobytes,

frigate/api/review.py

@@ -12,11 +12,18 @@ from fastapi.responses import JSONResponse
 from peewee import Case, DoesNotExist, fn, operator
 from playhouse.shortcuts import model_to_dict
 
+from frigate.api.defs.generic_response import GenericResponse
+from frigate.api.defs.review_body import ReviewModifyMultipleBody
 from frigate.api.defs.review_query_parameters import (
     ReviewActivityMotionQueryParams,
     ReviewQueryParams,
     ReviewSummaryQueryParams,
 )
+from frigate.api.defs.review_responses import (
+    ReviewActivityMotionResponse,
+    ReviewSegmentResponse,
+    ReviewSummaryResponse,
+)
 from frigate.api.defs.tags import Tags
 from frigate.models import Recordings, ReviewSegment
 from frigate.util.builtin import get_tz_modifiers
@@ -26,7 +33,7 @@ logger = logging.getLogger(__name__)
 router = APIRouter(tags=[Tags.review])
 
 
-@router.get("/review")
+@router.get("/review", response_model=list[ReviewSegmentResponse])
 def review(params: ReviewQueryParams = Depends()):
     cameras = params.cameras
     labels = params.labels
@@ -102,7 +109,7 @@ def review(params: ReviewQueryParams = Depends()):
     return JSONResponse(content=[r for r in review])
 
 
-@router.get("/review/summary")
+@router.get("/review/summary", response_model=ReviewSummaryResponse)
 def review_summary(params: ReviewSummaryQueryParams = Depends()):
     hour_modifier, minute_modifier, seconds_offset = get_tz_modifiers(params.timezone)
     day_ago = (datetime.datetime.now() - datetime.timedelta(hours=24)).timestamp()
@@ -173,18 +180,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("reviewed_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            ReviewSegment.has_been_reviewed,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("reviewed_motion"),
             fn.SUM(
                 Case(
                     None,
@@ -209,18 +204,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("total_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            1,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("total_motion"),
         )
         .where(reduce(operator.and_, clauses))
         .dicts()
@@ -282,18 +265,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("reviewed_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            ReviewSegment.has_been_reviewed,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("reviewed_motion"),
             fn.SUM(
                 Case(
                     None,
@@ -318,18 +289,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("total_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            1,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("total_motion"),
         )
         .where(reduce(operator.and_, clauses))
         .group_by(
@@ -348,19 +307,10 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
     return JSONResponse(content=data)
 
 
-@router.post("/reviews/viewed")
-def set_multiple_reviewed(body: dict = None):
-    json: dict[str, any] = body or {}
-    list_of_ids = json.get("ids", "")
-
-    if not list_of_ids or len(list_of_ids) == 0:
-        return JSONResponse(
-            context=({"success": False, "message": "Not a valid list of ids"}),
-            status_code=404,
-        )
-
+@router.post("/reviews/viewed", response_model=GenericResponse)
+def set_multiple_reviewed(body: ReviewModifyMultipleBody):
     ReviewSegment.update(has_been_reviewed=True).where(
-        ReviewSegment.id << list_of_ids
+        ReviewSegment.id << body.ids
     ).execute()
 
     return JSONResponse(
@@ -369,17 +319,9 @@ def set_multiple_reviewed(body: dict = None):
     )
 
 
-@router.post("/reviews/delete")
-def delete_reviews(body: dict = None):
-    json: dict[str, any] = body or {}
-    list_of_ids = json.get("ids", "")
-
-    if not list_of_ids or len(list_of_ids) == 0:
-        return JSONResponse(
-            content=({"success": False, "message": "Not a valid list of ids"}),
-            status_code=404,
-        )
-
+@router.post("/reviews/delete", response_model=GenericResponse)
+def delete_reviews(body: ReviewModifyMultipleBody):
+    list_of_ids = body.ids
     reviews = (
         ReviewSegment.select(
             ReviewSegment.camera,
@@ -424,7 +366,9 @@ def delete_reviews(body: dict = None):
     )
 
 
-@router.get("/review/activity/motion")
+@router.get(
+    "/review/activity/motion", response_model=list[ReviewActivityMotionResponse]
+)
 def motion_activity(params: ReviewActivityMotionQueryParams = Depends()):
     """Get motion and audio activity."""
     cameras = params.cameras
@@ -498,98 +442,44 @@ def motion_activity(params: ReviewActivityMotionQueryParams = Depends()):
     return JSONResponse(content=normalized)
 
 
-@router.get("/review/activity/audio")
-def audio_activity(params: ReviewActivityMotionQueryParams = Depends()):
-    """Get motion and audio activity."""
-    cameras = params.cameras
-    before = params.before or datetime.datetime.now().timestamp()
-    after = (
-        params.after
-        or (datetime.datetime.now() - datetime.timedelta(hours=1)).timestamp()
-    )
-    # get scale in seconds
-    scale = params.scale
-
-    clauses = [(Recordings.start_time > after) & (Recordings.end_time < before)]
-
-    if cameras != "all":
-        camera_list = cameras.split(",")
-        clauses.append((Recordings.camera << camera_list))
-
-    all_recordings: list[Recordings] = (
-        Recordings.select(
-            Recordings.start_time,
-            Recordings.duration,
-            Recordings.objects,
-            Recordings.dBFS,
-        )
-        .where(reduce(operator.and_, clauses))
-        .order_by(Recordings.start_time.asc())
-        .iterator()
-    )
-
-    # format is: { timestamp: segment_start_ts, motion: [0-100], audio: [0 - -100] }
-    # periods where active objects / audio was detected will cause audio to be scaled down
-    data: list[dict[str, float]] = []
-
-    for rec in all_recordings:
-        data.append(
-            {
-                "start_time": rec.start_time,
-                "audio": rec.dBFS if rec.objects == 0 else 0,
-            }
-        )
-
-    # resample data using pandas to get activity on scaled basis
-    df = pd.DataFrame(data, columns=["start_time", "audio"])
-    df = df.astype(dtype={"audio": "float16"})
-
-    # set date as datetime index
-    df["start_time"] = pd.to_datetime(df["start_time"], unit="s")
-    df.set_index(["start_time"], inplace=True)
-
-    # normalize data
-    df = df.resample(f"{scale}S").mean().fillna(0.0)
-    df["audio"] = (
-        (df["audio"] - df["audio"].max())
-        / (df["audio"].min() - df["audio"].max())
-        * -100
-    )
-
-    # change types for output
-    df.index = df.index.astype(int) // (10**9)
-    normalized = df.reset_index().to_dict("records")
-    return JSONResponse(content=normalized)
-
-
-@router.get("/review/event/{event_id}")
+@router.get("/review/event/{event_id}", response_model=ReviewSegmentResponse)
 def get_review_from_event(event_id: str):
     try:
-        return model_to_dict(
-            ReviewSegment.get(
-                ReviewSegment.data["detections"].cast("text") % f'*"{event_id}"*'
+        return JSONResponse(
+            model_to_dict(
+                ReviewSegment.get(
+                    ReviewSegment.data["detections"].cast("text") % f'*"{event_id}"*'
+                )
             )
         )
     except DoesNotExist:
-        return "Review item not found", 404
+        return JSONResponse(
+            content={"success": False, "message": "Review item not found"},
+            status_code=404,
+        )
 
 
-@router.get("/review/{event_id}")
-def get_review(event_id: str):
+@router.get("/review/{review_id}", response_model=ReviewSegmentResponse)
+def get_review(review_id: str):
     try:
-        return model_to_dict(ReviewSegment.get(ReviewSegment.id == event_id))
+        return JSONResponse(
+            content=model_to_dict(ReviewSegment.get(ReviewSegment.id == review_id))
+        )
     except DoesNotExist:
-        return "Review item not found", 404
+        return JSONResponse(
+            content={"success": False, "message": "Review item not found"},
+            status_code=404,
+        )
 
 
-@router.delete("/review/{event_id}/viewed")
-def set_not_reviewed(event_id: str):
+@router.delete("/review/{review_id}/viewed", response_model=GenericResponse)
+def set_not_reviewed(review_id: str):
     try:
-        review: ReviewSegment = ReviewSegment.get(ReviewSegment.id == event_id)
+        review: ReviewSegment = ReviewSegment.get(ReviewSegment.id == review_id)
     except DoesNotExist:
         return JSONResponse(
             content=(
-                {"success": False, "message": "Review " + event_id + " not found"}
+                {"success": False, "message": "Review " + review_id + " not found"}
             ),
             status_code=404,
         )
@@ -598,6 +488,8 @@ def set_not_reviewed(event_id: str):
     review.save()
 
     return JSONResponse(
-        content=({"success": True, "message": "Reviewed " + event_id + " not viewed"}),
+        content=(
+            {"success": True, "message": "Set Review " + review_id + " as not viewed"}
+        ),
         status_code=200,
     )

frigate/app.py

@@ -63,6 +63,7 @@ from frigate.record.cleanup import RecordingCleanup
 from frigate.record.export import migrate_exports
 from frigate.record.record import manage_recordings
 from frigate.review.review import manage_review_segments
+from frigate.service_manager import ServiceManager
 from frigate.stats.emitter import StatsEmitter
 from frigate.stats.util import stats_init
 from frigate.storage import StorageMaintainer
@@ -78,7 +79,6 @@ logger = logging.getLogger(__name__)
 
 class FrigateApp:
     def __init__(self, config: FrigateConfig) -> None:
-        self.audio_process: Optional[mp.Process] = None
         self.stop_event: MpEvent = mp.Event()
         self.detection_queue: Queue = mp.Queue()
         self.detectors: dict[str, ObjectDetectProcess] = {}
@@ -449,9 +449,8 @@ class FrigateApp:
         ]
 
         if audio_cameras:
-            self.audio_process = AudioProcessor(audio_cameras, self.camera_metrics)
-            self.audio_process.start()
-            self.processes["audio_detector"] = self.audio_process.pid or 0
+            proc = AudioProcessor(audio_cameras, self.camera_metrics).start(wait=True)
+            self.processes["audio_detector"] = proc.pid or 0
 
     def start_timeline_processor(self) -> None:
         self.timeline_processor = TimelineProcessor(
@@ -581,12 +580,12 @@ class FrigateApp:
         self.init_recording_manager()
         self.init_review_segment_manager()
         self.init_go2rtc()
+        self.start_detectors()
+        self.init_embeddings_manager()
         self.bind_database()
         self.check_db_data_migrations()
         self.init_inter_process_communicator()
         self.init_dispatcher()
-        self.start_detectors()
-        self.init_embeddings_manager()
         self.init_embeddings_client()
         self.start_video_output_processor()
         self.start_ptz_autotracker()
@@ -639,11 +638,6 @@ class FrigateApp:
             ReviewSegment.end_time == None
         ).execute()
 
-        # stop the audio process
-        if self.audio_process:
-            self.audio_process.terminate()
-            self.audio_process.join()
-
         # ensure the capture processes are done
         for camera, metrics in self.camera_metrics.items():
             capture_process = metrics.capture_process
@@ -699,7 +693,7 @@ class FrigateApp:
 
         # Save embeddings stats to disk
         if self.embeddings:
-            self.embeddings.save_stats()
+            self.embeddings.stop()
 
         # Stop Communicators
         self.inter_process_communicator.stop()
@@ -712,4 +706,6 @@ class FrigateApp:
             shm.close()
             shm.unlink()
 
+        ServiceManager.current().shutdown(wait=True)
+
         os._exit(os.EX_OK)

frigate/comms/dispatcher.py

@@ -15,6 +15,7 @@ from frigate.const import (
     INSERT_PREVIEW,
     REQUEST_REGION_GRID,
     UPDATE_CAMERA_ACTIVITY,
+    UPDATE_EMBEDDINGS_REINDEX_PROGRESS,
     UPDATE_EVENT_DESCRIPTION,
     UPDATE_MODEL_STATE,
     UPSERT_REVIEW_SEGMENT,
@@ -63,6 +64,9 @@ class Dispatcher:
         self.onvif = onvif
         self.ptz_metrics = ptz_metrics
         self.comms = communicators
+        self.camera_activity = {}
+        self.model_state = {}
+        self.embeddings_reindex = {}
 
         self._camera_settings_handlers: dict[str, Callable] = {
             "audio": self._on_audio_command,
@@ -84,37 +88,25 @@ class Dispatcher:
         for comm in self.comms:
             comm.subscribe(self._receive)
 
-        self.camera_activity = {}
-        self.model_state = {}
-
     def _receive(self, topic: str, payload: str) -> Optional[Any]:
         """Handle receiving of payload from communicators."""
-        if topic.endswith("set"):
+
+        def handle_camera_command(command_type, camera_name, command, payload):
             try:
-                # example /cam_name/detect/set payload=ON|OFF
-                if topic.count("/") == 2:
-                    camera_name = topic.split("/")[-3]
-                    command = topic.split("/")[-2]
+                if command_type == "set":
                     self._camera_settings_handlers[command](camera_name, payload)
-                elif topic.count("/") == 1:
-                    command = topic.split("/")[-2]
-                    self._global_settings_handlers[command](payload)
-            except IndexError:
-                logger.error(f"Received invalid set command: {topic}")
-                return
-        elif topic.endswith("ptz"):
-            try:
-                # example /cam_name/ptz payload=MOVE_UP|MOVE_DOWN|STOP...
-                camera_name = topic.split("/")[-2]
-                self._on_ptz_command(camera_name, payload)
-            except IndexError:
-                logger.error(f"Received invalid ptz command: {topic}")
-                return
-        elif topic == "restart":
+                elif command_type == "ptz":
+                    self._on_ptz_command(camera_name, payload)
+            except KeyError:
+                logger.error(f"Invalid command type or handler: {command_type}")
+
+        def handle_restart():
             restart_frigate()
-        elif topic == INSERT_MANY_RECORDINGS:
+
+        def handle_insert_many_recordings():
             Recordings.insert_many(payload).execute()
-        elif topic == REQUEST_REGION_GRID:
+
+        def handle_request_region_grid():
             camera = payload
             grid = get_camera_regions_grid(
                 camera,
@@ -122,24 +114,25 @@ class Dispatcher:
                 max(self.config.model.width, self.config.model.height),
             )
             return grid
-        elif topic == INSERT_PREVIEW:
+
+        def handle_insert_preview():
             Previews.insert(payload).execute()
-        elif topic == UPSERT_REVIEW_SEGMENT:
-            (
-                ReviewSegment.insert(payload)
-                .on_conflict(
-                    conflict_target=[ReviewSegment.id],
-                    update=payload,
-                )
-                .execute()
-            )
-        elif topic == CLEAR_ONGOING_REVIEW_SEGMENTS:
-            ReviewSegment.update(end_time=datetime.datetime.now().timestamp()).where(
-                ReviewSegment.end_time == None
+
+        def handle_upsert_review_segment():
+            ReviewSegment.insert(payload).on_conflict(
+                conflict_target=[ReviewSegment.id],
+                update=payload,
             ).execute()
-        elif topic == UPDATE_CAMERA_ACTIVITY:
+
+        def handle_clear_ongoing_review_segments():
+            ReviewSegment.update(end_time=datetime.datetime.now().timestamp()).where(
+                ReviewSegment.end_time.is_null(True)
+            ).execute()
+
+        def handle_update_camera_activity():
             self.camera_activity = payload
-        elif topic == UPDATE_EVENT_DESCRIPTION:
+
+        def handle_update_event_description():
             event: Event = Event.get(Event.id == payload["id"])
             event.data["description"] = payload["description"]
             event.save()
@@ -147,15 +140,31 @@ class Dispatcher:
                 "event_update",
                 json.dumps({"id": event.id, "description": event.data["description"]}),
             )
-        elif topic == UPDATE_MODEL_STATE:
-            model = payload["model"]
-            state = payload["state"]
-            self.model_state[model] = ModelStatusTypesEnum[state]
-            self.publish("model_state", json.dumps(self.model_state))
-        elif topic == "modelState":
-            model_state = self.model_state.copy()
-            self.publish("model_state", json.dumps(model_state))
-        elif topic == "onConnect":
+
+        def handle_update_model_state():
+            if payload:
+                model = payload["model"]
+                state = payload["state"]
+                self.model_state[model] = ModelStatusTypesEnum[state]
+                self.publish("model_state", json.dumps(self.model_state))
+
+        def handle_model_state():
+            self.publish("model_state", json.dumps(self.model_state.copy()))
+
+        def handle_update_embeddings_reindex_progress():
+            self.embeddings_reindex = payload
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(payload),
+            )
+
+        def handle_embeddings_reindex_progress():
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(self.embeddings_reindex.copy()),
+            )
+
+        def handle_on_connect():
             camera_status = self.camera_activity.copy()
 
             for camera in camera_status.keys():
@@ -170,6 +179,51 @@ class Dispatcher:
                 }
 
             self.publish("camera_activity", json.dumps(camera_status))
+            self.publish("model_state", json.dumps(self.model_state.copy()))
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(self.embeddings_reindex.copy()),
+            )
+
+        # Dictionary mapping topic to handlers
+        topic_handlers = {
+            INSERT_MANY_RECORDINGS: handle_insert_many_recordings,
+            REQUEST_REGION_GRID: handle_request_region_grid,
+            INSERT_PREVIEW: handle_insert_preview,
+            UPSERT_REVIEW_SEGMENT: handle_upsert_review_segment,
+            CLEAR_ONGOING_REVIEW_SEGMENTS: handle_clear_ongoing_review_segments,
+            UPDATE_CAMERA_ACTIVITY: handle_update_camera_activity,
+            UPDATE_EVENT_DESCRIPTION: handle_update_event_description,
+            UPDATE_MODEL_STATE: handle_update_model_state,
+            UPDATE_EMBEDDINGS_REINDEX_PROGRESS: handle_update_embeddings_reindex_progress,
+            "restart": handle_restart,
+            "embeddingsReindexProgress": handle_embeddings_reindex_progress,
+            "modelState": handle_model_state,
+            "onConnect": handle_on_connect,
+        }
+
+        if topic.endswith("set") or topic.endswith("ptz"):
+            try:
+                parts = topic.split("/")
+                if len(parts) == 3 and topic.endswith("set"):
+                    # example /cam_name/detect/set payload=ON|OFF
+                    camera_name = parts[-3]
+                    command = parts[-2]
+                    handle_camera_command("set", camera_name, command, payload)
+                elif len(parts) == 2 and topic.endswith("set"):
+                    command = parts[-2]
+                    self._global_settings_handlers[command](payload)
+                elif len(parts) == 2 and topic.endswith("ptz"):
+                    # example /cam_name/ptz payload=MOVE_UP|MOVE_DOWN|STOP...
+                    camera_name = parts[-2]
+                    handle_camera_command("ptz", camera_name, "", payload)
+            except IndexError:
+                logger.error(
+                    f"Received invalid {topic.split('/')[-1]} command: {topic}"
+                )
+                return
+        elif topic in topic_handlers:
+            return topic_handlers[topic]()
         else:
             self.publish(topic, payload, retain=False)
 

frigate/comms/embeddings_updater.py

@@ -0,0 +1,65 @@
+"""Facilitates communication between processes."""
+
+from enum import Enum
+from typing import Callable
+
+import zmq
+
+SOCKET_REP_REQ = "ipc:///tmp/cache/embeddings"
+
+
+class EmbeddingsRequestEnum(Enum):
+    embed_description = "embed_description"
+    embed_thumbnail = "embed_thumbnail"
+    generate_search = "generate_search"
+
+
+class EmbeddingsResponder:
+    def __init__(self) -> None:
+        self.context = zmq.Context()
+        self.socket = self.context.socket(zmq.REP)
+        self.socket.bind(SOCKET_REP_REQ)
+
+    def check_for_request(self, process: Callable) -> None:
+        while True:  # load all messages that are queued
+            has_message, _, _ = zmq.select([self.socket], [], [], 0.1)
+
+            if not has_message:
+                break
+
+            try:
+                (topic, value) = self.socket.recv_json(flags=zmq.NOBLOCK)
+
+                response = process(topic, value)
+
+                if response is not None:
+                    self.socket.send_json(response)
+                else:
+                    self.socket.send_json([])
+            except zmq.ZMQError:
+                break
+
+    def stop(self) -> None:
+        self.socket.close()
+        self.context.destroy()
+
+
+class EmbeddingsRequestor:
+    """Simplifies sending data to EmbeddingsResponder and getting a reply."""
+
+    def __init__(self) -> None:
+        self.context = zmq.Context()
+        self.socket = self.context.socket(zmq.REQ)
+        self.socket.connect(SOCKET_REP_REQ)
+
+    def send_data(self, topic: str, data: any) -> str:
+        """Sends data and then waits for reply."""
+        try:
+            self.socket.send_json((topic, data))
+            return self.socket.recv_json()
+        except zmq.ZMQError:
+            return ""
+
+    def stop(self) -> None:
+        self.socket.close()
+        self.context.destroy()

frigate/comms/event_metadata_updater.py

@@ -39,7 +39,7 @@ class EventMetadataSubscriber(Subscriber):
         super().__init__(topic)
 
     def check_for_update(
-        self, timeout: float = None
+        self, timeout: float = 1
     ) -> Optional[tuple[EventMetadataTypeEnum, str, RegenerateDescriptionEnum]]:
         return super().check_for_update(timeout)
 

frigate/comms/inter_process.py

@@ -65,8 +65,11 @@ class InterProcessRequestor:
 
     def send_data(self, topic: str, data: any) -> any:
         """Sends data and then waits for reply."""
-        self.socket.send_json((topic, data))
-        return self.socket.recv_json()
+        try:
+            self.socket.send_json((topic, data))
+            return self.socket.recv_json()
+        except zmq.ZMQError:
+            return ""
 
     def stop(self) -> None:
         self.socket.close()

frigate/comms/mqtt.py

@@ -17,7 +17,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
     def __init__(self, config: FrigateConfig) -> None:
         self.config = config
         self.mqtt_config = config.mqtt
-        self.connected: bool = False
+        self.connected = False
 
     def subscribe(self, receiver: Callable) -> None:
         """Wrapper for allowing dispatcher to subscribe."""
@@ -27,7 +27,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
     def publish(self, topic: str, payload: Any, retain: bool = False) -> None:
         """Wrapper for publishing when client is in valid state."""
         if not self.connected:
-            logger.error(f"Unable to publish to {topic}: client is not connected")
+            logger.debug(f"Unable to publish to {topic}: client is not connected")
             return
 
         self.client.publish(
@@ -173,6 +173,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
             client_id=self.mqtt_config.client_id,
         )
         self.client.on_connect = self._on_connect
+        self.client.on_disconnect = self._on_disconnect
         self.client.will_set(
             self.mqtt_config.topic_prefix + "/available",
             payload="offline",
@@ -197,14 +198,6 @@ class MqttClient(Communicator):  # type: ignore[misc]
 
         for name in self.config.cameras.keys():
             for callback in callback_types:
-                # We need to pre-clear existing set topics because in previous
-                # versions the webUI retained on the /set topic but this is
-                # no longer the case.
-                self.client.publish(
-                    f"{self.mqtt_config.topic_prefix}/{name}/{callback}/set",
-                    None,
-                    retain=True,
-                )
                 self.client.message_callback_add(
                     f"{self.mqtt_config.topic_prefix}/{name}/{callback}/set",
                     self.on_mqtt_command,

frigate/config/camera/genai.py

@@ -23,7 +23,7 @@ class GenAICameraConfig(BaseModel):
         default=False, title="Use snapshots for generating descriptions."
     )
     prompt: str = Field(
-        default="Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.",
+        default="Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.",
         title="Default caption prompt.",
     )
     object_prompts: dict[str, str] = Field(
@@ -51,7 +51,7 @@ class GenAICameraConfig(BaseModel):
 class GenAIConfig(FrigateBaseModel):
     enabled: bool = Field(default=False, title="Enable GenAI.")
     prompt: str = Field(
-        default="Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.",
+        default="Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.",
         title="Default caption prompt.",
     )
     object_prompts: dict[str, str] = Field(

frigate/config/semantic_search.py

@@ -12,3 +12,6 @@ class SemanticSearchConfig(FrigateBaseModel):
     reindex: Optional[bool] = Field(
         default=False, title="Reindex all detections on startup."
     )
+    model_size: str = Field(
+        default="small", title="The size of the embeddings model used."
+    )

frigate/const.py

@@ -17,7 +17,21 @@ PLUS_API_HOST = "https://api.frigate.video"
 
 DEFAULT_ATTRIBUTE_LABEL_MAP = {
     "person": ["amazon", "face"],
-    "car": ["amazon", "fedex", "license_plate", "ups"],
+    "car": [
+        "amazon",
+        "an_post",
+        "dhl",
+        "dpd",
+        "fedex",
+        "gls",
+        "license_plate",
+        "nzpost",
+        "postnl",
+        "postnord",
+        "purolator",
+        "ups",
+        "usps",
+    ],
 }
 LABEL_CONSOLIDATION_MAP = {
     "car": 0.8,
@@ -85,6 +99,7 @@ CLEAR_ONGOING_REVIEW_SEGMENTS = "clear_ongoing_review_segments"
 UPDATE_CAMERA_ACTIVITY = "update_camera_activity"
 UPDATE_EVENT_DESCRIPTION = "update_event_description"
 UPDATE_MODEL_STATE = "update_model_state"
+UPDATE_EMBEDDINGS_REINDEX_PROGRESS = "handle_embeddings_reindex_progress"
 
 # Stats Values
 

frigate/db/sqlitevecq.py

@@ -20,3 +20,34 @@ class SqliteVecQueueDatabase(SqliteQueueDatabase):
         conn.enable_load_extension(True)
         conn.load_extension(self.sqlite_vec_path)
         conn.enable_load_extension(False)
+
+    def delete_embeddings_thumbnail(self, event_ids: list[str]) -> None:
+        ids = ",".join(["?" for _ in event_ids])
+        self.execute_sql(f"DELETE FROM vec_thumbnails WHERE id IN ({ids})", event_ids)
+
+    def delete_embeddings_description(self, event_ids: list[str]) -> None:
+        ids = ",".join(["?" for _ in event_ids])
+        self.execute_sql(f"DELETE FROM vec_descriptions WHERE id IN ({ids})", event_ids)
+
+    def drop_embeddings_tables(self) -> None:
+        self.execute_sql("""
+            DROP TABLE vec_descriptions;
+        """)
+        self.execute_sql("""
+            DROP TABLE vec_thumbnails;
+        """)
+
+    def create_embeddings_tables(self) -> None:
+        """Create vec0 virtual table for embeddings"""
+        self.execute_sql("""
+            CREATE VIRTUAL TABLE IF NOT EXISTS vec_thumbnails USING vec0(
+                id TEXT PRIMARY KEY,
+                thumbnail_embedding FLOAT[768] distance_metric=cosine
+            );
+        """)
+        self.execute_sql("""
+            CREATE VIRTUAL TABLE IF NOT EXISTS vec_descriptions USING vec0(
+                id TEXT PRIMARY KEY,
+                description_embedding FLOAT[768] distance_metric=cosine
+            );
+        """)

frigate/detectors/detector_config.py

@@ -59,6 +59,7 @@ class ModelConfig(BaseModel):
     _merged_labelmap: Optional[Dict[int, str]] = PrivateAttr()
     _colormap: Dict[int, Tuple[int, int, int]] = PrivateAttr()
     _all_attributes: list[str] = PrivateAttr()
+    _all_attribute_logos: list[str] = PrivateAttr()
     _model_hash: str = PrivateAttr()
 
     @property
@@ -73,6 +74,10 @@ class ModelConfig(BaseModel):
     def all_attributes(self) -> list[str]:
         return self._all_attributes
 
+    @property
+    def all_attribute_logos(self) -> list[str]:
+        return self._all_attribute_logos
+
     @property
     def model_hash(self) -> str:
         return self._model_hash
@@ -93,6 +98,9 @@ class ModelConfig(BaseModel):
             unique_attributes.update(attributes)
 
         self._all_attributes = list(unique_attributes)
+        self._all_attribute_logos = list(
+            unique_attributes - set(["face", "license_plate"])
+        )
 
     def check_and_load_plus_model(
         self, plus_api: PlusApi, detector: str = None
@@ -140,6 +148,9 @@ class ModelConfig(BaseModel):
             unique_attributes.update(attributes)
 
         self._all_attributes = list(unique_attributes)
+        self._all_attribute_logos = list(
+            unique_attributes - set(["face", "license_plate"])
+        )
 
         self._merged_labelmap = {
             **{int(key): val for key, val in model_info["labelMap"].items()},
@@ -157,10 +168,14 @@ class ModelConfig(BaseModel):
             self._model_hash = file_hash.hexdigest()
 
     def create_colormap(self, enabled_labels: set[str]) -> None:
-        """Get a list of colors for enabled labels."""
-        colors = generate_color_palette(len(enabled_labels))
-
-        self._colormap = {label: color for label, color in zip(enabled_labels, colors)}
+        """Get a list of colors for enabled labels that aren't attributes."""
+        enabled_trackable_labels = list(
+            filter(lambda label: label not in self._all_attributes, enabled_labels)
+        )
+        colors = generate_color_palette(len(enabled_trackable_labels))
+        self._colormap = {
+            label: color for label, color in zip(enabled_trackable_labels, colors)
+        }
 
     model_config = ConfigDict(extra="forbid", protected_namespaces=())
 

frigate/detectors/plugins/openvino.py

@@ -3,6 +3,7 @@ import os
 
 import numpy as np
 import openvino as ov
+import openvino.properties as props
 from pydantic import Field
 from typing_extensions import Literal
 
@@ -34,6 +35,8 @@ class OvDetector(DetectionApi):
             logger.error(f"OpenVino model file {detector_config.model.path} not found.")
             raise FileNotFoundError
 
+        os.makedirs("/config/model_cache/openvino", exist_ok=True)
+        self.ov_core.set_property({props.cache_dir: "/config/model_cache/openvino"})
         self.interpreter = self.ov_core.compile_model(
             model=detector_config.model.path, device_name=detector_config.device
         )

frigate/embeddings/__init__.py

@@ -7,17 +7,18 @@ import os
 import signal
 import threading
 from types import FrameType
-from typing import Optional
+from typing import Optional, Union
 
 from setproctitle import setproctitle
 
+from frigate.comms.embeddings_updater import EmbeddingsRequestEnum, EmbeddingsRequestor
 from frigate.config import FrigateConfig
 from frigate.const import CONFIG_DIR
 from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event
+from frigate.util.builtin import serialize
 from frigate.util.services import listen
 
-from .embeddings import Embeddings
 from .maintainer import EmbeddingMaintainer
 from .util import ZScoreNormalization
 
@@ -55,12 +56,6 @@ def manage_embeddings(config: FrigateConfig) -> None:
     models = [Event]
     db.bind(models)
 
-    embeddings = Embeddings(db)
-
-    # Check if we need to re-index events
-    if config.semantic_search.reindex:
-        embeddings.reindex()
-
     maintainer = EmbeddingMaintainer(
         db,
         config,
@@ -71,9 +66,10 @@ def manage_embeddings(config: FrigateConfig) -> None:
 
 class EmbeddingsContext:
     def __init__(self, db: SqliteVecQueueDatabase):
-        self.embeddings = Embeddings(db)
+        self.db = db
         self.thumb_stats = ZScoreNormalization()
-        self.desc_stats = ZScoreNormalization(scale_factor=3, bias=-2.5)
+        self.desc_stats = ZScoreNormalization()
+        self.requestor = EmbeddingsRequestor()
 
         # load stats from disk
         try:
@@ -84,7 +80,7 @@ class EmbeddingsContext:
         except FileNotFoundError:
             pass
 
-    def save_stats(self):
+    def stop(self):
         """Write the stats to disk as JSON on exit."""
         contents = {
             "thumb_stats": self.thumb_stats.to_dict(),
@@ -92,3 +88,109 @@ class EmbeddingsContext:
         }
         with open(os.path.join(CONFIG_DIR, ".search_stats.json"), "w") as f:
             json.dump(contents, f)
+        self.requestor.stop()
+
+    def search_thumbnail(
+        self, query: Union[Event, str], event_ids: list[str] = None
+    ) -> list[tuple[str, float]]:
+        if query.__class__ == Event:
+            cursor = self.db.execute_sql(
+                """
+                SELECT thumbnail_embedding FROM vec_thumbnails WHERE id = ?
+                """,
+                [query.id],
+            )
+
+            row = cursor.fetchone() if cursor else None
+
+            if row:
+                query_embedding = row[0]
+            else:
+                # If no embedding found, generate it and return it
+                data = self.requestor.send_data(
+                    EmbeddingsRequestEnum.embed_thumbnail.value,
+                    {"id": str(query.id), "thumbnail": str(query.thumbnail)},
+                )
+
+                if not data:
+                    return []
+
+                query_embedding = serialize(data)
+        else:
+            data = self.requestor.send_data(
+                EmbeddingsRequestEnum.generate_search.value, query
+            )
+
+            if not data:
+                return []
+
+            query_embedding = serialize(data)
+
+        sql_query = """
+            SELECT
+                id,
+                distance
+            FROM vec_thumbnails
+            WHERE thumbnail_embedding MATCH ?
+                AND k = 100
+        """
+
+        # Add the IN clause if event_ids is provided and not empty
+        # this is the only filter supported by sqlite-vec as of 0.1.3
+        # but it seems to be broken in this version
+        if event_ids:
+            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
+
+        # order by distance DESC is not implemented in this version of sqlite-vec
+        # when it's implemented, we can use cosine similarity
+        sql_query += " ORDER BY distance"
+
+        parameters = [query_embedding] + event_ids if event_ids else [query_embedding]
+
+        results = self.db.execute_sql(sql_query, parameters).fetchall()
+
+        return results
+
+    def search_description(
+        self, query_text: str, event_ids: list[str] = None
+    ) -> list[tuple[str, float]]:
+        data = self.requestor.send_data(
+            EmbeddingsRequestEnum.generate_search.value, query_text
+        )
+
+        if not data:
+            return []
+
+        query_embedding = serialize(data)
+
+        # Prepare the base SQL query
+        sql_query = """
+            SELECT
+                id,
+                distance
+            FROM vec_descriptions
+            WHERE description_embedding MATCH ?
+                AND k = 100
+        """
+
+        # Add the IN clause if event_ids is provided and not empty
+        # this is the only filter supported by sqlite-vec as of 0.1.3
+        # but it seems to be broken in this version
+        if event_ids:
+            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
+
+        # order by distance DESC is not implemented in this version of sqlite-vec
+        # when it's implemented, we can use cosine similarity
+        sql_query += " ORDER BY distance"
+
+        parameters = [query_embedding] + event_ids if event_ids else [query_embedding]
+
+        results = self.db.execute_sql(sql_query, parameters).fetchall()
+
+        return results
+
+    def update_description(self, event_id: str, description: str) -> None:
+        self.requestor.send_data(
+            EmbeddingsRequestEnum.embed_description.value,
+            {"id": event_id, "description": description},
+        )

frigate/embeddings/embeddings.py

@@ -1,23 +1,26 @@
 """SQLite-vec embeddings database."""
 
 import base64
-import io
 import logging
-import struct
+import os
 import time
-from typing import List, Tuple, Union
 
-from PIL import Image
+from numpy import ndarray
 from playhouse.shortcuts import model_to_dict
 
 from frigate.comms.inter_process import InterProcessRequestor
-from frigate.const import UPDATE_MODEL_STATE
+from frigate.config.semantic_search import SemanticSearchConfig
+from frigate.const import (
+    CONFIG_DIR,
+    UPDATE_EMBEDDINGS_REINDEX_PROGRESS,
+    UPDATE_MODEL_STATE,
+)
 from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event
 from frigate.types import ModelStatusTypesEnum
+from frigate.util.builtin import serialize
 
-from .functions.clip import ClipEmbedding
-from .functions.minilm_l6_v2 import MiniLMEmbedding
+from .functions.onnx import GenericONNXEmbedding, ModelTypeEnum
 
 logger = logging.getLogger(__name__)
 
@@ -53,31 +56,26 @@ def get_metadata(event: Event) -> dict:
     )
 
 
-def serialize(vector: List[float]) -> bytes:
-    """Serializes a list of floats into a compact "raw bytes" format"""
-    return struct.pack("%sf" % len(vector), *vector)
-
-
-def deserialize(bytes_data: bytes) -> List[float]:
-    """Deserializes a compact "raw bytes" format into a list of floats"""
-    return list(struct.unpack("%sf" % (len(bytes_data) // 4), bytes_data))
-
-
 class Embeddings:
     """SQLite-vec embeddings database."""
 
-    def __init__(self, db: SqliteVecQueueDatabase) -> None:
+    def __init__(
+        self, config: SemanticSearchConfig, db: SqliteVecQueueDatabase
+    ) -> None:
+        self.config = config
         self.db = db
         self.requestor = InterProcessRequestor()
 
         # Create tables if they don't exist
-        self._create_tables()
+        self.db.create_embeddings_tables()
 
         models = [
-            "sentence-transformers/all-MiniLM-L6-v2-model.onnx",
-            "sentence-transformers/all-MiniLM-L6-v2-tokenizer",
-            "clip-clip_image_model_vitb32.onnx",
-            "clip-clip_text_model_vitb32.onnx",
+            "jinaai/jina-clip-v1-text_model_fp16.onnx",
+            "jinaai/jina-clip-v1-tokenizer",
+            "jinaai/jina-clip-v1-vision_model_fp16.onnx"
+            if config.model_size == "large"
+            else "jinaai/jina-clip-v1-vision_model_quantized.onnx",
+            "jinaai/jina-clip-v1-preprocessor_config.json",
         ]
 
         for model in models:
@@ -89,171 +87,172 @@ class Embeddings:
                 },
             )
 
-        self.clip_embedding = ClipEmbedding(
-            preferred_providers=["CPUExecutionProvider"]
-        )
-        self.minilm_embedding = MiniLMEmbedding(
-            preferred_providers=["CPUExecutionProvider"],
+        self.text_embedding = GenericONNXEmbedding(
+            model_name="jinaai/jina-clip-v1",
+            model_file="text_model_fp16.onnx",
+            tokenizer_file="tokenizer",
+            download_urls={
+                "text_model_fp16.onnx": "https://huggingface.co/jinaai/jina-clip-v1/resolve/main/onnx/text_model_fp16.onnx",
+            },
+            model_size=config.model_size,
+            model_type=ModelTypeEnum.text,
+            requestor=self.requestor,
+            device="CPU",
         )
 
-    def _create_tables(self):
-        # Create vec0 virtual table for thumbnail embeddings
-        self.db.execute_sql("""
-            CREATE VIRTUAL TABLE IF NOT EXISTS vec_thumbnails USING vec0(
-                id TEXT PRIMARY KEY,
-                thumbnail_embedding FLOAT[512]
-            );
-        """)
-
-        # Create vec0 virtual table for description embeddings
-        self.db.execute_sql("""
-            CREATE VIRTUAL TABLE IF NOT EXISTS vec_descriptions USING vec0(
-                id TEXT PRIMARY KEY,
-                description_embedding FLOAT[384]
-            );
-        """)
-
-    def upsert_thumbnail(self, event_id: str, thumbnail: bytes):
-        # Convert thumbnail bytes to PIL Image
-        image = Image.open(io.BytesIO(thumbnail)).convert("RGB")
-        # Generate embedding using CLIP
-        embedding = self.clip_embedding([image])[0]
-
-        self.db.execute_sql(
-            """
-            INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
-            VALUES(?, ?)
-            """,
-            (event_id, serialize(embedding)),
+        model_file = (
+            "vision_model_fp16.onnx"
+            if self.config.model_size == "large"
+            else "vision_model_quantized.onnx"
         )
 
-        return embedding
-
-    def upsert_description(self, event_id: str, description: str):
-        # Generate embedding using MiniLM
-        embedding = self.minilm_embedding([description])[0]
-
-        self.db.execute_sql(
-            """
-            INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
-            VALUES(?, ?)
-            """,
-            (event_id, serialize(embedding)),
-        )
-
-        return embedding
-
-    def delete_thumbnail(self, event_ids: List[str]) -> None:
-        ids = ",".join(["?" for _ in event_ids])
-        self.db.execute_sql(
-            f"DELETE FROM vec_thumbnails WHERE id IN ({ids})", event_ids
-        )
-
-    def delete_description(self, event_ids: List[str]) -> None:
-        ids = ",".join(["?" for _ in event_ids])
-        self.db.execute_sql(
-            f"DELETE FROM vec_descriptions WHERE id IN ({ids})", event_ids
-        )
-
-    def search_thumbnail(
-        self, query: Union[Event, str], event_ids: List[str] = None
-    ) -> List[Tuple[str, float]]:
-        if query.__class__ == Event:
-            cursor = self.db.execute_sql(
-                """
-                SELECT thumbnail_embedding FROM vec_thumbnails WHERE id = ?
-                """,
-                [query.id],
-            )
-
-            row = cursor.fetchone() if cursor else None
-
-            if row:
-                query_embedding = deserialize(
-                    row[0]
-                )  # Deserialize the thumbnail embedding
-            else:
-                # If no embedding found, generate it and return it
-                thumbnail = base64.b64decode(query.thumbnail)
-                query_embedding = self.upsert_thumbnail(query.id, thumbnail)
-        else:
-            query_embedding = self.clip_embedding([query])[0]
-
-        sql_query = """
-            SELECT
-                id,
-                distance
-            FROM vec_thumbnails
-            WHERE thumbnail_embedding MATCH ?
-                AND k = 100
-        """
-
-        # Add the IN clause if event_ids is provided and not empty
-        # this is the only filter supported by sqlite-vec as of 0.1.3
-        # but it seems to be broken in this version
-        if event_ids:
-            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
-
-        # order by distance DESC is not implemented in this version of sqlite-vec
-        # when it's implemented, we can use cosine similarity
-        sql_query += " ORDER BY distance"
-
-        parameters = (
-            [serialize(query_embedding)] + event_ids
-            if event_ids
-            else [serialize(query_embedding)]
-        )
-
-        results = self.db.execute_sql(sql_query, parameters).fetchall()
-
-        return results
-
-    def search_description(
-        self, query_text: str, event_ids: List[str] = None
-    ) -> List[Tuple[str, float]]:
-        query_embedding = self.minilm_embedding([query_text])[0]
-
-        # Prepare the base SQL query
-        sql_query = """
-            SELECT
-                id,
-                distance
-            FROM vec_descriptions
-            WHERE description_embedding MATCH ?
-                AND k = 100
-        """
-
-        # Add the IN clause if event_ids is provided and not empty
-        # this is the only filter supported by sqlite-vec as of 0.1.3
-        # but it seems to be broken in this version
-        if event_ids:
-            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
-
-        # order by distance DESC is not implemented in this version of sqlite-vec
-        # when it's implemented, we can use cosine similarity
-        sql_query += " ORDER BY distance"
-
-        parameters = (
-            [serialize(query_embedding)] + event_ids
-            if event_ids
-            else [serialize(query_embedding)]
-        )
-
-        results = self.db.execute_sql(sql_query, parameters).fetchall()
-
-        return results
-
-    def reindex(self) -> None:
-        logger.info("Indexing event embeddings...")
-
-        st = time.time()
-        totals = {
-            "thumb": 0,
-            "desc": 0,
+        download_urls = {
+            model_file: f"https://huggingface.co/jinaai/jina-clip-v1/resolve/main/onnx/{model_file}",
+            "preprocessor_config.json": "https://huggingface.co/jinaai/jina-clip-v1/resolve/main/preprocessor_config.json",
         }
 
-        batch_size = 100
+        self.vision_embedding = GenericONNXEmbedding(
+            model_name="jinaai/jina-clip-v1",
+            model_file=model_file,
+            download_urls=download_urls,
+            model_size=config.model_size,
+            model_type=ModelTypeEnum.vision,
+            requestor=self.requestor,
+            device="GPU" if config.model_size == "large" else "CPU",
+        )
+
+    def embed_thumbnail(
+        self, event_id: str, thumbnail: bytes, upsert: bool = True
+    ) -> ndarray:
+        """Embed thumbnail and optionally insert into DB.
+
+        @param: event_id in Events DB
+        @param: thumbnail bytes in jpg format
+        @param: upsert If embedding should be upserted into vec DB
+        """
+        # Convert thumbnail bytes to PIL Image
+        embedding = self.vision_embedding([thumbnail])[0]
+
+        if upsert:
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
+                VALUES(?, ?)
+                """,
+                (event_id, serialize(embedding)),
+            )
+
+        return embedding
+
+    def batch_embed_thumbnail(
+        self, event_thumbs: dict[str, bytes], upsert: bool = True
+    ) -> list[ndarray]:
+        """Embed thumbnails and optionally insert into DB.
+
+        @param: event_thumbs Map of Event IDs in DB to thumbnail bytes in jpg format
+        @param: upsert If embedding should be upserted into vec DB
+        """
+        ids = list(event_thumbs.keys())
+        embeddings = self.vision_embedding(list(event_thumbs.values()))
+
+        if upsert:
+            items = []
+
+            for i in range(len(ids)):
+                items.append(ids[i])
+                items.append(serialize(embeddings[i]))
+
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
+                VALUES {}
+                """.format(", ".join(["(?, ?)"] * len(ids))),
+                items,
+            )
+
+        return embeddings
+
+    def embed_description(
+        self, event_id: str, description: str, upsert: bool = True
+    ) -> ndarray:
+        embedding = self.text_embedding([description])[0]
+
+        if upsert:
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
+                VALUES(?, ?)
+                """,
+                (event_id, serialize(embedding)),
+            )
+
+        return embedding
+
+    def batch_embed_description(
+        self, event_descriptions: dict[str, str], upsert: bool = True
+    ) -> ndarray:
+        # upsert embeddings one by one to avoid token limit
+        embeddings = []
+
+        for desc in event_descriptions.values():
+            embeddings.append(self.text_embedding([desc])[0])
+
+        if upsert:
+            ids = list(event_descriptions.keys())
+            items = []
+
+            for i in range(len(ids)):
+                items.append(ids[i])
+                items.append(serialize(embeddings[i]))
+
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
+                VALUES {}
+                """.format(", ".join(["(?, ?)"] * len(ids))),
+                items,
+            )
+
+        return embeddings
+
+    def reindex(self) -> None:
+        logger.info("Indexing tracked object embeddings...")
+
+        self.db.drop_embeddings_tables()
+        logger.debug("Dropped embeddings tables.")
+        self.db.create_embeddings_tables()
+        logger.debug("Created embeddings tables.")
+
+        # Delete the saved stats file
+        if os.path.exists(os.path.join(CONFIG_DIR, ".search_stats.json")):
+            os.remove(os.path.join(CONFIG_DIR, ".search_stats.json"))
+
+        st = time.time()
+
+        # Get total count of events to process
+        total_events = (
+            Event.select()
+            .where(
+                (Event.has_clip == True | Event.has_snapshot == True)
+                & Event.thumbnail.is_null(False)
+            )
+            .count()
+        )
+
+        batch_size = 32
         current_page = 1
+
+        totals = {
+            "thumbnails": 0,
+            "descriptions": 0,
+            "processed_objects": total_events - 1 if total_events < batch_size else 0,
+            "total_objects": total_events,
+            "time_remaining": 0 if total_events < batch_size else -1,
+            "status": "indexing",
+        }
+
+        self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)
+
         events = (
             Event.select()
             .where(
@@ -266,14 +265,45 @@ class Embeddings:
 
         while len(events) > 0:
             event: Event
+            batch_thumbs = {}
+            batch_descs = {}
             for event in events:
-                thumbnail = base64.b64decode(event.thumbnail)
-                self.upsert_thumbnail(event.id, thumbnail)
-                totals["thumb"] += 1
-                if description := event.data.get("description", "").strip():
-                    totals["desc"] += 1
-                    self.upsert_description(event.id, description)
+                batch_thumbs[event.id] = base64.b64decode(event.thumbnail)
+                totals["thumbnails"] += 1
 
+                if description := event.data.get("description", "").strip():
+                    batch_descs[event.id] = description
+                    totals["descriptions"] += 1
+
+                totals["processed_objects"] += 1
+
+            # run batch embedding
+            self.batch_embed_thumbnail(batch_thumbs)
+
+            if batch_descs:
+                self.batch_embed_description(batch_descs)
+
+            # report progress every batch so we don't spam the logs
+            progress = (totals["processed_objects"] / total_events) * 100
+            logger.debug(
+                "Processed %d/%d events (%.2f%% complete) | Thumbnails: %d, Descriptions: %d",
+                totals["processed_objects"],
+                total_events,
+                progress,
+                totals["thumbnails"],
+                totals["descriptions"],
+            )
+
+            # Calculate time remaining
+            elapsed_time = time.time() - st
+            avg_time_per_event = elapsed_time / totals["processed_objects"]
+            remaining_events = total_events - totals["processed_objects"]
+            time_remaining = avg_time_per_event * remaining_events
+            totals["time_remaining"] = int(time_remaining)
+
+            self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)
+
+            # Move to the next page
             current_page += 1
             events = (
                 Event.select()
@@ -287,7 +317,10 @@ class Embeddings:
 
         logger.info(
             "Embedded %d thumbnails and %d descriptions in %s seconds",
-            totals["thumb"],
-            totals["desc"],
-            time.time() - st,
+            totals["thumbnails"],
+            totals["descriptions"],
+            round(time.time() - st, 1),
         )
+        totals["status"] = "completed"
+
+        self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)

frigate/embeddings/functions/clip.py

@@ -1,166 +0,0 @@
-import logging
-import os
-from typing import List, Optional, Union
-
-import numpy as np
-import onnxruntime as ort
-from onnx_clip import OnnxClip, Preprocessor, Tokenizer
-from PIL import Image
-
-from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
-from frigate.types import ModelStatusTypesEnum
-from frigate.util.downloader import ModelDownloader
-
-logger = logging.getLogger(__name__)
-
-
-class Clip(OnnxClip):
-    """Override load models to use pre-downloaded models from cache directory."""
-
-    def __init__(
-        self,
-        model: str = "ViT-B/32",
-        batch_size: Optional[int] = None,
-        providers: List[str] = ["CPUExecutionProvider"],
-    ):
-        """
-        Instantiates the model and required encoding classes.
-
-        Args:
-            model: The model to utilize. Currently ViT-B/32 and RN50 are
-                allowed.
-            batch_size: If set, splits the lists in `get_image_embeddings`
-                and `get_text_embeddings` into batches of this size before
-                passing them to the model. The embeddings are then concatenated
-                back together before being returned. This is necessary when
-                passing large amounts of data (perhaps ~100 or more).
-        """
-        allowed_models = ["ViT-B/32", "RN50"]
-        if model not in allowed_models:
-            raise ValueError(f"`model` must be in {allowed_models}. Got {model}.")
-        if model == "ViT-B/32":
-            self.embedding_size = 512
-        elif model == "RN50":
-            self.embedding_size = 1024
-        self.image_model, self.text_model = self._load_models(model, providers)
-        self._tokenizer = Tokenizer()
-        self._preprocessor = Preprocessor()
-        self._batch_size = batch_size
-
-    @staticmethod
-    def _load_models(
-        model: str,
-        providers: List[str],
-    ) -> tuple[ort.InferenceSession, ort.InferenceSession]:
-        """
-        Load models from cache directory.
-        """
-        if model == "ViT-B/32":
-            IMAGE_MODEL_FILE = "clip_image_model_vitb32.onnx"
-            TEXT_MODEL_FILE = "clip_text_model_vitb32.onnx"
-        elif model == "RN50":
-            IMAGE_MODEL_FILE = "clip_image_model_rn50.onnx"
-            TEXT_MODEL_FILE = "clip_text_model_rn50.onnx"
-        else:
-            raise ValueError(f"Unexpected model {model}. No `.onnx` file found.")
-
-        models = []
-        for model_file in [IMAGE_MODEL_FILE, TEXT_MODEL_FILE]:
-            path = os.path.join(MODEL_CACHE_DIR, "clip", model_file)
-            models.append(Clip._load_model(path, providers))
-
-        return models[0], models[1]
-
-    @staticmethod
-    def _load_model(path: str, providers: List[str]):
-        if os.path.exists(path):
-            return ort.InferenceSession(path, providers=providers)
-        else:
-            logger.warning(f"CLIP model file {path} not found.")
-            return None
-
-
-class ClipEmbedding:
-    """Embedding function for CLIP model."""
-
-    def __init__(
-        self,
-        model: str = "ViT-B/32",
-        silent: bool = False,
-        preferred_providers: List[str] = ["CPUExecutionProvider"],
-    ):
-        self.model_name = model
-        self.silent = silent
-        self.preferred_providers = preferred_providers
-        self.model_files = self._get_model_files()
-        self.model = None
-
-        self.downloader = ModelDownloader(
-            model_name="clip",
-            download_path=os.path.join(MODEL_CACHE_DIR, "clip"),
-            file_names=self.model_files,
-            download_func=self._download_model,
-            silent=self.silent,
-        )
-        self.downloader.ensure_model_files()
-
-    def _get_model_files(self):
-        if self.model_name == "ViT-B/32":
-            return ["clip_image_model_vitb32.onnx", "clip_text_model_vitb32.onnx"]
-        elif self.model_name == "RN50":
-            return ["clip_image_model_rn50.onnx", "clip_text_model_rn50.onnx"]
-        else:
-            raise ValueError(
-                f"Unexpected model {self.model_name}. No `.onnx` file found."
-            )
-
-    def _download_model(self, path: str):
-        s3_url = (
-            f"https://lakera-clip.s3.eu-west-1.amazonaws.com/{os.path.basename(path)}"
-        )
-        try:
-            ModelDownloader.download_from_url(s3_url, path, self.silent)
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.model_name}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.downloaded,
-                },
-            )
-        except Exception:
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.model_name}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.error,
-                },
-            )
-
-    def _load_model(self):
-        if self.model is None:
-            self.downloader.wait_for_download()
-            self.model = Clip(self.model_name, providers=self.preferred_providers)
-
-    def __call__(self, input: Union[List[str], List[Image.Image]]) -> List[np.ndarray]:
-        self._load_model()
-        if (
-            self.model is None
-            or self.model.image_model is None
-            or self.model.text_model is None
-        ):
-            logger.info(
-                "CLIP model is not fully loaded. Please wait for the download to complete."
-            )
-            return []
-
-        embeddings = []
-        for item in input:
-            if isinstance(item, Image.Image):
-                result = self.model.get_image_embeddings([item])
-                embeddings.append(result[0])
-            elif isinstance(item, str):
-                result = self.model.get_text_embeddings([item])
-                embeddings.append(result[0])
-            else:
-                raise ValueError(f"Unsupported input type: {type(item)}")
-        return embeddings

frigate/embeddings/functions/minilm_l6_v2.py

@@ -1,107 +0,0 @@
-import logging
-import os
-from typing import List
-
-import numpy as np
-import onnxruntime as ort
-
-# importing this without pytorch or others causes a warning
-# https://github.com/huggingface/transformers/issues/27214
-# suppressed by setting env TRANSFORMERS_NO_ADVISORY_WARNINGS=1
-from transformers import AutoTokenizer
-
-from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
-from frigate.types import ModelStatusTypesEnum
-from frigate.util.downloader import ModelDownloader
-
-logger = logging.getLogger(__name__)
-
-
-class MiniLMEmbedding:
-    """Embedding function for ONNX MiniLM-L6 model."""
-
-    DOWNLOAD_PATH = f"{MODEL_CACHE_DIR}/all-MiniLM-L6-v2"
-    MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
-    IMAGE_MODEL_FILE = "model.onnx"
-    TOKENIZER_FILE = "tokenizer"
-
-    def __init__(self, preferred_providers=["CPUExecutionProvider"]):
-        self.preferred_providers = preferred_providers
-        self.tokenizer = None
-        self.session = None
-
-        self.downloader = ModelDownloader(
-            model_name=self.MODEL_NAME,
-            download_path=self.DOWNLOAD_PATH,
-            file_names=[self.IMAGE_MODEL_FILE, self.TOKENIZER_FILE],
-            download_func=self._download_model,
-        )
-        self.downloader.ensure_model_files()
-
-    def _download_model(self, path: str):
-        try:
-            if os.path.basename(path) == self.IMAGE_MODEL_FILE:
-                s3_url = f"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/onnx/{self.IMAGE_MODEL_FILE}"
-                ModelDownloader.download_from_url(s3_url, path)
-            elif os.path.basename(path) == self.TOKENIZER_FILE:
-                logger.info("Downloading MiniLM tokenizer")
-                tokenizer = AutoTokenizer.from_pretrained(
-                    self.MODEL_NAME, clean_up_tokenization_spaces=True
-                )
-                tokenizer.save_pretrained(path)
-
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.MODEL_NAME}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.downloaded,
-                },
-            )
-        except Exception:
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.MODEL_NAME}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.error,
-                },
-            )
-
-    def _load_model_and_tokenizer(self):
-        if self.tokenizer is None or self.session is None:
-            self.downloader.wait_for_download()
-            self.tokenizer = self._load_tokenizer()
-            self.session = self._load_model(
-                os.path.join(self.DOWNLOAD_PATH, self.IMAGE_MODEL_FILE),
-                self.preferred_providers,
-            )
-
-    def _load_tokenizer(self):
-        tokenizer_path = os.path.join(self.DOWNLOAD_PATH, self.TOKENIZER_FILE)
-        return AutoTokenizer.from_pretrained(
-            tokenizer_path, clean_up_tokenization_spaces=True
-        )
-
-    def _load_model(self, path: str, providers: List[str]):
-        if os.path.exists(path):
-            return ort.InferenceSession(path, providers=providers)
-        else:
-            logger.warning(f"MiniLM model file {path} not found.")
-            return None
-
-    def __call__(self, texts: List[str]) -> List[np.ndarray]:
-        self._load_model_and_tokenizer()
-
-        if self.session is None or self.tokenizer is None:
-            logger.error("MiniLM model or tokenizer is not loaded.")
-            return []
-
-        inputs = self.tokenizer(
-            texts, padding=True, truncation=True, return_tensors="np"
-        )
-        input_names = [input.name for input in self.session.get_inputs()]
-        onnx_inputs = {name: inputs[name] for name in input_names if name in inputs}
-
-        outputs = self.session.run(None, onnx_inputs)
-        embeddings = outputs[0].mean(axis=1)
-
-        return [embedding for embedding in embeddings]

frigate/embeddings/functions/onnx.py

@@ -0,0 +1,216 @@
+import logging
+import os
+import warnings
+from enum import Enum
+from io import BytesIO
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+import requests
+from PIL import Image
+
+# importing this without pytorch or others causes a warning
+# https://github.com/huggingface/transformers/issues/27214
+# suppressed by setting env TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+from transformers import AutoFeatureExtractor, AutoTokenizer
+from transformers.utils.logging import disable_progress_bar
+
+from frigate.comms.inter_process import InterProcessRequestor
+from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
+from frigate.types import ModelStatusTypesEnum
+from frigate.util.downloader import ModelDownloader
+from frigate.util.model import ONNXModelRunner
+
+warnings.filterwarnings(
+    "ignore",
+    category=FutureWarning,
+    message="The class CLIPFeatureExtractor is deprecated",
+)
+
+# disables the progress bar for downloading tokenizers and feature extractors
+disable_progress_bar()
+logger = logging.getLogger(__name__)
+
+
+class ModelTypeEnum(str, Enum):
+    face = "face"
+    vision = "vision"
+    text = "text"
+
+
+class GenericONNXEmbedding:
+    """Generic embedding function for ONNX models (text and vision)."""
+
+    def __init__(
+        self,
+        model_name: str,
+        model_file: str,
+        download_urls: Dict[str, str],
+        model_size: str,
+        model_type: str,
+        requestor: InterProcessRequestor,
+        tokenizer_file: Optional[str] = None,
+        device: str = "AUTO",
+    ):
+        self.model_name = model_name
+        self.model_file = model_file
+        self.tokenizer_file = tokenizer_file
+        self.requestor = requestor
+        self.download_urls = download_urls
+        self.model_type = model_type  # 'text' or 'vision'
+        self.model_size = model_size
+        self.device = device
+        self.download_path = os.path.join(MODEL_CACHE_DIR, self.model_name)
+        self.tokenizer = None
+        self.feature_extractor = None
+        self.runner = None
+        files_names = list(self.download_urls.keys()) + (
+            [self.tokenizer_file] if self.tokenizer_file else []
+        )
+
+        if not all(
+            os.path.exists(os.path.join(self.download_path, n)) for n in files_names
+        ):
+            logger.debug(f"starting model download for {self.model_name}")
+            self.downloader = ModelDownloader(
+                model_name=self.model_name,
+                download_path=self.download_path,
+                file_names=files_names,
+                download_func=self._download_model,
+            )
+            self.downloader.ensure_model_files()
+        else:
+            self.downloader = None
+            ModelDownloader.mark_files_state(
+                self.requestor,
+                self.model_name,
+                files_names,
+                ModelStatusTypesEnum.downloaded,
+            )
+            self._load_model_and_tokenizer()
+            logger.debug(f"models are already downloaded for {self.model_name}")
+
+    def _download_model(self, path: str):
+        try:
+            file_name = os.path.basename(path)
+            if file_name in self.download_urls:
+                ModelDownloader.download_from_url(self.download_urls[file_name], path)
+            elif (
+                file_name == self.tokenizer_file
+                and self.model_type == ModelTypeEnum.text
+            ):
+                if not os.path.exists(path + "/" + self.model_name):
+                    logger.info(f"Downloading {self.model_name} tokenizer")
+                tokenizer = AutoTokenizer.from_pretrained(
+                    self.model_name,
+                    trust_remote_code=True,
+                    cache_dir=f"{MODEL_CACHE_DIR}/{self.model_name}/tokenizer",
+                    clean_up_tokenization_spaces=True,
+                )
+                tokenizer.save_pretrained(path)
+
+            self.downloader.requestor.send_data(
+                UPDATE_MODEL_STATE,
+                {
+                    "model": f"{self.model_name}-{file_name}",
+                    "state": ModelStatusTypesEnum.downloaded,
+                },
+            )
+        except Exception:
+            self.downloader.requestor.send_data(
+                UPDATE_MODEL_STATE,
+                {
+                    "model": f"{self.model_name}-{file_name}",
+                    "state": ModelStatusTypesEnum.error,
+                },
+            )
+
+    def _load_model_and_tokenizer(self):
+        if self.runner is None:
+            if self.downloader:
+                self.downloader.wait_for_download()
+            if self.model_type == ModelTypeEnum.text:
+                self.tokenizer = self._load_tokenizer()
+            else:
+                self.feature_extractor = self._load_feature_extractor()
+            self.runner = ONNXModelRunner(
+                os.path.join(self.download_path, self.model_file),
+                self.device,
+                self.model_size,
+            )
+
+    def _load_tokenizer(self):
+        tokenizer_path = os.path.join(f"{MODEL_CACHE_DIR}/{self.model_name}/tokenizer")
+        return AutoTokenizer.from_pretrained(
+            self.model_name,
+            cache_dir=tokenizer_path,
+            trust_remote_code=True,
+            clean_up_tokenization_spaces=True,
+        )
+
+    def _load_feature_extractor(self):
+        return AutoFeatureExtractor.from_pretrained(
+            f"{MODEL_CACHE_DIR}/{self.model_name}",
+        )
+
+    def _preprocess_inputs(self, raw_inputs: any) -> any:
+        if self.model_type == ModelTypeEnum.text:
+            max_length = max(len(self.tokenizer.encode(text)) for text in raw_inputs)
+            return [
+                self.tokenizer(
+                    text,
+                    padding="max_length",
+                    truncation=True,
+                    max_length=max_length,
+                    return_tensors="np",
+                )
+                for text in raw_inputs
+            ]
+        elif self.model_type == ModelTypeEnum.vision:
+            processed_images = [self._process_image(img) for img in raw_inputs]
+            return [
+                self.feature_extractor(images=image, return_tensors="np")
+                for image in processed_images
+            ]
+        else:
+            raise ValueError(f"Unable to preprocess inputs for {self.model_type}")
+
+    def _process_image(self, image):
+        if isinstance(image, str):
+            if image.startswith("http"):
+                response = requests.get(image)
+                image = Image.open(BytesIO(response.content)).convert("RGB")
+        elif isinstance(image, bytes):
+            image = Image.open(BytesIO(image)).convert("RGB")
+
+        return image
+
+    def __call__(
+        self, inputs: Union[List[str], List[Image.Image], List[str]]
+    ) -> List[np.ndarray]:
+        self._load_model_and_tokenizer()
+        if self.runner is None or (
+            self.tokenizer is None and self.feature_extractor is None
+        ):
+            logger.error(
+                f"{self.model_name} model or tokenizer/feature extractor is not loaded."
+            )
+            return []
+
+        processed_inputs = self._preprocess_inputs(inputs)
+        input_names = self.runner.get_input_names()
+        onnx_inputs = {name: [] for name in input_names}
+        input: dict[str, any]
+        for input in processed_inputs:
+            for key, value in input.items():
+                if key in input_names:
+                    onnx_inputs[key].append(value[0])
+
+        for key in input_names:
+            if onnx_inputs.get(key):
+                onnx_inputs[key] = np.stack(onnx_inputs[key])
+            else:
+                logger.warning(f"Expected input '{key}' not found in onnx_inputs")
+
+        embeddings = self.runner.run(onnx_inputs)[0]
+        return [embedding for embedding in embeddings]

frigate/embeddings/maintainer.py

@@ -12,6 +12,7 @@ import numpy as np
 from peewee import DoesNotExist
 from playhouse.sqliteq import SqliteQueueDatabase
 
+from frigate.comms.embeddings_updater import EmbeddingsRequestEnum, EmbeddingsResponder
 from frigate.comms.event_metadata_updater import (
     EventMetadataSubscriber,
     EventMetadataTypeEnum,
@@ -23,12 +24,15 @@ from frigate.const import CLIPS_DIR, UPDATE_EVENT_DESCRIPTION
 from frigate.events.types import EventTypeEnum
 from frigate.genai import get_genai_client
 from frigate.models import Event
+from frigate.util.builtin import serialize
 from frigate.util.image import SharedMemoryFrameManager, calculate_region
 
 from .embeddings import Embeddings
 
 logger = logging.getLogger(__name__)
 
+MAX_THUMBNAILS = 10
+
 
 class EmbeddingMaintainer(threading.Thread):
     """Handle embedding queue and post event updates."""
@@ -39,15 +43,20 @@ class EmbeddingMaintainer(threading.Thread):
         config: FrigateConfig,
         stop_event: MpEvent,
     ) -> None:
-        threading.Thread.__init__(self)
-        self.name = "embeddings_maintainer"
+        super().__init__(name="embeddings_maintainer")
         self.config = config
-        self.embeddings = Embeddings(db)
+        self.embeddings = Embeddings(config.semantic_search, db)
+
+        # Check if we need to re-index events
+        if config.semantic_search.reindex:
+            self.embeddings.reindex()
+
         self.event_subscriber = EventUpdateSubscriber()
         self.event_end_subscriber = EventEndSubscriber()
         self.event_metadata_subscriber = EventMetadataSubscriber(
             EventMetadataTypeEnum.regenerate_description
         )
+        self.embeddings_responder = EmbeddingsResponder()
         self.frame_manager = SharedMemoryFrameManager()
         # create communication for updating event descriptions
         self.requestor = InterProcessRequestor()
@@ -58,6 +67,7 @@ class EmbeddingMaintainer(threading.Thread):
     def run(self) -> None:
         """Maintain a SQLite-vec database for semantic search."""
         while not self.stop_event.is_set():
+            self._process_requests()
             self._process_updates()
             self._process_finalized()
             self._process_event_metadata()
@@ -65,12 +75,40 @@ class EmbeddingMaintainer(threading.Thread):
         self.event_subscriber.stop()
         self.event_end_subscriber.stop()
         self.event_metadata_subscriber.stop()
+        self.embeddings_responder.stop()
         self.requestor.stop()
         logger.info("Exiting embeddings maintenance...")
 
+    def _process_requests(self) -> None:
+        """Process embeddings requests"""
+
+        def _handle_request(topic: str, data: str) -> str:
+            try:
+                if topic == EmbeddingsRequestEnum.embed_description.value:
+                    return serialize(
+                        self.embeddings.embed_description(
+                            data["id"], data["description"]
+                        ),
+                        pack=False,
+                    )
+                elif topic == EmbeddingsRequestEnum.embed_thumbnail.value:
+                    thumbnail = base64.b64decode(data["thumbnail"])
+                    return serialize(
+                        self.embeddings.embed_thumbnail(data["id"], thumbnail),
+                        pack=False,
+                    )
+                elif topic == EmbeddingsRequestEnum.generate_search.value:
+                    return serialize(
+                        self.embeddings.text_embedding([data])[0], pack=False
+                    )
+            except Exception as e:
+                logger.error(f"Unable to handle embeddings request {e}")
+
+        self.embeddings_responder.check_for_request(_handle_request)
+
     def _process_updates(self) -> None:
         """Process event updates"""
-        update = self.event_subscriber.check_for_update()
+        update = self.event_subscriber.check_for_update(timeout=0.1)
 
         if update is None:
             return
@@ -81,6 +119,15 @@ class EmbeddingMaintainer(threading.Thread):
             return
 
         camera_config = self.config.cameras[camera]
+        # no need to save our own thumbnails if genai is not enabled
+        # or if the object has become stationary
+        if (
+            not camera_config.genai.enabled
+            or self.genai_client is None
+            or data["stationary"]
+        ):
+            return
+
         if data["id"] not in self.tracked_events:
             self.tracked_events[data["id"]] = []
 
@@ -91,7 +138,14 @@ class EmbeddingMaintainer(threading.Thread):
 
             if yuv_frame is not None:
                 data["thumbnail"] = self._create_thumbnail(yuv_frame, data["box"])
+
+                # Limit the number of thumbnails saved
+                if len(self.tracked_events[data["id"]]) >= MAX_THUMBNAILS:
+                    # Always keep the first thumbnail for the event
+                    self.tracked_events[data["id"]].pop(1)
+
                 self.tracked_events[data["id"]].append(data)
+
                 self.frame_manager.close(frame_id)
         except FileNotFoundError:
             pass
@@ -99,7 +153,7 @@ class EmbeddingMaintainer(threading.Thread):
     def _process_finalized(self) -> None:
         """Process the end of an event."""
         while True:
-            ended = self.event_end_subscriber.check_for_update()
+            ended = self.event_end_subscriber.check_for_update(timeout=0.1)
 
             if ended == None:
                 break
@@ -136,9 +190,6 @@ class EmbeddingMaintainer(threading.Thread):
                         or set(event.zones) & set(camera_config.genai.required_zones)
                     )
                 ):
-                    logger.debug(
-                        f"Description generation for {event}, has_snapshot: {event.has_snapshot}"
-                    )
                     if event.has_snapshot and camera_config.genai.use_snapshot:
                         with open(
                             os.path.join(CLIPS_DIR, f"{event.camera}-{event.id}.jpg"),
@@ -192,7 +243,7 @@ class EmbeddingMaintainer(threading.Thread):
     def _process_event_metadata(self):
         # Check for regenerate description requests
         (topic, event_id, source) = self.event_metadata_subscriber.check_for_update(
-            timeout=1
+            timeout=0.1
         )
 
         if topic is None:
@@ -219,14 +270,14 @@ class EmbeddingMaintainer(threading.Thread):
 
     def _embed_thumbnail(self, event_id: str, thumbnail: bytes) -> None:
         """Embed the thumbnail for an event."""
-        self.embeddings.upsert_thumbnail(event_id, thumbnail)
+        self.embeddings.embed_thumbnail(event_id, thumbnail)
 
     def _embed_description(self, event: Event, thumbnails: list[bytes]) -> None:
         """Embed the description for an event."""
         camera_config = self.config.cameras[event.camera]
 
         description = self.genai_client.generate_description(
-            camera_config, thumbnails, event.label
+            camera_config, thumbnails, event
         )
 
         if not description:
@@ -239,8 +290,8 @@ class EmbeddingMaintainer(threading.Thread):
             {"id": event.id, "description": description},
         )
 
-        # Encode the description
-        self.embeddings.upsert_description(event.id, description)
+        # Embed the description
+        self.embeddings.embed_description(event.id, description)
 
         logger.debug(
             "Generated description for %s (%d images): %s",

frigate/embeddings/util.py

@@ -20,10 +20,11 @@ class ZScoreNormalization:
 
     @property
     def stddev(self):
-        return math.sqrt(self.variance)
+        return math.sqrt(self.variance) if self.variance > 0 else 0.0
 
-    def normalize(self, distances: list[float]):
-        self._update(distances)
+    def normalize(self, distances: list[float], save_stats: bool):
+        if save_stats:
+            self._update(distances)
         if self.stddev == 0:
             return distances
         return [

frigate/events/audio.py

@@ -9,7 +9,6 @@ from typing import Tuple
 import numpy as np
 import requests
 
-import frigate.util as util
 from frigate.camera import CameraMetrics
 from frigate.comms.config_updater import ConfigSubscriber
 from frigate.comms.detections_updater import DetectionPublisher, DetectionTypeEnum
@@ -26,6 +25,7 @@ from frigate.const import (
 from frigate.ffmpeg_presets import parse_preset_input
 from frigate.log import LogPipe
 from frigate.object_detection import load_labels
+from frigate.service_manager import ServiceProcess
 from frigate.util.builtin import get_ffmpeg_arg_list
 from frigate.video import start_or_restart_ffmpeg, stop_ffmpeg
 
@@ -63,13 +63,15 @@ def get_ffmpeg_command(ffmpeg: FfmpegConfig) -> list[str]:
     )
 
 
-class AudioProcessor(util.Process):
+class AudioProcessor(ServiceProcess):
+    name = "frigate.audio_manager"
+
     def __init__(
         self,
         cameras: list[CameraConfig],
         camera_metrics: dict[str, CameraMetrics],
     ):
-        super().__init__(name="frigate.audio_manager", daemon=True)
+        super().__init__()
 
         self.camera_metrics = camera_metrics
         self.cameras = cameras

frigate/events/cleanup.py

@@ -8,11 +8,9 @@ from enum import Enum
 from multiprocessing.synchronize import Event as MpEvent
 from pathlib import Path
 
-from playhouse.sqliteq import SqliteQueueDatabase
-
 from frigate.config import FrigateConfig
 from frigate.const import CLIPS_DIR
-from frigate.embeddings.embeddings import Embeddings
+from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event, Timeline
 
 logger = logging.getLogger(__name__)
@@ -25,7 +23,7 @@ class EventCleanupType(str, Enum):
 
 class EventCleanup(threading.Thread):
     def __init__(
-        self, config: FrigateConfig, stop_event: MpEvent, db: SqliteQueueDatabase
+        self, config: FrigateConfig, stop_event: MpEvent, db: SqliteVecQueueDatabase
     ):
         super().__init__(name="event_cleanup")
         self.config = config
@@ -35,9 +33,6 @@ class EventCleanup(threading.Thread):
         self.removed_camera_labels: list[str] = None
         self.camera_labels: dict[str, dict[str, any]] = {}
 
-        if self.config.semantic_search.enabled:
-            self.embeddings = Embeddings(self.db)
-
     def get_removed_camera_labels(self) -> list[Event]:
         """Get a list of distinct labels for removed cameras."""
         if self.removed_camera_labels is None:
@@ -234,8 +229,8 @@ class EventCleanup(threading.Thread):
                     Event.delete().where(Event.id << chunk).execute()
 
                     if self.config.semantic_search.enabled:
-                        self.embeddings.delete_description(chunk)
-                        self.embeddings.delete_thumbnail(chunk)
+                        self.db.delete_embeddings_description(event_ids=chunk)
+                        self.db.delete_embeddings_thumbnail(event_ids=chunk)
                         logger.debug(f"Deleted {len(events_to_delete)} embeddings")
 
         logger.info("Exiting event cleanup...")

frigate/genai/__init__.py

@@ -4,7 +4,10 @@ import importlib
 import os
 from typing import Optional
 
+from playhouse.shortcuts import model_to_dict
+
 from frigate.config import CameraConfig, GenAIConfig, GenAIProviderEnum
+from frigate.models import Event
 
 PROVIDERS = {}
 
@@ -31,12 +34,13 @@ class GenAIClient:
         self,
         camera_config: CameraConfig,
         thumbnails: list[bytes],
-        label: str,
+        event: Event,
     ) -> Optional[str]:
         """Generate a description for the frame."""
         prompt = camera_config.genai.object_prompts.get(
-            label, camera_config.genai.prompt
-        )
+            event.label,
+            camera_config.genai.prompt,
+        ).format(**model_to_dict(event))
         return self._send(prompt, thumbnails)
 
     def _init_provider(self):

frigate/genai/ollama.py

@@ -21,12 +21,20 @@ class OllamaClient(GenAIClient):
 
     def _init_provider(self):
         """Initialize the client."""
-        client = ApiClient(host=self.genai_config.base_url, timeout=self.timeout)
-        response = client.pull(self.genai_config.model)
-        if response["status"] != "success":
-            logger.error("Failed to pull %s model from Ollama", self.genai_config.model)
+        try:
+            client = ApiClient(host=self.genai_config.base_url, timeout=self.timeout)
+            # ensure the model is available locally
+            response = client.show(self.genai_config.model)
+            if response.get("error"):
+                logger.error(
+                    "Ollama error: %s",
+                    response["error"],
+                )
+                return None
+            return client
+        except Exception as e:
+            logger.warning("Error initializing Ollama: %s", str(e))
             return None
-        return client
 
     def _send(self, prompt: str, images: list[bytes]) -> Optional[str]:
         """Submit a request to Ollama"""

frigate/models.py

@@ -93,7 +93,7 @@ class ReviewSegment(Model):  # type: ignore[misc]
     start_time = DateTimeField()
     end_time = DateTimeField()
     has_been_reviewed = BooleanField(default=False)
-    severity = CharField(max_length=30)  # alert, detection, significant_motion
+    severity = CharField(max_length=30)  # alert, detection
     thumb_path = CharField(unique=True)
     data = JSONField()  # additional data about detection like list of labels, zone, areas of significant motion
 

frigate/mypy.ini

@@ -59,3 +59,7 @@ ignore_errors = false
 [mypy-frigate.watchdog]
 ignore_errors = false
 disallow_untyped_calls = false
+
+
+[mypy-frigate.service_manager.*]
+ignore_errors = false

frigate/object_processing.py

@@ -1,4 +1,3 @@
-import base64
 import datetime
 import json
 import logging
@@ -7,7 +6,6 @@ import queue
 import threading
 from collections import Counter, defaultdict
 from multiprocessing.synchronize import Event as MpEvent
-from statistics import median
 from typing import Callable
 
 import cv2
@@ -18,7 +16,6 @@ from frigate.comms.dispatcher import Dispatcher
 from frigate.comms.events_updater import EventEndSubscriber, EventUpdatePublisher
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import (
-    CameraConfig,
     FrigateConfig,
     MqttConfig,
     RecordConfig,
@@ -28,458 +25,18 @@ from frigate.config import (
 from frigate.const import CLIPS_DIR, UPDATE_CAMERA_ACTIVITY
 from frigate.events.types import EventStateEnum, EventTypeEnum
 from frigate.ptz.autotrack import PtzAutoTrackerThread
+from frigate.track.tracked_object import TrackedObject
 from frigate.util.image import (
     SharedMemoryFrameManager,
-    area,
-    calculate_region,
     draw_box_with_label,
     draw_timestamp,
+    is_better_thumbnail,
     is_label_printable,
 )
 
 logger = logging.getLogger(__name__)
 
 
-def on_edge(box, frame_shape):
-    if (
-        box[0] == 0
-        or box[1] == 0
-        or box[2] == frame_shape[1] - 1
-        or box[3] == frame_shape[0] - 1
-    ):
-        return True
-
-
-def has_better_attr(current_thumb, new_obj, attr_label) -> bool:
-    max_new_attr = max(
-        [0]
-        + [area(a["box"]) for a in new_obj["attributes"] if a["label"] == attr_label]
-    )
-    max_current_attr = max(
-        [0]
-        + [
-            area(a["box"])
-            for a in current_thumb["attributes"]
-            if a["label"] == attr_label
-        ]
-    )
-
-    # if the thumb has a higher scoring attr
-    return max_new_attr > max_current_attr
-
-
-def is_better_thumbnail(label, current_thumb, new_obj, frame_shape) -> bool:
-    # larger is better
-    # cutoff images are less ideal, but they should also be smaller?
-    # better scores are obviously better too
-
-    # check face on person
-    if label == "person":
-        if has_better_attr(current_thumb, new_obj, "face"):
-            return True
-        # if the current thumb has a face attr, dont update unless it gets better
-        if any([a["label"] == "face" for a in current_thumb["attributes"]]):
-            return False
-
-    # check license_plate on car
-    if label == "car":
-        if has_better_attr(current_thumb, new_obj, "license_plate"):
-            return True
-        # if the current thumb has a license_plate attr, dont update unless it gets better
-        if any([a["label"] == "license_plate" for a in current_thumb["attributes"]]):
-            return False
-
-    # if the new_thumb is on an edge, and the current thumb is not
-    if on_edge(new_obj["box"], frame_shape) and not on_edge(
-        current_thumb["box"], frame_shape
-    ):
-        return False
-
-    # if the score is better by more than 5%
-    if new_obj["score"] > current_thumb["score"] + 0.05:
-        return True
-
-    # if the area is 10% larger
-    if new_obj["area"] > current_thumb["area"] * 1.1:
-        return True
-
-    return False
-
-
-class TrackedObject:
-    def __init__(
-        self,
-        camera,
-        colormap,
-        camera_config: CameraConfig,
-        frame_cache,
-        obj_data: dict[str, any],
-    ):
-        # set the score history then remove as it is not part of object state
-        self.score_history = obj_data["score_history"]
-        del obj_data["score_history"]
-
-        self.obj_data = obj_data
-        self.camera = camera
-        self.colormap = colormap
-        self.camera_config = camera_config
-        self.frame_cache = frame_cache
-        self.zone_presence: dict[str, int] = {}
-        self.zone_loitering: dict[str, int] = {}
-        self.current_zones = []
-        self.entered_zones = []
-        self.attributes = defaultdict(float)
-        self.false_positive = True
-        self.has_clip = False
-        self.has_snapshot = False
-        self.top_score = self.computed_score = 0.0
-        self.thumbnail_data = None
-        self.last_updated = 0
-        self.last_published = 0
-        self.frame = None
-        self.active = True
-        self.previous = self.to_dict()
-
-    def _is_false_positive(self):
-        # once a true positive, always a true positive
-        if not self.false_positive:
-            return False
-
-        threshold = self.camera_config.objects.filters[self.obj_data["label"]].threshold
-        return self.computed_score < threshold
-
-    def compute_score(self):
-        """get median of scores for object."""
-        return median(self.score_history)
-
-    def update(self, current_frame_time: float, obj_data, has_valid_frame: bool):
-        thumb_update = False
-        significant_change = False
-        autotracker_update = False
-        # if the object is not in the current frame, add a 0.0 to the score history
-        if obj_data["frame_time"] != current_frame_time:
-            self.score_history.append(0.0)
-        else:
-            self.score_history.append(obj_data["score"])
-
-        # only keep the last 10 scores
-        if len(self.score_history) > 10:
-            self.score_history = self.score_history[-10:]
-
-        # calculate if this is a false positive
-        self.computed_score = self.compute_score()
-        if self.computed_score > self.top_score:
-            self.top_score = self.computed_score
-        self.false_positive = self._is_false_positive()
-        self.active = self.is_active()
-
-        if not self.false_positive and has_valid_frame:
-            # determine if this frame is a better thumbnail
-            if self.thumbnail_data is None or is_better_thumbnail(
-                self.obj_data["label"],
-                self.thumbnail_data,
-                obj_data,
-                self.camera_config.frame_shape,
-            ):
-                self.thumbnail_data = {
-                    "frame_time": current_frame_time,
-                    "box": obj_data["box"],
-                    "area": obj_data["area"],
-                    "region": obj_data["region"],
-                    "score": obj_data["score"],
-                    "attributes": obj_data["attributes"],
-                }
-                thumb_update = True
-
-        # check zones
-        current_zones = []
-        bottom_center = (obj_data["centroid"][0], obj_data["box"][3])
-        # check each zone
-        for name, zone in self.camera_config.zones.items():
-            # if the zone is not for this object type, skip
-            if len(zone.objects) > 0 and obj_data["label"] not in zone.objects:
-                continue
-            contour = zone.contour
-            zone_score = self.zone_presence.get(name, 0) + 1
-            # check if the object is in the zone
-            if cv2.pointPolygonTest(contour, bottom_center, False) >= 0:
-                # if the object passed the filters once, dont apply again
-                if name in self.current_zones or not zone_filtered(self, zone.filters):
-                    # an object is only considered present in a zone if it has a zone inertia of 3+
-                    if zone_score >= zone.inertia:
-                        loitering_score = self.zone_loitering.get(name, 0) + 1
-
-                        # loitering time is configured as seconds, convert to count of frames
-                        if loitering_score >= (
-                            self.camera_config.zones[name].loitering_time
-                            * self.camera_config.detect.fps
-                        ):
-                            current_zones.append(name)
-
-                            if name not in self.entered_zones:
-                                self.entered_zones.append(name)
-                        else:
-                            self.zone_loitering[name] = loitering_score
-                    else:
-                        self.zone_presence[name] = zone_score
-            else:
-                # once an object has a zone inertia of 3+ it is not checked anymore
-                if 0 < zone_score < zone.inertia:
-                    self.zone_presence[name] = zone_score - 1
-
-        # maintain attributes
-        for attr in obj_data["attributes"]:
-            if self.attributes[attr["label"]] < attr["score"]:
-                self.attributes[attr["label"]] = attr["score"]
-
-        # populate the sub_label for object with highest scoring logo
-        if self.obj_data["label"] in ["car", "package", "person"]:
-            recognized_logos = {
-                k: self.attributes[k]
-                for k in ["ups", "fedex", "amazon"]
-                if k in self.attributes
-            }
-            if len(recognized_logos) > 0:
-                max_logo = max(recognized_logos, key=recognized_logos.get)
-
-                # don't overwrite sub label if it is already set
-                if (
-                    self.obj_data.get("sub_label") is None
-                    or self.obj_data["sub_label"][0] == max_logo
-                ):
-                    self.obj_data["sub_label"] = (max_logo, recognized_logos[max_logo])
-
-        # check for significant change
-        if not self.false_positive:
-            # if the zones changed, signal an update
-            if set(self.current_zones) != set(current_zones):
-                significant_change = True
-
-            # if the position changed, signal an update
-            if self.obj_data["position_changes"] != obj_data["position_changes"]:
-                significant_change = True
-
-            if self.obj_data["attributes"] != obj_data["attributes"]:
-                significant_change = True
-
-            # if the state changed between stationary and active
-            if self.previous["active"] != self.active:
-                significant_change = True
-
-            # update at least once per minute
-            if self.obj_data["frame_time"] - self.previous["frame_time"] > 60:
-                significant_change = True
-
-            # update autotrack at most 3 objects per second
-            if self.obj_data["frame_time"] - self.previous["frame_time"] >= (1 / 3):
-                autotracker_update = True
-
-        self.obj_data.update(obj_data)
-        self.current_zones = current_zones
-        return (thumb_update, significant_change, autotracker_update)
-
-    def to_dict(self, include_thumbnail: bool = False):
-        event = {
-            "id": self.obj_data["id"],
-            "camera": self.camera,
-            "frame_time": self.obj_data["frame_time"],
-            "snapshot": self.thumbnail_data,
-            "label": self.obj_data["label"],
-            "sub_label": self.obj_data.get("sub_label"),
-            "top_score": self.top_score,
-            "false_positive": self.false_positive,
-            "start_time": self.obj_data["start_time"],
-            "end_time": self.obj_data.get("end_time", None),
-            "score": self.obj_data["score"],
-            "box": self.obj_data["box"],
-            "area": self.obj_data["area"],
-            "ratio": self.obj_data["ratio"],
-            "region": self.obj_data["region"],
-            "active": self.active,
-            "stationary": not self.active,
-            "motionless_count": self.obj_data["motionless_count"],
-            "position_changes": self.obj_data["position_changes"],
-            "current_zones": self.current_zones.copy(),
-            "entered_zones": self.entered_zones.copy(),
-            "has_clip": self.has_clip,
-            "has_snapshot": self.has_snapshot,
-            "attributes": self.attributes,
-            "current_attributes": self.obj_data["attributes"],
-        }
-
-        if include_thumbnail:
-            event["thumbnail"] = base64.b64encode(self.get_thumbnail()).decode("utf-8")
-
-        return event
-
-    def is_active(self):
-        return not self.is_stationary()
-
-    def is_stationary(self):
-        return (
-            self.obj_data["motionless_count"]
-            > self.camera_config.detect.stationary.threshold
-        )
-
-    def get_thumbnail(self):
-        if (
-            self.thumbnail_data is None
-            or self.thumbnail_data["frame_time"] not in self.frame_cache
-        ):
-            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
-
-        jpg_bytes = self.get_jpg_bytes(
-            timestamp=False, bounding_box=False, crop=True, height=175
-        )
-
-        if jpg_bytes:
-            return jpg_bytes
-        else:
-            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
-            return jpg.tobytes()
-
-    def get_clean_png(self):
-        if self.thumbnail_data is None:
-            return None
-
-        try:
-            best_frame = cv2.cvtColor(
-                self.frame_cache[self.thumbnail_data["frame_time"]],
-                cv2.COLOR_YUV2BGR_I420,
-            )
-        except KeyError:
-            logger.warning(
-                f"Unable to create clean png because frame {self.thumbnail_data['frame_time']} is not in the cache"
-            )
-            return None
-
-        ret, png = cv2.imencode(".png", best_frame)
-        if ret:
-            return png.tobytes()
-        else:
-            return None
-
-    def get_jpg_bytes(
-        self, timestamp=False, bounding_box=False, crop=False, height=None, quality=70
-    ):
-        if self.thumbnail_data is None:
-            return None
-
-        try:
-            best_frame = cv2.cvtColor(
-                self.frame_cache[self.thumbnail_data["frame_time"]],
-                cv2.COLOR_YUV2BGR_I420,
-            )
-        except KeyError:
-            logger.warning(
-                f"Unable to create jpg because frame {self.thumbnail_data['frame_time']} is not in the cache"
-            )
-            return None
-
-        if bounding_box:
-            thickness = 2
-            color = self.colormap[self.obj_data["label"]]
-
-            # draw the bounding boxes on the frame
-            box = self.thumbnail_data["box"]
-            draw_box_with_label(
-                best_frame,
-                box[0],
-                box[1],
-                box[2],
-                box[3],
-                self.obj_data["label"],
-                f"{int(self.thumbnail_data['score']*100)}% {int(self.thumbnail_data['area'])}",
-                thickness=thickness,
-                color=color,
-            )
-
-            # draw any attributes
-            for attribute in self.thumbnail_data["attributes"]:
-                box = attribute["box"]
-                draw_box_with_label(
-                    best_frame,
-                    box[0],
-                    box[1],
-                    box[2],
-                    box[3],
-                    attribute["label"],
-                    f"{attribute['score']:.0%}",
-                    thickness=thickness,
-                    color=color,
-                )
-
-        if crop:
-            box = self.thumbnail_data["box"]
-            box_size = 300
-            region = calculate_region(
-                best_frame.shape,
-                box[0],
-                box[1],
-                box[2],
-                box[3],
-                box_size,
-                multiplier=1.1,
-            )
-            best_frame = best_frame[region[1] : region[3], region[0] : region[2]]
-
-        if height:
-            width = int(height * best_frame.shape[1] / best_frame.shape[0])
-            best_frame = cv2.resize(
-                best_frame, dsize=(width, height), interpolation=cv2.INTER_AREA
-            )
-        if timestamp:
-            color = self.camera_config.timestamp_style.color
-            draw_timestamp(
-                best_frame,
-                self.thumbnail_data["frame_time"],
-                self.camera_config.timestamp_style.format,
-                font_effect=self.camera_config.timestamp_style.effect,
-                font_thickness=self.camera_config.timestamp_style.thickness,
-                font_color=(color.blue, color.green, color.red),
-                position=self.camera_config.timestamp_style.position,
-            )
-
-        ret, jpg = cv2.imencode(
-            ".jpg", best_frame, [int(cv2.IMWRITE_JPEG_QUALITY), quality]
-        )
-        if ret:
-            return jpg.tobytes()
-        else:
-            return None
-
-
-def zone_filtered(obj: TrackedObject, object_config):
-    object_name = obj.obj_data["label"]
-
-    if object_name in object_config:
-        obj_settings = object_config[object_name]
-
-        # if the min area is larger than the
-        # detected object, don't add it to detected objects
-        if obj_settings.min_area > obj.obj_data["area"]:
-            return True
-
-        # if the detected object is larger than the
-        # max area, don't add it to detected objects
-        if obj_settings.max_area < obj.obj_data["area"]:
-            return True
-
-        # if the score is lower than the threshold, skip
-        if obj_settings.threshold > obj.computed_score:
-            return True
-
-        # if the object is not proportionally wide enough
-        if obj_settings.min_ratio > obj.obj_data["ratio"]:
-            return True
-
-        # if the object is proportionally too wide
-        if obj_settings.max_ratio < obj.obj_data["ratio"]:
-            return True
-
-    return False
-
-
 # Maintains the state of a camera
 class CameraState:
     def __init__(
@@ -696,8 +253,7 @@ class CameraState:
 
         for id in new_ids:
             new_obj = tracked_objects[id] = TrackedObject(
-                self.name,
-                self.config.model.colormap,
+                self.config.model,
                 self.camera_config,
                 self.frame_cache,
                 current_detections[id],
@@ -788,6 +344,7 @@ class CameraState:
             # if the object's thumbnail is not from the current frame, skip
             if (
                 current_frame is None
+                or obj.thumbnail_data is None
                 or obj.false_positive
                 or obj.thumbnail_data["frame_time"] != frame_time
             ):

frigate/ptz/autotrack.py

@@ -32,6 +32,7 @@ from frigate.const import (
     CONFIG_DIR,
 )
 from frigate.ptz.onvif import OnvifController
+from frigate.track.tracked_object import TrackedObject
 from frigate.util.builtin import update_yaml_file
 from frigate.util.image import SharedMemoryFrameManager, intersection_over_union
 
@@ -214,7 +215,7 @@ class PtzAutoTracker:
             ):
                 self._autotracker_setup(camera_config, camera)
 
-    def _autotracker_setup(self, camera_config, camera):
+    def _autotracker_setup(self, camera_config: CameraConfig, camera: str):
         logger.debug(f"{camera}: Autotracker init")
 
         self.object_types[camera] = camera_config.onvif.autotracking.track
@@ -852,7 +853,7 @@ class PtzAutoTracker:
             logger.debug(f"{camera}: Valid velocity ")
             return True, velocities.flatten()
 
-    def _get_distance_threshold(self, camera, obj):
+    def _get_distance_threshold(self, camera: str, obj: TrackedObject):
         # Returns true if Euclidean distance from object to center of frame is
         # less than 10% of the of the larger dimension (width or height) of the frame,
         # multiplied by a scaling factor for object size.
@@ -888,7 +889,9 @@ class PtzAutoTracker:
 
         return distance_threshold
 
-    def _should_zoom_in(self, camera, obj, box, predicted_time, debug_zooming=False):
+    def _should_zoom_in(
+        self, camera: str, obj: TrackedObject, box, predicted_time, debug_zooming=False
+    ):
         # returns True if we should zoom in, False if we should zoom out, None to do nothing
         camera_config = self.config.cameras[camera]
         camera_width = camera_config.frame_shape[1]
@@ -1019,7 +1022,7 @@ class PtzAutoTracker:
         # Don't zoom at all
         return None
 
-    def _autotrack_move_ptz(self, camera, obj):
+    def _autotrack_move_ptz(self, camera: str, obj: TrackedObject):
         camera_config = self.config.cameras[camera]
         camera_width = camera_config.frame_shape[1]
         camera_height = camera_config.frame_shape[0]
@@ -1090,7 +1093,12 @@ class PtzAutoTracker:
                 self._enqueue_move(camera, obj.obj_data["frame_time"], 0, 0, zoom)
 
     def _get_zoom_amount(
-        self, camera, obj, predicted_box, predicted_movement_time, debug_zoom=True
+        self,
+        camera: str,
+        obj: TrackedObject,
+        predicted_box,
+        predicted_movement_time,
+        debug_zoom=True,
     ):
         camera_config = self.config.cameras[camera]
 
@@ -1186,13 +1194,13 @@ class PtzAutoTracker:
 
         return zoom
 
-    def is_autotracking(self, camera):
+    def is_autotracking(self, camera: str):
         return self.tracked_object[camera] is not None
 
-    def autotracked_object_region(self, camera):
+    def autotracked_object_region(self, camera: str):
         return self.tracked_object[camera]["region"]
 
-    def autotrack_object(self, camera, obj):
+    def autotrack_object(self, camera: str, obj: TrackedObject):
         camera_config = self.config.cameras[camera]
 
         if camera_config.onvif.autotracking.enabled:
@@ -1208,7 +1216,7 @@ class PtzAutoTracker:
             if (
                 # new object
                 self.tracked_object[camera] is None
-                and obj.camera == camera
+                and obj.camera_config.name == camera
                 and obj.obj_data["label"] in self.object_types[camera]
                 and set(obj.entered_zones) & set(self.required_zones[camera])
                 and not obj.previous["false_positive"]
@@ -1267,7 +1275,7 @@ class PtzAutoTracker:
                 # If it's within bounds, start tracking that object.
                 # Should we check region (maybe too broad) or expand the previous object's box a bit and check that?
                 self.tracked_object[camera] is None
-                and obj.camera == camera
+                and obj.camera_config.name == camera
                 and obj.obj_data["label"] in self.object_types[camera]
                 and not obj.previous["false_positive"]
                 and not obj.false_positive

frigate/record/export.py

@@ -43,6 +43,11 @@ class PlaybackFactorEnum(str, Enum):
     timelapse_25x = "timelapse_25x"
 
 
+class PlaybackSourceEnum(str, Enum):
+    recordings = "recordings"
+    preview = "preview"
+
+
 class RecordingExporter(threading.Thread):
     """Exports a specific set of recordings for a camera to storage as a single file."""
 
@@ -56,6 +61,7 @@ class RecordingExporter(threading.Thread):
         start_time: int,
         end_time: int,
         playback_factor: PlaybackFactorEnum,
+        playback_source: PlaybackSourceEnum,
     ) -> None:
         super().__init__()
         self.config = config
@@ -66,6 +72,7 @@ class RecordingExporter(threading.Thread):
         self.start_time = start_time
         self.end_time = end_time
         self.playback_factor = playback_factor
+        self.playback_source = playback_source
 
         # ensure export thumb dir
         Path(os.path.join(CLIPS_DIR, "export")).mkdir(exist_ok=True)
@@ -170,30 +177,7 @@ class RecordingExporter(threading.Thread):
 
         return thumb_path
 
-    def run(self) -> None:
-        logger.debug(
-            f"Beginning export for {self.camera} from {self.start_time} to {self.end_time}"
-        )
-        export_name = (
-            self.user_provided_name
-            or f"{self.camera.replace('_', ' ')} {self.get_datetime_from_timestamp(self.start_time)} {self.get_datetime_from_timestamp(self.end_time)}"
-        )
-        video_path = f"{EXPORT_DIR}/{self.export_id}.mp4"
-
-        thumb_path = self.save_thumbnail(self.export_id)
-
-        Export.insert(
-            {
-                Export.id: self.export_id,
-                Export.camera: self.camera,
-                Export.name: export_name,
-                Export.date: self.start_time,
-                Export.video_path: video_path,
-                Export.thumb_path: thumb_path,
-                Export.in_progress: True,
-            }
-        ).execute()
-
+    def get_record_export_command(self, video_path: str) -> list[str]:
         if (self.end_time - self.start_time) <= MAX_PLAYLIST_SECONDS:
             playlist_lines = f"http://127.0.0.1:5000/vod/{self.camera}/start/{self.start_time}/end/{self.end_time}/index.m3u8"
             ffmpeg_input = (
@@ -204,7 +188,10 @@ class RecordingExporter(threading.Thread):
 
             # get full set of recordings
             export_recordings = (
-                Recordings.select()
+                Recordings.select(
+                    Recordings.start_time,
+                    Recordings.end_time,
+                )
                 .where(
                     Recordings.start_time.between(self.start_time, self.end_time)
                     | Recordings.end_time.between(self.start_time, self.end_time)
@@ -229,6 +216,65 @@ class RecordingExporter(threading.Thread):
 
             ffmpeg_input = "-y -protocol_whitelist pipe,file,http,tcp -f concat -safe 0 -i /dev/stdin"
 
+        if self.playback_factor == PlaybackFactorEnum.realtime:
+            ffmpeg_cmd = (
+                f"{self.config.ffmpeg.ffmpeg_path} -hide_banner {ffmpeg_input} -c copy -movflags +faststart {video_path}"
+            ).split(" ")
+        elif self.playback_factor == PlaybackFactorEnum.timelapse_25x:
+            ffmpeg_cmd = (
+                parse_preset_hardware_acceleration_encode(
+                    self.config.ffmpeg.ffmpeg_path,
+                    self.config.ffmpeg.hwaccel_args,
+                    f"-an {ffmpeg_input}",
+                    f"{self.config.cameras[self.camera].record.export.timelapse_args} -movflags +faststart {video_path}",
+                    EncodeTypeEnum.timelapse,
+                )
+            ).split(" ")
+
+        return ffmpeg_cmd, playlist_lines
+
+    def get_preview_export_command(self, video_path: str) -> list[str]:
+        playlist_lines = []
+
+        # get full set of previews
+        export_previews = (
+            Previews.select(
+                Previews.path,
+                Previews.start_time,
+                Previews.end_time,
+            )
+            .where(
+                Previews.start_time.between(self.start_time, self.end_time)
+                | Previews.end_time.between(self.start_time, self.end_time)
+                | (
+                    (self.start_time > Previews.start_time)
+                    & (self.end_time < Previews.end_time)
+                )
+            )
+            .where(Previews.camera == self.camera)
+            .order_by(Previews.start_time.asc())
+            .namedtuples()
+            .iterator()
+        )
+
+        preview: Previews
+        for preview in export_previews:
+            playlist_lines.append(f"file '{preview.path}'")
+
+            if preview.start_time < self.start_time:
+                playlist_lines.append(
+                    f"inpoint {int(self.start_time - preview.start_time)}"
+                )
+
+            if preview.end_time > self.end_time:
+                playlist_lines.append(
+                    f"outpoint {int(preview.end_time - self.end_time)}"
+                )
+
+        ffmpeg_input = (
+            "-y -protocol_whitelist pipe,file,tcp -f concat -safe 0 -i /dev/stdin"
+        )
+
         if self.playback_factor == PlaybackFactorEnum.realtime:
             ffmpeg_cmd = (
                 f"{self.config.ffmpeg.ffmpeg_path} -hide_banner {ffmpeg_input} -c copy -movflags +faststart {video_path}"
@@ -244,6 +290,36 @@ class RecordingExporter(threading.Thread):
                 )
             ).split(" ")
 
+        return ffmpeg_cmd, playlist_lines
+
+    def run(self) -> None:
+        logger.debug(
+            f"Beginning export for {self.camera} from {self.start_time} to {self.end_time}"
+        )
+        export_name = (
+            self.user_provided_name
+            or f"{self.camera.replace('_', ' ')} {self.get_datetime_from_timestamp(self.start_time)} {self.get_datetime_from_timestamp(self.end_time)}"
+        )
+        video_path = f"{EXPORT_DIR}/{self.export_id}.mp4"
+        thumb_path = self.save_thumbnail(self.export_id)
+
+        Export.insert(
+            {
+                Export.id: self.export_id,
+                Export.camera: self.camera,
+                Export.name: export_name,
+                Export.date: self.start_time,
+                Export.video_path: video_path,
+                Export.thumb_path: thumb_path,
+                Export.in_progress: True,
+            }
+        ).execute()
+
+        if self.playback_source == PlaybackSourceEnum.recordings:
+            ffmpeg_cmd, playlist_lines = self.get_record_export_command(video_path)
+        else:
+            ffmpeg_cmd, playlist_lines = self.get_preview_export_command(video_path)
+
         p = sp.run(
             ffmpeg_cmd,
             input="\n".join(playlist_lines),
@@ -254,7 +330,7 @@ class RecordingExporter(threading.Thread):
 
         if p.returncode != 0:
             logger.error(
-                f"Failed to export recording for command {' '.join(ffmpeg_cmd)}"
+                f"Failed to export {self.playback_source.value} for command {' '.join(ffmpeg_cmd)}"
             )
             logger.error(p.stderr)
             Path(video_path).unlink(missing_ok=True)

frigate/record/maintainer.py

@@ -142,6 +142,8 @@ class RecordingMaintainer(threading.Thread):
                     )
                 )
             )
+
+            # see if the recording mover is too slow and segments need to be deleted
             if processed_segment_count > keep_count:
                 logger.warning(
                     f"Unable to keep up with recording segments in cache for {camera}. Keeping the {keep_count} most recent segments out of {processed_segment_count} and discarding the rest..."
@@ -153,6 +155,21 @@ class RecordingMaintainer(threading.Thread):
                     self.end_time_cache.pop(cache_path, None)
                 grouped_recordings[camera] = grouped_recordings[camera][-keep_count:]
 
+            # see if detection has failed and unprocessed segments need to be deleted
+            unprocessed_segment_count = (
+                len(grouped_recordings[camera]) - processed_segment_count
+            )
+            if unprocessed_segment_count > keep_count:
+                logger.warning(
+                    f"Too many unprocessed recording segments in cache for {camera}. This likely indicates an issue with the detect stream, keeping the {keep_count} most recent segments out of {unprocessed_segment_count} and discarding the rest..."
+                )
+                to_remove = grouped_recordings[camera][:-keep_count]
+                for rec in to_remove:
+                    cache_path = rec["cache_path"]
+                    Path(cache_path).unlink(missing_ok=True)
+                    self.end_time_cache.pop(cache_path, None)
+                grouped_recordings[camera] = grouped_recordings[camera][-keep_count:]
+
         tasks = []
         for camera, recordings in grouped_recordings.items():
             # clear out all the object recording info for old frames

frigate/review/maintainer.py

@@ -51,7 +51,7 @@ class PendingReviewSegment:
         frame_time: float,
         severity: SeverityEnum,
         detections: dict[str, str],
-        sub_labels: set[str],
+        sub_labels: dict[str, str],
         zones: list[str],
         audio: set[str],
     ):
@@ -135,7 +135,7 @@ class PendingReviewSegment:
             ReviewSegment.data.name: {
                 "detections": list(set(self.detections.keys())),
                 "objects": list(set(self.detections.values())),
-                "sub_labels": list(self.sub_labels),
+                "sub_labels": list(self.sub_labels.values()),
                 "zones": self.zones,
                 "audio": list(self.audio),
             },
@@ -167,7 +167,7 @@ class ReviewSegmentMaintainer(threading.Thread):
         # clear ongoing review segments from last instance
         self.requestor.send_data(CLEAR_ONGOING_REVIEW_SEGMENTS, "")
 
-    def new_segment(
+    def _publish_segment_start(
         self,
         segment: PendingReviewSegment,
     ) -> None:
@@ -186,7 +186,7 @@ class ReviewSegmentMaintainer(threading.Thread):
             ),
         )
 
-    def update_segment(
+    def _publish_segment_update(
         self,
         segment: PendingReviewSegment,
         camera_config: CameraConfig,
@@ -211,7 +211,7 @@ class ReviewSegmentMaintainer(threading.Thread):
             ),
         )
 
-    def end_segment(
+    def _publish_segment_end(
         self,
         segment: PendingReviewSegment,
         prev_data: dict[str, any],
@@ -239,10 +239,16 @@ class ReviewSegmentMaintainer(threading.Thread):
     ) -> None:
         """Validate if existing review segment should continue."""
         camera_config = self.config.cameras[segment.camera]
-        active_objects = get_active_objects(frame_time, camera_config, objects)
+
+        # get active objects + objects loitering in loitering zones
+        active_objects = get_active_objects(
+            frame_time, camera_config, objects
+        ) + get_loitering_objects(frame_time, camera_config, objects)
         prev_data = segment.get_data(False)
+        has_activity = False
 
         if len(active_objects) > 0:
+            has_activity = True
             should_update = False
 
             if frame_time > segment.last_update:
@@ -255,7 +261,7 @@ class ReviewSegmentMaintainer(threading.Thread):
                     segment.detections[object["id"]] = object["sub_label"][0]
                 else:
                     segment.detections[object["id"]] = f'{object["label"]}-verified'
-                    segment.sub_labels.add(object["sub_label"][0])
+                    segment.sub_labels[object["id"]] = object["sub_label"][0]
 
                 # if object is alert label
                 # and has entered required zones or required zones is not set
@@ -295,13 +301,14 @@ class ReviewSegmentMaintainer(threading.Thread):
                         logger.debug(f"Failed to get frame {frame_id} from SHM")
                         return
 
-                    self.update_segment(
+                    self._publish_segment_update(
                         segment, camera_config, yuv_frame, active_objects, prev_data
                     )
                     self.frame_manager.close(frame_id)
                 except FileNotFoundError:
                     return
-        else:
+
+        if not has_activity:
             if not segment.has_frame:
                 try:
                     frame_id = f"{camera_config.name}{frame_time}"
@@ -315,16 +322,18 @@ class ReviewSegmentMaintainer(threading.Thread):
 
                     segment.save_full_frame(camera_config, yuv_frame)
                     self.frame_manager.close(frame_id)
-                    self.update_segment(segment, camera_config, None, [], prev_data)
+                    self._publish_segment_update(
+                        segment, camera_config, None, [], prev_data
+                    )
                 except FileNotFoundError:
                     return
 
             if segment.severity == SeverityEnum.alert and frame_time > (
                 segment.last_update + THRESHOLD_ALERT_ACTIVITY
             ):
-                self.end_segment(segment, prev_data)
+                self._publish_segment_end(segment, prev_data)
             elif frame_time > (segment.last_update + THRESHOLD_DETECTION_ACTIVITY):
-                self.end_segment(segment, prev_data)
+                self._publish_segment_end(segment, prev_data)
 
     def check_if_new_segment(
         self,
@@ -338,7 +347,7 @@ class ReviewSegmentMaintainer(threading.Thread):
 
         if len(active_objects) > 0:
             detections: dict[str, str] = {}
-            sub_labels = set()
+            sub_labels: dict[str, str] = {}
             zones: list[str] = []
             severity = None
 
@@ -349,7 +358,7 @@ class ReviewSegmentMaintainer(threading.Thread):
                     detections[object["id"]] = object["sub_label"][0]
                 else:
                     detections[object["id"]] = f'{object["label"]}-verified'
-                    sub_labels.add(object["sub_label"][0])
+                    sub_labels[object["id"]] = object["sub_label"][0]
 
                 # if object is alert label
                 # and has entered required zones or required zones is not set
@@ -418,7 +427,7 @@ class ReviewSegmentMaintainer(threading.Thread):
                         camera_config, yuv_frame, active_objects
                     )
                     self.frame_manager.close(frame_id)
-                    self.new_segment(self.active_review_segments[camera])
+                    self._publish_segment_start(self.active_review_segments[camera])
                 except FileNotFoundError:
                     return
 
@@ -557,7 +566,7 @@ class ReviewSegmentMaintainer(threading.Thread):
                             frame_time,
                             severity,
                             {},
-                            set(),
+                            {},
                             [],
                             detections,
                         )
@@ -567,7 +576,7 @@ class ReviewSegmentMaintainer(threading.Thread):
                         frame_time,
                         SeverityEnum.alert,
                         {manual_info["event_id"]: manual_info["label"]},
-                        set(),
+                        {},
                         [],
                         set(),
                     )
@@ -609,3 +618,24 @@ def get_active_objects(
             )
         )  # object must be in the alerts or detections label list
     ]
+
+
+def get_loitering_objects(
+    frame_time: float, camera_config: CameraConfig, all_objects: list[TrackedObject]
+) -> list[TrackedObject]:
+    """get loitering objects for detection."""
+    return [
+        o
+        for o in all_objects
+        if o["pending_loitering"]  # object must be pending loitering
+        and o["position_changes"] > 0  # object must have moved at least once
+        and o["frame_time"] == frame_time  # object must be detected in this frame
+        and not o["false_positive"]  # object must not be a false positive
+        and (
+            o["label"] in camera_config.review.alerts.labels
+            or (
+                camera_config.review.detections.labels is None
+                or o["label"] in camera_config.review.detections.labels
+            )
+        )  # object must be in the alerts or detections label list
+    ]

frigate/service_manager/__init__.py

@@ -0,0 +1,4 @@
+from .multiprocessing import ServiceProcess
+from .service import Service, ServiceManager
+
+__all__ = ["Service", "ServiceProcess", "ServiceManager"]

frigate/service_manager/multiprocessing.py

@@ -0,0 +1,164 @@
+import asyncio
+import faulthandler
+import logging
+import multiprocessing as mp
+import signal
+import sys
+import threading
+from abc import ABC, abstractmethod
+from asyncio.exceptions import TimeoutError
+from logging.handlers import QueueHandler
+from types import FrameType
+from typing import Optional
+
+import frigate.log
+
+from .multiprocessing_waiter import wait as mp_wait
+from .service import Service, ServiceManager
+
+DEFAULT_STOP_TIMEOUT = 10  # seconds
+
+
+class BaseServiceProcess(Service, ABC):
+    """A Service the manages a multiprocessing.Process."""
+
+    _process: Optional[mp.Process]
+
+    def __init__(
+        self,
+        *,
+        name: Optional[str] = None,
+        manager: Optional[ServiceManager] = None,
+    ) -> None:
+        super().__init__(name=name, manager=manager)
+
+        self._process = None
+
+    async def on_start(self) -> None:
+        if self._process is not None:
+            if self._process.is_alive():
+                return  # Already started.
+            else:
+                self._process.close()
+
+        # At this point, the process is either stopped or dead, so we can recreate it.
+        self._process = mp.Process(target=self._run)
+        self._process.name = self.name
+        self._process.daemon = True
+        self.before_start()
+        self._process.start()
+        self.after_start()
+
+        self.manager.logger.info(f"Started {self.name} (pid: {self._process.pid})")
+
+    async def on_stop(
+        self,
+        *,
+        force: bool = False,
+        timeout: Optional[float] = None,
+    ) -> None:
+        if timeout is None:
+            timeout = DEFAULT_STOP_TIMEOUT
+
+        if self._process is None:
+            return  # Already stopped.
+
+        running = True
+
+        if not force:
+            self._process.terminate()
+            try:
+                await asyncio.wait_for(mp_wait(self._process), timeout)
+                running = False
+            except TimeoutError:
+                self.manager.logger.warning(
+                    f"{self.name} is still running after "
+                    f"{timeout} seconds. Killing."
+                )
+
+        if running:
+            self._process.kill()
+            await mp_wait(self._process)
+
+        self._process.close()
+        self._process = None
+
+        self.manager.logger.info(f"{self.name} stopped")
+
+    @property
+    def pid(self) -> Optional[int]:
+        return self._process.pid if self._process else None
+
+    def _run(self) -> None:
+        self.before_run()
+        self.run()
+        self.after_run()
+
+    def before_start(self) -> None:
+        pass
+
+    def after_start(self) -> None:
+        pass
+
+    def before_run(self) -> None:
+        pass
+
+    def after_run(self) -> None:
+        pass
+
+    @abstractmethod
+    def run(self) -> None:
+        pass
+
+    def __getstate__(self) -> dict:
+        return {
+            k: v
+            for k, v in self.__dict__.items()
+            if not (k.startswith("_Service__") or k == "_process")
+        }
+
+
+class ServiceProcess(BaseServiceProcess):
+    logger: logging.Logger
+
+    @property
+    def stop_event(self) -> threading.Event:
+        # Lazily create the stop_event. This allows the signal handler to tell if anyone is
+        # monitoring the stop event, and to raise a SystemExit if not.
+        if "stop_event" not in self.__dict__:
+            stop_event = threading.Event()
+            self.__dict__["stop_event"] = stop_event
+        else:
+            stop_event = self.__dict__["stop_event"]
+            assert isinstance(stop_event, threading.Event)
+
+        return stop_event
+
+    def before_start(self) -> None:
+        if frigate.log.log_listener is None:
+            raise RuntimeError("Logging has not yet been set up.")
+        self.__log_queue = frigate.log.log_listener.queue
+
+    def before_run(self) -> None:
+        super().before_run()
+
+        faulthandler.enable()
+
+        def receiveSignal(signalNumber: int, frame: Optional[FrameType]) -> None:
+            # Get the stop_event through the dict to bypass lazy initialization.
+            stop_event = self.__dict__.get("stop_event")
+            if stop_event is not None:
+                # Someone is monitoring stop_event. We should set it.
+                stop_event.set()
+            else:
+                # Nobody is monitoring stop_event. We should raise SystemExit.
+                sys.exit()
+
+        signal.signal(signal.SIGTERM, receiveSignal)
+        signal.signal(signal.SIGINT, receiveSignal)
+
+        self.logger = logging.getLogger(self.name)
+
+        logging.basicConfig(handlers=[], force=True)
+        logging.getLogger().addHandler(QueueHandler(self.__log_queue))
+        del self.__log_queue

frigate/service_manager/multiprocessing_waiter.py

@@ -0,0 +1,150 @@
+import asyncio
+import functools
+import logging
+import multiprocessing as mp
+import queue
+import threading
+from multiprocessing.connection import Connection
+from multiprocessing.connection import wait as mp_wait
+from socket import socket
+from typing import Any, Optional, Union
+
+logger = logging.getLogger(__name__)
+
+
+class MultiprocessingWaiter(threading.Thread):
+    """A background thread that manages futures for the multiprocessing.connection.wait() method."""
+
+    def __init__(self) -> None:
+        super().__init__(daemon=True)
+
+        # Queue of objects to wait for and futures to set results for.
+        self._queue: queue.Queue[tuple[Any, asyncio.Future[None]]] = queue.Queue()
+
+        # This is required to get mp_wait() to wake up when new objects to wait for are received.
+        receive, send = mp.Pipe(duplex=False)
+        self._receive_connection = receive
+        self._send_connection = send
+
+    def wait_for_sentinel(self, sentinel: Any) -> asyncio.Future[None]:
+        """Create an asyncio.Future tracking a sentinel for multiprocessing.connection.wait()
+
+        Warning: This method is NOT thread-safe.
+        """
+        # This would be incredibly stupid, but you never know.
+        assert sentinel != self._receive_connection
+
+        # Send the future to the background thread for processing.
+        future = asyncio.get_running_loop().create_future()
+        self._queue.put((sentinel, future))
+
+        # Notify the background thread.
+        #
+        # This is the non-thread-safe part, but since this method is not really meant to be called
+        # by users, we can get away with not adding a lock at this point (to avoid adding 2 locks).
+        self._send_connection.send_bytes(b".")
+
+        return future
+
+    def run(self) -> None:
+        logger.debug("Started background thread")
+
+        wait_dict: dict[Any, set[asyncio.Future[None]]] = {
+            self._receive_connection: set()
+        }
+        while True:
+            for ready_obj in mp_wait(wait_dict.keys()):
+                # Make sure we never remove the receive connection from the wait dict
+                if ready_obj is self._receive_connection:
+                    continue
+
+                logger.debug(
+                    f"Sentinel {ready_obj!r} is ready. "
+                    f"Notifying {len(wait_dict[ready_obj])} future(s)."
+                )
+
+                # Go over all the futures attached to this object and mark them as ready.
+                for fut in wait_dict.pop(ready_obj):
+                    if fut.cancelled():
+                        logger.debug(
+                            f"A future for sentinel {ready_obj!r} is ready, "
+                            "but the future is cancelled. Skipping."
+                        )
+                    else:
+                        fut.get_loop().call_soon_threadsafe(
+                            # Note: We need to check fut.cancelled() again, since it might
+                            # have been set before the event loop's definition of "soon".
+                            functools.partial(
+                                lambda fut: fut.cancelled() or fut.set_result(None), fut
+                            )
+                        )
+
+            # Check for cancellations in the remaining futures.
+            done_objects = []
+            for obj, fut_set in wait_dict.items():
+                if obj is self._receive_connection:
+                    continue
+
+                # Find any cancelled futures and remove them.
+                cancelled = [fut for fut in fut_set if fut.cancelled()]
+                fut_set.difference_update(cancelled)
+                logger.debug(
+                    f"Removing {len(cancelled)} future(s) from sentinel: {obj!r}"
+                )
+
+                # Mark objects with no remaining futures for removal.
+                if len(fut_set) == 0:
+                    done_objects.append(obj)
+
+            # Remove any objects that are done after removing cancelled futures.
+            for obj in done_objects:
+                logger.debug(
+                    f"Sentinel {obj!r} no longer has any futures waiting for it."
+                )
+                del wait_dict[obj]
+
+            # Get new objects to wait for from the queue.
+            while True:
+                try:
+                    obj, fut = self._queue.get_nowait()
+                    self._receive_connection.recv_bytes(maxlength=1)
+                    self._queue.task_done()
+
+                    logger.debug(f"Received new sentinel: {obj!r}")
+
+                    wait_dict.setdefault(obj, set()).add(fut)
+                except queue.Empty:
+                    break
+
+
+waiter_lock = threading.Lock()
+waiter_thread: Optional[MultiprocessingWaiter] = None
+
+
+async def wait(object: Union[mp.Process, Connection, socket]) -> None:
+    """Wait for the supplied object to be ready.
+
+    Under the hood, this uses multiprocessing.connection.wait() and a background thread manage the
+    returned futures.
+    """
+    global waiter_thread, waiter_lock
+
+    sentinel: Union[Connection, socket, int]
+    if isinstance(object, mp.Process):
+        sentinel = object.sentinel
+    elif isinstance(object, Connection) or isinstance(object, socket):
+        sentinel = object
+    else:
+        raise ValueError(f"Cannot wait for object of type {type(object).__qualname__}")
+
+    with waiter_lock:
+        if waiter_thread is None:
+            # Start a new waiter thread.
+            waiter_thread = MultiprocessingWaiter()
+            waiter_thread.start()
+
+        # Create the future while still holding the lock,
+        # since wait_for_sentinel() is not thread safe.
+        fut = waiter_thread.wait_for_sentinel(sentinel)
+
+    await fut

frigate/service_manager/service.py

@@ -0,0 +1,446 @@
+from __future__ import annotations
+
+import asyncio
+import atexit
+import logging
+import threading
+from abc import ABC, abstractmethod
+from contextvars import ContextVar
+from dataclasses import dataclass
+from functools import partial
+from typing import Coroutine, Optional, Union, cast
+
+from typing_extensions import Self
+
+
+class Service(ABC):
+    """An abstract service instance."""
+
+    def __init__(
+        self,
+        *,
+        name: Optional[str] = None,
+        manager: Optional[ServiceManager] = None,
+    ):
+        if name:
+            self.__dict__["name"] = name
+
+        self.__manager = manager or ServiceManager.current()
+        self.__lock = asyncio.Lock(loop=self.__manager._event_loop)
+        self.__manager._register(self)
+
+    @property
+    def name(self) -> str:
+        try:
+            return cast(str, self.__dict__["name"])
+        except KeyError:
+            return type(self).__qualname__
+
+    @property
+    def manager(self) -> ServiceManager:
+        """The service manager this service is registered with."""
+        try:
+            return self.__manager
+        except AttributeError:
+            raise RuntimeError("Cannot access associated service manager")
+
+    def start(
+        self,
+        *,
+        wait: bool = False,
+        wait_timeout: Optional[float] = None,
+    ) -> Self:
+        """Start this service.
+
+        :param wait: If set, this function will block until the task is complete.
+        :param wait_timeout: If set, this function will not return until the task is complete or the
+            specified timeout has elapsed.
+        """
+
+        self.manager.run_task(
+            self.on_start(),
+            wait=wait,
+            wait_timeout=wait_timeout,
+            lock=self.__lock,
+        )
+
+        return self
+
+    def stop(
+        self,
+        *,
+        force: bool = False,
+        timeout: Optional[float] = None,
+        wait: bool = False,
+        wait_timeout: Optional[float] = None,
+    ) -> Self:
+        """Stop this service.
+
+        :param force: If set, the service will be killed immediately.
+        :param timeout: Maximum amount of time to wait before force-killing the service.
+
+        :param wait: If set, this function will block until the task is complete.
+        :param wait_timeout: If set, this function will not return until the task is complete or the
+            specified timeout has elapsed.
+        """
+
+        self.manager.run_task(
+            self.on_stop(force=force, timeout=timeout),
+            wait=wait,
+            wait_timeout=wait_timeout,
+            lock=self.__lock,
+        )
+
+        return self
+
+    def restart(
+        self,
+        *,
+        force: bool = False,
+        stop_timeout: Optional[float] = None,
+        wait: bool = False,
+        wait_timeout: Optional[float] = None,
+    ) -> Self:
+        """Restart this service.
+
+        :param force: If set, the service will be killed immediately.
+        :param timeout: Maximum amount of time to wait before force-killing the service.
+
+        :param wait: If set, this function will block until the task is complete.
+        :param wait_timeout: If set, this function will not return until the task is complete or the
+            specified timeout has elapsed.
+        """
+
+        self.manager.run_task(
+            self.on_restart(force=force, stop_timeout=stop_timeout),
+            wait=wait,
+            wait_timeout=wait_timeout,
+            lock=self.__lock,
+        )
+
+        return self
+
+    @abstractmethod
+    async def on_start(self) -> None:
+        pass
+
+    @abstractmethod
+    async def on_stop(
+        self,
+        *,
+        force: bool = False,
+        timeout: Optional[float] = None,
+    ) -> None:
+        pass
+
+    async def on_restart(
+        self,
+        *,
+        force: bool = False,
+        stop_timeout: Optional[float] = None,
+    ) -> None:
+        await self.on_stop(force=force, timeout=stop_timeout)
+        await self.on_start()
+
+
+default_service_manager_lock = threading.Lock()
+default_service_manager: Optional[ServiceManager] = None
+
+current_service_manager: ContextVar[ServiceManager] = ContextVar(
+    "current_service_manager"
+)
+
+
+@dataclass
+class Command:
+    """A coroutine to execute in the service manager thread.
+
+    Attributes:
+        coro: The coroutine to execute.
+        lock: An async lock to acquire before calling the coroutine.
+        done: If specified, the service manager will set this event after the command completes.
+    """
+
+    coro: Coroutine
+    lock: Optional[asyncio.Lock] = None
+    done: Optional[threading.Event] = None
+
+
+class ServiceManager:
+    """A set of services, along with the global state required to manage them efficiently.
+
+    Typically users of the service infrastructure will not interact with a service manager directly,
+    but rather through individual Service subclasses that will automatically manage a service
+    manager instance.
+
+    Each service manager instance has a background thread in which service lifecycle tasks are
+    executed in an async executor. This is done to avoid head-of-line blocking in the business logic
+    that spins up individual services. This thread is automatically started when the service manager
+    is created and stopped either manually, or on application exit.
+
+    All (public) service manager methods are thread-safe.
+    """
+
+    _name: str
+    _logger: logging.Logger
+
+    # The set of services this service manager knows about.
+    _services: dict[str, Service]
+    _services_lock: threading.Lock
+
+    # Commands will be queued with associated event loop. Queueing `None` signals shutdown.
+    _command_queue: asyncio.Queue[Union[Command, None]]
+    _event_loop: asyncio.AbstractEventLoop
+
+    # The pending command counter is used to ensure all commands have been queued before shutdown.
+    _pending_commands: AtomicCounter
+
+    # The set of pending tasks after they have been received by the background thread and spawned.
+    _tasks: set
+
+    # Fired after the async runtime starts. Object initialization completes after this is set.
+    _setup_event: threading.Event
+
+    # Will be acquired to ensure the shutdown sentinel is sent only once. Never released.
+    _shutdown_lock: threading.Lock
+
+    def __init__(self, *, name: Optional[str] = None):
+        self._name = name if name is not None else (__package__ or __name__)
+        self._logger = logging.getLogger(self.name)
+
+        self._services = dict()
+        self._services_lock = threading.Lock()
+
+        self._pending_commands = AtomicCounter()
+        self._tasks = set()
+
+        self._shutdown_lock = threading.Lock()
+
+        # --- Start the manager thread and wait for it to be ready. ---
+
+        self._setup_event = threading.Event()
+
+        async def start_manager() -> None:
+            self._event_loop = asyncio.get_running_loop()
+            self._command_queue = asyncio.Queue()
+
+            self._setup_event.set()
+            await self._monitor_command_queue()
+
+        self._manager_thread = threading.Thread(
+            name=self.name,
+            target=lambda: asyncio.run(start_manager()),
+            daemon=True,
+        )
+
+        self._manager_thread.start()
+        atexit.register(partial(self.shutdown, wait=True))
+
+        self._setup_event.wait()
+
+    @property
+    def name(self) -> str:
+        """The name of this service manager. Primarily intended for logging purposes."""
+        return self._name
+
+    @property
+    def logger(self) -> logging.Logger:
+        """The logger used by this service manager."""
+        return self._logger
+
+    @classmethod
+    def current(cls) -> ServiceManager:
+        """The service manager set in the current context (async task or thread).
+
+        A global default service manager will be automatically created on first access."""
+
+        global default_service_manager
+
+        current = current_service_manager.get(None)
+        if current is None:
+            with default_service_manager_lock:
+                if default_service_manager is None:
+                    default_service_manager = cls()
+
+            current = default_service_manager
+            current_service_manager.set(current)
+        return current
+
+    def make_current(self) -> None:
+        """Make this the current service manager."""
+
+        current_service_manager.set(self)
+
+    def run_task(
+        self,
+        coro: Coroutine,
+        *,
+        wait: bool = False,
+        wait_timeout: Optional[float] = None,
+        lock: Optional[asyncio.Lock] = None,
+    ) -> None:
+        """Run an async task in the service manager thread.
+
+        :param wait: If set, this function will block until the task is complete.
+        :param wait_timeout: If set, this function will not return until the task is complete or the
+            specified timeout has elapsed.
+        """
+
+        if not isinstance(coro, Coroutine):
+            raise TypeError(f"Cannot schedule task for object of type {type(coro)}")
+
+        cmd = Command(coro=coro, lock=lock)
+        if wait or wait_timeout is not None:
+            cmd.done = threading.Event()
+
+        self._send_command(cmd)
+
+        if cmd.done is not None:
+            cmd.done.wait(timeout=wait_timeout)
+
+    def shutdown(
+        self, *, wait: bool = False, wait_timeout: Optional[float] = None
+    ) -> None:
+        """Shutdown the service manager thread.
+
+        After the shutdown process completes, any subsequent calls to the service manager will
+        produce an error.
+
+        :param wait: If set, this function will block until the shutdown process is complete.
+        :param wait_timeout: If set, this function will not return until the shutdown process is
+            complete or the specified timeout has elapsed.
+        """
+
+        if self._shutdown_lock.acquire(blocking=False):
+            self._send_command(None)
+        if wait:
+            self._manager_thread.join(timeout=wait_timeout)
+
+    def _ensure_running(self) -> None:
+        self._setup_event.wait()
+        if not self._manager_thread.is_alive():
+            raise RuntimeError(f"ServiceManager {self.name} is not running")
+
+    def _send_command(self, command: Union[Command, None]) -> None:
+        self._ensure_running()
+
+        async def queue_command() -> None:
+            await self._command_queue.put(command)
+            self._pending_commands.sub()
+
+        self._pending_commands.add()
+        asyncio.run_coroutine_threadsafe(queue_command(), self._event_loop)
+
+    def _register(self, service: Service) -> None:
+        """Register a service with the service manager. This is done by the service constructor."""
+
+        self._ensure_running()
+        with self._services_lock:
+            name_conflict: Optional[Service] = next(
+                (
+                    existing
+                    for name, existing in self._services.items()
+                    if name == service.name
+                ),
+                None,
+            )
+
+            if name_conflict is service:
+                raise RuntimeError(f"Attempt to re-register service: {service.name}")
+            elif name_conflict is not None:
+                raise RuntimeError(f"Duplicate service name: {service.name}")
+
+            self.logger.debug(f"Registering service: {service.name}")
+            self._services[service.name] = service
+
+    def _run_command(self, command: Command) -> None:
+        """Execute a command and add it to the tasks set."""
+
+        def task_done(task: asyncio.Task) -> None:
+            exc = task.exception()
+            if exc:
+                self.logger.exception("Exception in service manager task", exc_info=exc)
+            self._tasks.discard(task)
+            if command.done is not None:
+                command.done.set()
+
+        async def task_harness() -> None:
+            if command.lock is not None:
+                async with command.lock:
+                    await command.coro
+            else:
+                await command.coro
+
+        task = asyncio.create_task(task_harness())
+        task.add_done_callback(task_done)
+        self._tasks.add(task)
+
+    async def _monitor_command_queue(self) -> None:
+        """The main function of the background thread."""
+
+        self.logger.info("Started service manager")
+
+        # Main command processing loop.
+        while (command := await self._command_queue.get()) is not None:
+            self._run_command(command)
+
+        # Send a stop command to all services. We don't have a status command yet, so we can just
+        # stop everything and be done with it.
+        with self._services_lock:
+            self.logger.debug(f"Stopping {len(self._services)} services")
+            for service in self._services.values():
+                service.stop()
+
+        # Wait for all commands to finish executing.
+        await self._shutdown()
+
+        self.logger.info("Exiting service manager")
+
+    async def _shutdown(self) -> None:
+        """Ensure all commands have been queued & executed."""
+
+        while True:
+            command = None
+            try:
+                # Try and get a command from the queue.
+                command = self._command_queue.get_nowait()
+            except asyncio.QueueEmpty:
+                if self._pending_commands.value > 0:
+                    # If there are pending commands to queue, await them.
+                    command = await self._command_queue.get()
+                elif self._tasks:
+                    # If there are still pending tasks, wait for them. These tasks might queue
+                    # commands though, so we have to loop again.
+                    await asyncio.wait(self._tasks)
+                else:
+                    # Nothing is pending at this point, so we're done here.
+                    break
+
+            # If we got a command, run it.
+            if command is not None:
+                self._run_command(command)
+
+
+class AtomicCounter:
+    """A lock-protected atomic counter."""
+
+    # Modern CPUs have atomics, but python doesn't seem to include them in the standard library.
+    # Besides, the performance penalty is negligible compared to, well, using python.
+    # So this will do just fine.
+
+    def __init__(self, initial: int = 0):
+        self._lock = threading.Lock()
+        self._value = initial
+
+    def add(self, value: int = 1) -> None:
+        with self._lock:
+            self._value += value
+
+    def sub(self, value: int = 1) -> None:
+        with self._lock:
+            self._value -= value
+
+    @property
+    def value(self) -> int:
+        with self._lock:
+            return self._value

frigate/stats/util.py

@@ -197,8 +197,8 @@ async def set_gpu_stats(
             # intel QSV GPU
             intel_usage = get_intel_gpu_stats()
 
-            if intel_usage:
-                stats["intel-qsv"] = intel_usage
+            if intel_usage is not None:
+                stats["intel-qsv"] = intel_usage or {"gpu": "", "mem": ""}
             else:
                 stats["intel-qsv"] = {"gpu": "", "mem": ""}
                 hwaccel_errors.append(args)
@@ -222,8 +222,8 @@ async def set_gpu_stats(
                 # intel VAAPI GPU
                 intel_usage = get_intel_gpu_stats()
 
-                if intel_usage:
-                    stats["intel-vaapi"] = intel_usage
+                if intel_usage is not None:
+                    stats["intel-vaapi"] = intel_usage or {"gpu": "", "mem": ""}
                 else:
                     stats["intel-vaapi"] = {"gpu": "", "mem": ""}
                     hwaccel_errors.append(args)

frigate/test/test_obects.py

@@ -1,11 +1,11 @@
 import unittest
 
-from frigate.track.object_attribute import ObjectAttribute
+from frigate.track.tracked_object import TrackedObjectAttribute
 
 
 class TestAttribute(unittest.TestCase):
     def test_overlapping_object_selection(self) -> None:
-        attribute = ObjectAttribute(
+        attribute = TrackedObjectAttribute(
             (
                 "amazon",
                 0.80078125,

frigate/track/object_attribute.py

@@ -1,44 +0,0 @@
-"""Object attribute."""
-
-from frigate.util.object import area, box_inside
-
-
-class ObjectAttribute:
-    def __init__(self, raw_data: tuple) -> None:
-        self.label = raw_data[0]
-        self.score = raw_data[1]
-        self.box = raw_data[2]
-        self.area = raw_data[3]
-        self.ratio = raw_data[4]
-        self.region = raw_data[5]
-
-    def get_tracking_data(self) -> dict[str, any]:
-        """Return data saved to the object."""
-        return {
-            "label": self.label,
-            "score": self.score,
-            "box": self.box,
-        }
-
-    def find_best_object(self, objects: list[dict[str, any]]) -> str:
-        """Find the best attribute for each object and return its ID."""
-        best_object_area = None
-        best_object_id = None
-
-        for obj in objects:
-            if not box_inside(obj["box"], self.box):
-                continue
-
-            object_area = area(obj["box"])
-
-            # if multiple objects have the same attribute then they
-            # are overlapping, it is most likely that the smaller object
-            # is the one with the attribute
-            if best_object_area is None:
-                best_object_area = object_area
-                best_object_id = obj["id"]
-            elif object_area < best_object_area:
-                best_object_area = object_area
-                best_object_id = obj["id"]
-
-        return best_object_id

frigate/track/tracked_object.py

@@ -0,0 +1,447 @@
+"""Object attribute."""
+
+import base64
+import logging
+from collections import defaultdict
+from statistics import median
+
+import cv2
+import numpy as np
+
+from frigate.config import (
+    CameraConfig,
+    ModelConfig,
+)
+from frigate.util.image import (
+    area,
+    calculate_region,
+    draw_box_with_label,
+    draw_timestamp,
+    is_better_thumbnail,
+)
+from frigate.util.object import box_inside
+
+logger = logging.getLogger(__name__)
+
+
+class TrackedObject:
+    def __init__(
+        self,
+        model_config: ModelConfig,
+        camera_config: CameraConfig,
+        frame_cache,
+        obj_data: dict[str, any],
+    ):
+        # set the score history then remove as it is not part of object state
+        self.score_history = obj_data["score_history"]
+        del obj_data["score_history"]
+
+        self.obj_data = obj_data
+        self.colormap = model_config.colormap
+        self.logos = model_config.all_attribute_logos
+        self.camera_config = camera_config
+        self.frame_cache = frame_cache
+        self.zone_presence: dict[str, int] = {}
+        self.zone_loitering: dict[str, int] = {}
+        self.current_zones = []
+        self.entered_zones = []
+        self.attributes = defaultdict(float)
+        self.false_positive = True
+        self.has_clip = False
+        self.has_snapshot = False
+        self.top_score = self.computed_score = 0.0
+        self.thumbnail_data = None
+        self.last_updated = 0
+        self.last_published = 0
+        self.frame = None
+        self.active = True
+        self.pending_loitering = False
+        self.previous = self.to_dict()
+
+    def _is_false_positive(self):
+        # once a true positive, always a true positive
+        if not self.false_positive:
+            return False
+
+        threshold = self.camera_config.objects.filters[self.obj_data["label"]].threshold
+        return self.computed_score < threshold
+
+    def compute_score(self):
+        """get median of scores for object."""
+        return median(self.score_history)
+
+    def update(self, current_frame_time: float, obj_data, has_valid_frame: bool):
+        thumb_update = False
+        significant_change = False
+        autotracker_update = False
+        # if the object is not in the current frame, add a 0.0 to the score history
+        if obj_data["frame_time"] != current_frame_time:
+            self.score_history.append(0.0)
+        else:
+            self.score_history.append(obj_data["score"])
+
+        # only keep the last 10 scores
+        if len(self.score_history) > 10:
+            self.score_history = self.score_history[-10:]
+
+        # calculate if this is a false positive
+        self.computed_score = self.compute_score()
+        if self.computed_score > self.top_score:
+            self.top_score = self.computed_score
+        self.false_positive = self._is_false_positive()
+        self.active = self.is_active()
+
+        if not self.false_positive and has_valid_frame:
+            # determine if this frame is a better thumbnail
+            if self.thumbnail_data is None or is_better_thumbnail(
+                self.obj_data["label"],
+                self.thumbnail_data,
+                obj_data,
+                self.camera_config.frame_shape,
+            ):
+                self.thumbnail_data = {
+                    "frame_time": current_frame_time,
+                    "box": obj_data["box"],
+                    "area": obj_data["area"],
+                    "region": obj_data["region"],
+                    "score": obj_data["score"],
+                    "attributes": obj_data["attributes"],
+                }
+                thumb_update = True
+
+        # check zones
+        current_zones = []
+        bottom_center = (obj_data["centroid"][0], obj_data["box"][3])
+        in_loitering_zone = False
+
+        # check each zone
+        for name, zone in self.camera_config.zones.items():
+            # if the zone is not for this object type, skip
+            if len(zone.objects) > 0 and obj_data["label"] not in zone.objects:
+                continue
+            contour = zone.contour
+            zone_score = self.zone_presence.get(name, 0) + 1
+            # check if the object is in the zone
+            if cv2.pointPolygonTest(contour, bottom_center, False) >= 0:
+                # if the object passed the filters once, dont apply again
+                if name in self.current_zones or not zone_filtered(self, zone.filters):
+                    # an object is only considered present in a zone if it has a zone inertia of 3+
+                    if zone_score >= zone.inertia:
+                        # if the zone has loitering time, update loitering status
+                        if zone.loitering_time > 0:
+                            in_loitering_zone = True
+
+                        loitering_score = self.zone_loitering.get(name, 0) + 1
+
+                        # loitering time is configured as seconds, convert to count of frames
+                        if loitering_score >= (
+                            self.camera_config.zones[name].loitering_time
+                            * self.camera_config.detect.fps
+                        ):
+                            current_zones.append(name)
+
+                            if name not in self.entered_zones:
+                                self.entered_zones.append(name)
+                        else:
+                            self.zone_loitering[name] = loitering_score
+                    else:
+                        self.zone_presence[name] = zone_score
+            else:
+                # once an object has a zone inertia of 3+ it is not checked anymore
+                if 0 < zone_score < zone.inertia:
+                    self.zone_presence[name] = zone_score - 1
+
+        # update loitering status
+        self.pending_loitering = in_loitering_zone
+
+        # maintain attributes
+        for attr in obj_data["attributes"]:
+            if self.attributes[attr["label"]] < attr["score"]:
+                self.attributes[attr["label"]] = attr["score"]
+
+        # populate the sub_label for object with highest scoring logo
+        if self.obj_data["label"] in ["car", "package", "person"]:
+            recognized_logos = {
+                k: self.attributes[k] for k in self.logos if k in self.attributes
+            }
+            if len(recognized_logos) > 0:
+                max_logo = max(recognized_logos, key=recognized_logos.get)
+
+                # don't overwrite sub label if it is already set
+                if (
+                    self.obj_data.get("sub_label") is None
+                    or self.obj_data["sub_label"][0] == max_logo
+                ):
+                    self.obj_data["sub_label"] = (max_logo, recognized_logos[max_logo])
+
+        # check for significant change
+        if not self.false_positive:
+            # if the zones changed, signal an update
+            if set(self.current_zones) != set(current_zones):
+                significant_change = True
+
+            # if the position changed, signal an update
+            if self.obj_data["position_changes"] != obj_data["position_changes"]:
+                significant_change = True
+
+            if self.obj_data["attributes"] != obj_data["attributes"]:
+                significant_change = True
+
+            # if the state changed between stationary and active
+            if self.previous["active"] != self.active:
+                significant_change = True
+
+            # update at least once per minute
+            if self.obj_data["frame_time"] - self.previous["frame_time"] > 60:
+                significant_change = True
+
+            # update autotrack at most 3 objects per second
+            if self.obj_data["frame_time"] - self.previous["frame_time"] >= (1 / 3):
+                autotracker_update = True
+
+        self.obj_data.update(obj_data)
+        self.current_zones = current_zones
+        return (thumb_update, significant_change, autotracker_update)
+
+    def to_dict(self, include_thumbnail: bool = False):
+        event = {
+            "id": self.obj_data["id"],
+            "camera": self.camera_config.name,
+            "frame_time": self.obj_data["frame_time"],
+            "snapshot": self.thumbnail_data,
+            "label": self.obj_data["label"],
+            "sub_label": self.obj_data.get("sub_label"),
+            "top_score": self.top_score,
+            "false_positive": self.false_positive,
+            "start_time": self.obj_data["start_time"],
+            "end_time": self.obj_data.get("end_time", None),
+            "score": self.obj_data["score"],
+            "box": self.obj_data["box"],
+            "area": self.obj_data["area"],
+            "ratio": self.obj_data["ratio"],
+            "region": self.obj_data["region"],
+            "active": self.active,
+            "stationary": not self.active,
+            "motionless_count": self.obj_data["motionless_count"],
+            "position_changes": self.obj_data["position_changes"],
+            "current_zones": self.current_zones.copy(),
+            "entered_zones": self.entered_zones.copy(),
+            "has_clip": self.has_clip,
+            "has_snapshot": self.has_snapshot,
+            "attributes": self.attributes,
+            "current_attributes": self.obj_data["attributes"],
+            "pending_loitering": self.pending_loitering,
+        }
+
+        if include_thumbnail:
+            event["thumbnail"] = base64.b64encode(self.get_thumbnail()).decode("utf-8")
+
+        return event
+
+    def is_active(self):
+        return not self.is_stationary()
+
+    def is_stationary(self):
+        return (
+            self.obj_data["motionless_count"]
+            > self.camera_config.detect.stationary.threshold
+        )
+
+    def get_thumbnail(self):
+        if (
+            self.thumbnail_data is None
+            or self.thumbnail_data["frame_time"] not in self.frame_cache
+        ):
+            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
+
+        jpg_bytes = self.get_jpg_bytes(
+            timestamp=False, bounding_box=False, crop=True, height=175
+        )
+
+        if jpg_bytes:
+            return jpg_bytes
+        else:
+            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
+            return jpg.tobytes()
+
+    def get_clean_png(self):
+        if self.thumbnail_data is None:
+            return None
+
+        try:
+            best_frame = cv2.cvtColor(
+                self.frame_cache[self.thumbnail_data["frame_time"]],
+                cv2.COLOR_YUV2BGR_I420,
+            )
+        except KeyError:
+            logger.warning(
+                f"Unable to create clean png because frame {self.thumbnail_data['frame_time']} is not in the cache"
+            )
+            return None
+
+        ret, png = cv2.imencode(".png", best_frame)
+        if ret:
+            return png.tobytes()
+        else:
+            return None
+
+    def get_jpg_bytes(
+        self, timestamp=False, bounding_box=False, crop=False, height=None, quality=70
+    ):
+        if self.thumbnail_data is None:
+            return None
+
+        try:
+            best_frame = cv2.cvtColor(
+                self.frame_cache[self.thumbnail_data["frame_time"]],
+                cv2.COLOR_YUV2BGR_I420,
+            )
+        except KeyError:
+            logger.warning(
+                f"Unable to create jpg because frame {self.thumbnail_data['frame_time']} is not in the cache"
+            )
+            return None
+
+        if bounding_box:
+            thickness = 2
+            color = self.colormap[self.obj_data["label"]]
+
+            # draw the bounding boxes on the frame
+            box = self.thumbnail_data["box"]
+            draw_box_with_label(
+                best_frame,
+                box[0],
+                box[1],
+                box[2],
+                box[3],
+                self.obj_data["label"],
+                f"{int(self.thumbnail_data['score']*100)}% {int(self.thumbnail_data['area'])}",
+                thickness=thickness,
+                color=color,
+            )
+
+            # draw any attributes
+            for attribute in self.thumbnail_data["attributes"]:
+                box = attribute["box"]
+                draw_box_with_label(
+                    best_frame,
+                    box[0],
+                    box[1],
+                    box[2],
+                    box[3],
+                    attribute["label"],
+                    f"{attribute['score']:.0%}",
+                    thickness=thickness,
+                    color=color,
+                )
+
+        if crop:
+            box = self.thumbnail_data["box"]
+            box_size = 300
+            region = calculate_region(
+                best_frame.shape,
+                box[0],
+                box[1],
+                box[2],
+                box[3],
+                box_size,
+                multiplier=1.1,
+            )
+            best_frame = best_frame[region[1] : region[3], region[0] : region[2]]
+
+        if height:
+            width = int(height * best_frame.shape[1] / best_frame.shape[0])
+            best_frame = cv2.resize(
+                best_frame, dsize=(width, height), interpolation=cv2.INTER_AREA
+            )
+        if timestamp:
+            color = self.camera_config.timestamp_style.color
+            draw_timestamp(
+                best_frame,
+                self.thumbnail_data["frame_time"],
+                self.camera_config.timestamp_style.format,
+                font_effect=self.camera_config.timestamp_style.effect,
+                font_thickness=self.camera_config.timestamp_style.thickness,
+                font_color=(color.blue, color.green, color.red),
+                position=self.camera_config.timestamp_style.position,
+            )
+
+        ret, jpg = cv2.imencode(
+            ".jpg", best_frame, [int(cv2.IMWRITE_JPEG_QUALITY), quality]
+        )
+        if ret:
+            return jpg.tobytes()
+        else:
+            return None
+
+
+def zone_filtered(obj: TrackedObject, object_config):
+    object_name = obj.obj_data["label"]
+
+    if object_name in object_config:
+        obj_settings = object_config[object_name]
+
+        # if the min area is larger than the
+        # detected object, don't add it to detected objects
+        if obj_settings.min_area > obj.obj_data["area"]:
+            return True
+
+        # if the detected object is larger than the
+        # max area, don't add it to detected objects
+        if obj_settings.max_area < obj.obj_data["area"]:
+            return True
+
+        # if the score is lower than the threshold, skip
+        if obj_settings.threshold > obj.computed_score:
+            return True
+
+        # if the object is not proportionally wide enough
+        if obj_settings.min_ratio > obj.obj_data["ratio"]:
+            return True
+
+        # if the object is proportionally too wide
+        if obj_settings.max_ratio < obj.obj_data["ratio"]:
+            return True
+
+    return False
+
+
+class TrackedObjectAttribute:
+    def __init__(self, raw_data: tuple) -> None:
+        self.label = raw_data[0]
+        self.score = raw_data[1]
+        self.box = raw_data[2]
+        self.area = raw_data[3]
+        self.ratio = raw_data[4]
+        self.region = raw_data[5]
+
+    def get_tracking_data(self) -> dict[str, any]:
+        """Return data saved to the object."""
+        return {
+            "label": self.label,
+            "score": self.score,
+            "box": self.box,
+        }
+
+    def find_best_object(self, objects: list[dict[str, any]]) -> str:
+        """Find the best attribute for each object and return its ID."""
+        best_object_area = None
+        best_object_id = None
+
+        for obj in objects:
+            if not box_inside(obj["box"], self.box):
+                continue
+
+            object_area = area(obj["box"])
+
+            # if multiple objects have the same attribute then they
+            # are overlapping, it is most likely that the smaller object
+            # is the one with the attribute
+            if best_object_area is None:
+                best_object_area = object_area
+                best_object_id = obj["id"]
+            elif object_area < best_object_area:
+                best_object_area = object_area
+                best_object_id = obj["id"]
+
+        return best_object_id

frigate/util/builtin.py

@@ -8,10 +8,11 @@ import multiprocessing as mp
 import queue
 import re
 import shlex
+import struct
 import urllib.parse
 from collections.abc import Mapping
 from pathlib import Path
-from typing import Any, Optional, Tuple
+from typing import Any, Optional, Tuple, Union
 
 import numpy as np
 import pytz
@@ -182,16 +183,11 @@ def update_yaml_from_url(file_path, url):
             update_yaml_file(file_path, key_path, new_value_list)
         else:
             value = new_value_list[0]
-            if "," in value:
-                # Skip conversion if we're a mask or zone string
-                update_yaml_file(file_path, key_path, value)
-            else:
-                try:
-                    value = ast.literal_eval(value)
-                except (ValueError, SyntaxError):
-                    pass
-                update_yaml_file(file_path, key_path, value)
-
+            try:
+                # no need to convert if we have a mask/zone string
+                value = ast.literal_eval(value) if "," not in value else value
+            except (ValueError, SyntaxError):
+                pass
             update_yaml_file(file_path, key_path, value)
 
 
@@ -342,3 +338,32 @@ def generate_color_palette(n):
         colors.append(interpolate(color1, color2, factor))
 
     return colors
+
+
+def serialize(
+    vector: Union[list[float], np.ndarray, float], pack: bool = True
+) -> bytes:
+    """Serializes a list of floats, numpy array, or single float into a compact "raw bytes" format"""
+    if isinstance(vector, np.ndarray):
+        # Convert numpy array to list of floats
+        vector = vector.flatten().tolist()
+    elif isinstance(vector, (float, np.float32, np.float64)):
+        # Handle single float values
+        vector = [vector]
+    elif not isinstance(vector, list):
+        raise TypeError(
+            f"Input must be a list of floats, a numpy array, or a single float. Got {type(vector)}"
+        )
+
+    try:
+        if pack:
+            return struct.pack("%sf" % len(vector), *vector)
+        else:
+            return vector
+    except struct.error as e:
+        raise ValueError(f"Failed to pack vector: {e}. Vector: {vector}")
+
+
+def deserialize(bytes_data: bytes) -> list[float]:
+    """Deserializes a compact "raw bytes" format into a list of floats"""
+    return list(struct.unpack("%sf" % (len(bytes_data) // 4), bytes_data))

frigate/util/downloader.py

@@ -19,6 +19,13 @@ class FileLock:
         self.path = path
         self.lock_file = f"{path}.lock"
 
+        # we have not acquired the lock yet so it should not exist
+        if os.path.exists(self.lock_file):
+            try:
+                os.remove(self.lock_file)
+            except Exception:
+                pass
+
     def acquire(self):
         parent_dir = os.path.dirname(self.lock_file)
         os.makedirs(parent_dir, exist_ok=True)
@@ -56,14 +63,12 @@ class ModelDownloader:
         self.download_complete = threading.Event()
 
     def ensure_model_files(self):
-        for file in self.file_names:
-            self.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.model_name}-{file}",
-                    "state": ModelStatusTypesEnum.downloading,
-                },
-            )
+        self.mark_files_state(
+            self.requestor,
+            self.model_name,
+            self.file_names,
+            ModelStatusTypesEnum.downloading,
+        )
         self.download_thread = threading.Thread(
             target=self._download_models,
             name=f"_download_model_{self.model_name}",
@@ -92,6 +97,7 @@ class ModelDownloader:
                 },
             )
 
+        self.requestor.stop()
         self.download_complete.set()
 
     @staticmethod
@@ -119,5 +125,21 @@ class ModelDownloader:
         if not silent:
             logger.info(f"Downloading complete: {url}")
 
+    @staticmethod
+    def mark_files_state(
+        requestor: InterProcessRequestor,
+        model_name: str,
+        files: list[str],
+        state: ModelStatusTypesEnum,
+    ) -> None:
+        for file_name in files:
+            requestor.send_data(
+                UPDATE_MODEL_STATE,
+                {
+                    "model": f"{model_name}-{file_name}",
+                    "state": state,
+                },
+            )
+
     def wait_for_download(self):
         self.download_complete.wait()

frigate/util/image.py

@@ -36,6 +36,72 @@ def transliterate_to_latin(text: str) -> str:
     return unidecode(text)
 
 
+def on_edge(box, frame_shape):
+    if (
+        box[0] == 0
+        or box[1] == 0
+        or box[2] == frame_shape[1] - 1
+        or box[3] == frame_shape[0] - 1
+    ):
+        return True
+
+
+def has_better_attr(current_thumb, new_obj, attr_label) -> bool:
+    max_new_attr = max(
+        [0]
+        + [area(a["box"]) for a in new_obj["attributes"] if a["label"] == attr_label]
+    )
+    max_current_attr = max(
+        [0]
+        + [
+            area(a["box"])
+            for a in current_thumb["attributes"]
+            if a["label"] == attr_label
+        ]
+    )
+
+    # if the thumb has a higher scoring attr
+    return max_new_attr > max_current_attr
+
+
+def is_better_thumbnail(label, current_thumb, new_obj, frame_shape) -> bool:
+    # larger is better
+    # cutoff images are less ideal, but they should also be smaller?
+    # better scores are obviously better too
+
+    # check face on person
+    if label == "person":
+        if has_better_attr(current_thumb, new_obj, "face"):
+            return True
+        # if the current thumb has a face attr, dont update unless it gets better
+        if any([a["label"] == "face" for a in current_thumb["attributes"]]):
+            return False
+
+    # check license_plate on car
+    if label == "car":
+        if has_better_attr(current_thumb, new_obj, "license_plate"):
+            return True
+        # if the current thumb has a license_plate attr, dont update unless it gets better
+        if any([a["label"] == "license_plate" for a in current_thumb["attributes"]]):
+            return False
+
+    # if the new_thumb is on an edge, and the current thumb is not
+    if on_edge(new_obj["box"], frame_shape) and not on_edge(
+        current_thumb["box"], frame_shape
+    ):
+        return False
+
+    # if the score is better by more than 5%
+    if new_obj["score"] > current_thumb["score"] + 0.05:
+        return True
+
+    # if the area is 10% larger
+    if new_obj["area"] > current_thumb["area"] * 1.1:
+        return True
+
+    return False
+
+
 def draw_timestamp(
     frame,
     timestamp,

frigate/util/model.py

@@ -1,39 +1,118 @@
 """Model Utils"""
 
 import os
+from typing import Any
 
 import onnxruntime as ort
 
+try:
+    import openvino as ov
+except ImportError:
+    # openvino is not included
+    pass
+
 
 def get_ort_providers(
-    force_cpu: bool = False, openvino_device: str = "AUTO"
+    force_cpu: bool = False, openvino_device: str = "AUTO", requires_fp16: bool = False
 ) -> tuple[list[str], list[dict[str, any]]]:
     if force_cpu:
-        return (["CPUExecutionProvider"], [{}])
+        return (
+            ["CPUExecutionProvider"],
+            [
+                {
+                    "enable_cpu_mem_arena": False,
+                }
+            ],
+        )
 
-    providers = ort.get_available_providers()
+    providers = []
     options = []
 
-    for provider in providers:
-        if provider == "TensorrtExecutionProvider":
-            os.makedirs("/config/model_cache/tensorrt/ort/trt-engines", exist_ok=True)
+    for provider in ort.get_available_providers():
+        if provider == "CUDAExecutionProvider":
+            providers.append(provider)
             options.append(
                 {
-                    "trt_timing_cache_enable": True,
-                    "trt_engine_cache_enable": True,
-                    "trt_timing_cache_path": "/config/model_cache/tensorrt/ort",
-                    "trt_engine_cache_path": "/config/model_cache/tensorrt/ort/trt-engines",
+                    "arena_extend_strategy": "kSameAsRequested",
                 }
             )
+        elif provider == "TensorrtExecutionProvider":
+            # TensorrtExecutionProvider uses too much memory without options to control it
+            pass
         elif provider == "OpenVINOExecutionProvider":
             os.makedirs("/config/model_cache/openvino/ort", exist_ok=True)
+            providers.append(provider)
             options.append(
                 {
+                    "arena_extend_strategy": "kSameAsRequested",
                     "cache_dir": "/config/model_cache/openvino/ort",
                     "device_type": openvino_device,
                 }
             )
+        elif provider == "CPUExecutionProvider":
+            providers.append(provider)
+            options.append(
+                {
+                    "enable_cpu_mem_arena": False,
+                }
+            )
         else:
+            providers.append(provider)
             options.append({})
 
     return (providers, options)
+
+
+class ONNXModelRunner:
+    """Run onnx models optimally based on available hardware."""
+
+    def __init__(self, model_path: str, device: str, requires_fp16: bool = False):
+        self.model_path = model_path
+        self.ort: ort.InferenceSession = None
+        self.ov: ov.Core = None
+        providers, options = get_ort_providers(device == "CPU", device, requires_fp16)
+
+        if "OpenVINOExecutionProvider" in providers:
+            # use OpenVINO directly
+            self.type = "ov"
+            self.ov = ov.Core()
+            self.ov.set_property(
+                {ov.properties.cache_dir: "/config/model_cache/openvino"}
+            )
+            self.interpreter = self.ov.compile_model(
+                model=model_path, device_name=device
+            )
+        else:
+            # Use ONNXRuntime
+            self.type = "ort"
+            self.ort = ort.InferenceSession(
+                model_path,
+                providers=providers,
+                provider_options=options,
+            )
+
+    def get_input_names(self) -> list[str]:
+        if self.type == "ov":
+            input_names = []
+
+            for input in self.interpreter.inputs:
+                input_names.extend(input.names)
+
+            return input_names
+        elif self.type == "ort":
+            return [input.name for input in self.ort.get_inputs()]
+
+    def run(self, input: dict[str, Any]) -> Any:
+        if self.type == "ov":
+            infer_request = self.interpreter.create_infer_request()
+            input_tensor = list(input.values())
+
+            if len(input_tensor) == 1:
+                input_tensor = ov.Tensor(array=input_tensor[0])
+            else:
+                input_tensor = ov.Tensor(array=input_tensor)
+
+            infer_request.infer(input_tensor)
+            return [infer_request.get_output_tensor().data]
+        elif self.type == "ort":
+            return self.ort.run(None, input)

frigate/util/services.py

@@ -257,6 +257,40 @@ def get_amd_gpu_stats() -> dict[str, str]:
 
 def get_intel_gpu_stats() -> dict[str, str]:
     """Get stats using intel_gpu_top."""
+
+    def get_stats_manually(output: str) -> dict[str, str]:
+        """Find global stats via regex when json fails to parse."""
+        reading = "".join(output)
+        results: dict[str, str] = {}
+
+        # render is used for qsv
+        render = []
+        for result in re.findall(r'"Render/3D/0":{[a-z":\d.,%]+}', reading):
+            packet = json.loads(result[14:])
+            single = packet.get("busy", 0.0)
+            render.append(float(single))
+
+        if render:
+            render_avg = sum(render) / len(render)
+        else:
+            render_avg = 1
+
+        # video is used for vaapi
+        video = []
+        for result in re.findall(r'"Video/\d":{[a-z":\d.,%]+}', reading):
+            packet = json.loads(result[10:])
+            single = packet.get("busy", 0.0)
+            video.append(float(single))
+
+        if video:
+            video_avg = sum(video) / len(video)
+        else:
+            video_avg = 1
+
+        results["gpu"] = f"{round((video_avg + render_avg) / 2, 2)}%"
+        results["mem"] = "-%"
+        return results
+
     intel_gpu_top_command = [
         "timeout",
         "0.5s",
@@ -279,7 +313,13 @@ def get_intel_gpu_stats() -> dict[str, str]:
         logger.error(f"Unable to poll intel GPU stats: {p.stderr}")
         return None
     else:
-        data = json.loads(f'[{"".join(p.stdout.split())}]')
+        output = "".join(p.stdout.split())
+
+        try:
+            data = json.loads(f"[{output}]")
+        except json.JSONDecodeError:
+            return get_stats_manually(output)
+
         results: dict[str, str] = {}
         render = {"global": []}
         video = {"global": []}
@@ -318,16 +358,17 @@ def get_intel_gpu_stats() -> dict[str, str]:
                     if video_frame is not None:
                         video[key].append(float(video_frame))
 
-        results["gpu"] = (
-            f"{round(((sum(render['global']) / len(render['global'])) + (sum(video['global']) / len(video['global']))) / 2, 2)}%"
-        )
-        results["mem"] = "-%"
+        if render["global"]:
+            results["gpu"] = (
+                f"{round(((sum(render['global']) / len(render['global'])) + (sum(video['global']) / len(video['global']))) / 2, 2)}%"
+            )
+            results["mem"] = "-%"
 
         if len(render.keys()) > 1:
             results["clients"] = {}
 
             for key in render.keys():
-                if key == "global":
+                if key == "global" or not render[key] or not video[key]:
                     continue
 
                 results["clients"][key] = (

frigate/video.py

@@ -27,7 +27,7 @@ from frigate.object_detection import RemoteObjectDetector
 from frigate.ptz.autotrack import ptz_moving_at_frame_time
 from frigate.track import ObjectTracker
 from frigate.track.norfair_tracker import NorfairTracker
-from frigate.track.object_attribute import ObjectAttribute
+from frigate.track.tracked_object import TrackedObjectAttribute
 from frigate.util.builtin import EventsPerSecond, get_tomorrow_at_time
 from frigate.util.image import (
     FrameManager,
@@ -734,10 +734,10 @@ def process_frames(
                 object_tracker.update_frame_times(frame_time)
 
         # group the attribute detections based on what label they apply to
-        attribute_detections: dict[str, list[ObjectAttribute]] = {}
+        attribute_detections: dict[str, list[TrackedObjectAttribute]] = {}
         for label, attribute_labels in model_config.attributes_map.items():
             attribute_detections[label] = [
-                ObjectAttribute(d)
+                TrackedObjectAttribute(d)
                 for d in consolidated_detections
                 if d[0] in attribute_labels
             ]

web/package.json

@@ -16,28 +16,28 @@
   "dependencies": {
     "@cycjimmy/jsmpeg-player": "^6.1.1",
     "@hookform/resolvers": "^3.9.0",
-    "@radix-ui/react-alert-dialog": "^1.1.1",
+    "@radix-ui/react-alert-dialog": "^1.1.2",
     "@radix-ui/react-aspect-ratio": "^1.1.0",
-    "@radix-ui/react-checkbox": "^1.1.1",
-    "@radix-ui/react-context-menu": "^2.2.1",
-    "@radix-ui/react-dialog": "^1.1.1",
-    "@radix-ui/react-dropdown-menu": "^2.1.1",
-    "@radix-ui/react-hover-card": "^1.1.1",
+    "@radix-ui/react-checkbox": "^1.1.2",
+    "@radix-ui/react-context-menu": "^2.2.2",
+    "@radix-ui/react-dialog": "^1.1.2",
+    "@radix-ui/react-dropdown-menu": "^2.1.2",
+    "@radix-ui/react-hover-card": "^1.1.2",
     "@radix-ui/react-label": "^2.1.0",
-    "@radix-ui/react-popover": "^1.1.1",
-    "@radix-ui/react-radio-group": "^1.2.0",
-    "@radix-ui/react-scroll-area": "^1.1.0",
-    "@radix-ui/react-select": "^2.1.1",
+    "@radix-ui/react-popover": "^1.1.2",
+    "@radix-ui/react-radio-group": "^1.2.1",
+    "@radix-ui/react-scroll-area": "^1.2.0",
+    "@radix-ui/react-select": "^2.1.2",
     "@radix-ui/react-separator": "^1.1.0",
-    "@radix-ui/react-slider": "^1.2.0",
+    "@radix-ui/react-slider": "^1.2.1",
     "@radix-ui/react-slot": "^1.1.0",
-    "@radix-ui/react-switch": "^1.1.0",
-    "@radix-ui/react-tabs": "^1.1.0",
+    "@radix-ui/react-switch": "^1.1.1",
+    "@radix-ui/react-tabs": "^1.1.1",
     "@radix-ui/react-toggle": "^1.1.0",
     "@radix-ui/react-toggle-group": "^1.1.0",
-    "@radix-ui/react-tooltip": "^1.1.2",
+    "@radix-ui/react-tooltip": "^1.1.3",
     "apexcharts": "^3.52.0",
-    "axios": "^1.7.3",
+    "axios": "^1.7.7",
     "class-variance-authority": "^0.7.0",
     "clsx": "^2.1.1",
     "cmdk": "^1.0.0",
@@ -45,10 +45,10 @@
     "date-fns": "^3.6.0",
     "embla-carousel-react": "^8.2.0",
     "framer-motion": "^11.5.4",
-    "hls.js": "^1.5.14",
+    "hls.js": "^1.5.17",
     "idb-keyval": "^6.2.1",
     "immer": "^10.1.1",
-    "konva": "^9.3.14",
+    "konva": "^9.3.16",
     "lodash": "^4.17.21",
     "lucide-react": "^0.407.0",
     "monaco-yaml": "^5.2.2",
@@ -65,7 +65,7 @@
     "react-konva": "^18.2.10",
     "react-router-dom": "^6.26.0",
     "react-swipeable": "^7.0.1",
-    "react-tracked": "^2.0.0",
+    "react-tracked": "^2.0.1",
     "react-transition-group": "^4.4.5",
     "react-use-websocket": "^4.8.1",
     "react-zoom-pan-pinch": "3.4.4",
@@ -83,9 +83,9 @@
     "zod": "^3.23.8"
   },
   "devDependencies": {
-    "@tailwindcss/forms": "^0.5.7",
-    "@testing-library/jest-dom": "^6.4.6",
-    "@types/lodash": "^4.17.7",
+    "@tailwindcss/forms": "^0.5.9",
+    "@testing-library/jest-dom": "^6.6.2",
+    "@types/lodash": "^4.17.12",
     "@types/node": "^20.14.10",
     "@types/react": "^18.3.2",
     "@types/react-dom": "^18.3.0",
@@ -95,7 +95,7 @@
     "@types/strftime": "^0.9.8",
     "@typescript-eslint/eslint-plugin": "^7.5.0",
     "@typescript-eslint/parser": "^7.5.0",
-    "@vitejs/plugin-react-swc": "^3.6.0",
+    "@vitejs/plugin-react-swc": "^3.7.1",
     "@vitest/coverage-v8": "^2.0.5",
     "autoprefixer": "^10.4.20",
     "eslint": "^8.57.0",
@@ -109,8 +109,8 @@
     "jest-websocket-mock": "^2.5.0",
     "jsdom": "^24.1.1",
     "msw": "^2.3.5",
-    "postcss": "^8.4.39",
-    "prettier": "^3.3.2",
+    "postcss": "^8.4.47",
+    "prettier": "^3.3.3",
     "prettier-plugin-tailwindcss": "^0.6.5",
     "tailwindcss": "^3.4.9",
     "typescript": "^5.5.4",

web/src/api/ws.tsx

@@ -2,6 +2,7 @@ import { baseUrl } from "./baseUrl";
 import { useCallback, useEffect, useState } from "react";
 import useWebSocket, { ReadyState } from "react-use-websocket";
 import {
+  EmbeddingsReindexProgressType,
   FrigateCameraState,
   FrigateEvent,
   FrigateReview,
@@ -64,7 +65,10 @@ function useValue(): useValueReturn {
         : "OFF";
     });
 
-    setWsState({ ...wsState, ...cameraStates });
+    setWsState((prevState) => ({
+      ...prevState,
+      ...cameraStates,
+    }));
     setHasCameraState(true);
     // we only want this to run initially when the config is loaded
     // eslint-disable-next-line react-hooks/exhaustive-deps
@@ -76,7 +80,10 @@ function useValue(): useValueReturn {
       const data: Update = JSON.parse(event.data);
 
       if (data) {
-        setWsState({ ...wsState, [data.topic]: data.payload });
+        setWsState((prevState) => ({
+          ...prevState,
+          [data.topic]: data.payload,
+        }));
       }
     },
     onOpen: () => {
@@ -302,6 +309,42 @@ export function useModelState(
   return { payload: data ? data[model] : undefined };
 }
 
+export function useEmbeddingsReindexProgress(
+  revalidateOnFocus: boolean = true,
+): {
+  payload: EmbeddingsReindexProgressType;
+} {
+  const {
+    value: { payload },
+    send: sendCommand,
+  } = useWs("embeddings_reindex_progress", "embeddingsReindexProgress");
+
+  const data = useDeepMemo(JSON.parse(payload as string));
+
+  useEffect(() => {
+    let listener = undefined;
+    if (revalidateOnFocus) {
+      sendCommand("embeddingsReindexProgress");
+      listener = () => {
+        if (document.visibilityState == "visible") {
+          sendCommand("embeddingsReindexProgress");
+        }
+      };
+      addEventListener("visibilitychange", listener);
+    }
+
+    return () => {
+      if (listener) {
+        removeEventListener("visibilitychange", listener);
+      }
+    };
+    // we know that these deps are correct
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [revalidateOnFocus]);
+
+  return { payload: data };
+}
+
 export function useMotionActivity(camera: string): { payload: string } {
   const {
     value: { payload },

web/src/components/Statusbar.tsx

@@ -1,3 +1,4 @@
+import { useEmbeddingsReindexProgress } from "@/api/ws";
 import {
   StatusBarMessagesContext,
   StatusMessage,
@@ -41,6 +42,23 @@ export default function Statusbar() {
     });
   }, [potentialProblems, addMessage, clearMessages]);
 
+  const { payload: reindexState } = useEmbeddingsReindexProgress();
+
+  useEffect(() => {
+    if (reindexState) {
+      if (reindexState.status == "indexing") {
+        clearMessages("embeddings-reindex");
+        addMessage(
+          "embeddings-reindex",
+          `Reindexing embeddings (${Math.floor((reindexState.processed_objects / reindexState.total_objects) * 100)}% complete)`,
+        );
+      }
+      if (reindexState.status === "completed") {
+        clearMessages("embeddings-reindex");
+      }
+    }
+  }, [reindexState, addMessage, clearMessages]);
+
   return (
     <div className="absolute bottom-0 left-0 right-0 z-10 flex h-8 w-full items-center justify-between border-t border-secondary-highlight bg-background_alt px-4 dark:text-secondary-foreground">
       <div className="flex h-full items-center gap-2">

web/src/components/auth/AuthForm.tsx

@@ -121,6 +121,7 @@ export function UserAuthForm({ className, ...props }: UserAuthFormProps) {
               variant="select"
               disabled={isLoading}
               className="flex flex-1"
+              aria-label="Login"
             >
               {isLoading && <ActivityIndicator className="mr-2 h-4 w-4" />}
               Login

web/src/components/button/DownloadVideoButton.tsx

@@ -0,0 +1,59 @@
+import { useState } from "react";
+import { Button } from "@/components/ui/button";
+import { toast } from "sonner";
+import ActivityIndicator from "../indicators/activity-indicator";
+import { FaDownload } from "react-icons/fa";
+import { formatUnixTimestampToDateTime } from "@/utils/dateUtil";
+import { cn } from "@/lib/utils";
+
+type DownloadVideoButtonProps = {
+  source: string;
+  camera: string;
+  startTime: number;
+  className?: string;
+};
+
+export function DownloadVideoButton({
+  source,
+  camera,
+  startTime,
+  className,
+}: DownloadVideoButtonProps) {
+  const [isDownloading, setIsDownloading] = useState(false);
+
+  const formattedDate = formatUnixTimestampToDateTime(startTime, {
+    strftime_fmt: "%D-%T",
+    time_style: "medium",
+    date_style: "medium",
+  });
+  const filename = `${camera}_${formattedDate}.mp4`;
+
+  const handleDownloadStart = () => {
+    setIsDownloading(true);
+    toast.success("Your review item video has started downloading.", {
+      position: "top-center",
+    });
+  };
+
+  return (
+    <div className="flex justify-center">
+      <Button
+        asChild
+        disabled={isDownloading}
+        className="flex items-center gap-2"
+        size="sm"
+        aria-label="Download Video"
+      >
+        <a href={source} download={filename} onClick={handleDownloadStart}>
+          {isDownloading ? (
+            <ActivityIndicator className="size-4" />
+          ) : (
+            <FaDownload
+              className={cn("size-4 text-secondary-foreground", className)}
+            />
+          )}
+        </a>
+      </Button>
+    </div>
+  );
+}

web/src/components/camera/DebugCameraImage.tsx

@@ -55,7 +55,12 @@ export default function DebugCameraImage({
         searchParams={searchParams}
         cameraClasses="relative w-full h-full flex justify-center"
       />
-      <Button onClick={handleToggleSettings} variant="link" size="sm">
+      <Button
+        onClick={handleToggleSettings}
+        variant="link"
+        size="sm"
+        aria-label="Settings"
+      >
         <span className="h-5 w-5">
           <LuSettings />
         </span>{" "}

web/src/components/card/AnimatedEventCard.tsx

@@ -121,6 +121,7 @@ export function AnimatedEventCard({
                 <Button
                   className="absolute right-2 top-1 z-40 bg-gray-500 bg-gradient-to-br from-gray-400 to-gray-500"
                   size="xs"
+                  aria-label="Mark as Reviewed"
                   onClick={async () => {
                     await axios.post(`reviews/viewed`, { ids: [event.id] });
                     updateEvents();

web/src/components/card/ExportCard.tsx

@@ -113,6 +113,7 @@ export default function ExportCard({
               />
               <DialogFooter>
                 <Button
+                  aria-label="Save Export"
                   size="sm"
                   variant="select"
                   disabled={(editName?.update?.length ?? 0) == 0}
@@ -206,6 +207,7 @@ export default function ExportCard({
             {!exportedRecording.in_progress && (
               <Button
                 className="absolute left-1/2 top-1/2 h-20 w-20 -translate-x-1/2 -translate-y-1/2 cursor-pointer text-white hover:bg-transparent hover:text-white"
+                aria-label="Play"
                 variant="ghost"
                 onClick={() => {
                   onSelect(exportedRecording);

web/src/components/card/ReviewCard.tsx

@@ -34,6 +34,7 @@ import { toast } from "sonner";
 import useKeyboardListener from "@/hooks/use-keyboard-listener";
 import { Tooltip, TooltipContent, TooltipTrigger } from "../ui/tooltip";
 import { capitalizeFirstLetter } from "@/utils/stringUtil";
+import { buttonVariants } from "../ui/button";
 
 type ReviewCardProps = {
   event: ReviewSegment;
@@ -228,7 +229,10 @@ export default function ReviewCard({
               <AlertDialogCancel onClick={() => setOptionsOpen(false)}>
                 Cancel
               </AlertDialogCancel>
-              <AlertDialogAction className="bg-destructive" onClick={onDelete}>
+              <AlertDialogAction
+                className={buttonVariants({ variant: "destructive" })}
+                onClick={onDelete}
+              >
                 Delete
               </AlertDialogAction>
             </AlertDialogFooter>
@@ -295,7 +299,10 @@ export default function ReviewCard({
             <AlertDialogCancel onClick={() => setOptionsOpen(false)}>
               Cancel
             </AlertDialogCancel>
-            <AlertDialogAction className="bg-destructive" onClick={onDelete}>
+            <AlertDialogAction
+              className={buttonVariants({ variant: "destructive" })}
+              onClick={onDelete}
+            >
               Delete
             </AlertDialogAction>
           </AlertDialogFooter>

web/src/components/card/SearchThumbnail.tsx

@@ -1,50 +1,56 @@
-import { useCallback } from "react";
+import { useCallback, useMemo } from "react";
 import { useApiHost } from "@/api";
 import { getIconForLabel } from "@/utils/iconUtil";
-import TimeAgo from "../dynamic/TimeAgo";
 import useSWR from "swr";
 import { FrigateConfig } from "@/types/frigateConfig";
 import { isIOS, isSafari } from "react-device-detect";
 import Chip from "@/components/indicators/Chip";
-import { useFormattedTimestamp } from "@/hooks/use-date-utils";
 import useImageLoaded from "@/hooks/use-image-loaded";
 import { Tooltip, TooltipContent, TooltipTrigger } from "../ui/tooltip";
 import ImageLoadingIndicator from "../indicators/ImageLoadingIndicator";
-import ActivityIndicator from "../indicators/activity-indicator";
 import { capitalizeFirstLetter } from "@/utils/stringUtil";
 import { SearchResult } from "@/types/search";
-import useContextMenu from "@/hooks/use-contextmenu";
 import { cn } from "@/lib/utils";
 import { TooltipPortal } from "@radix-ui/react-tooltip";
 
 type SearchThumbnailProps = {
   searchResult: SearchResult;
-  findSimilar: () => void;
   onClick: (searchResult: SearchResult) => void;
 };
 
 export default function SearchThumbnail({
   searchResult,
-  findSimilar,
   onClick,
 }: SearchThumbnailProps) {
   const apiHost = useApiHost();
   const { data: config } = useSWR<FrigateConfig>("config");
   const [imgRef, imgLoaded, onImgLoad] = useImageLoaded();
 
-  useContextMenu(imgRef, findSimilar);
+  // interactions
 
   const handleOnClick = useCallback(() => {
     onClick(searchResult);
   }, [searchResult, onClick]);
 
-  // date
+  const objectLabel = useMemo(() => {
+    if (
+      !config ||
+      !searchResult.sub_label ||
+      !config.model.attributes_map[searchResult.label]
+    ) {
+      return searchResult.label;
+    }
 
-  const formattedDate = useFormattedTimestamp(
-    searchResult.start_time,
-    config?.ui.time_format == "24hour" ? "%b %-d, %H:%M" : "%b %-d, %I:%M %p",
-    config?.ui.timezone,
-  );
+    if (
+      config.model.attributes_map[searchResult.label].includes(
+        searchResult.sub_label,
+      )
+    ) {
+      return searchResult.sub_label;
+    }
+
+    return `${searchResult.label}-verified`;
+  }, [config, searchResult]);
 
   return (
     <div className="relative size-full cursor-pointer" onClick={handleOnClick}>
@@ -80,17 +86,23 @@ export default function SearchThumbnail({
               <TooltipTrigger asChild>
                 <div className="mx-3 pb-1 text-sm text-white">
                   <Chip
-                    className={`z-0 flex items-start justify-between space-x-1 bg-gray-500 bg-gradient-to-br from-gray-400 to-gray-500`}
+                    className={`z-0 flex items-center justify-between gap-1 space-x-1 bg-gray-500 bg-gradient-to-br from-gray-400 to-gray-500 text-xs`}
                     onClick={() => onClick(searchResult)}
                   >
-                    {getIconForLabel(searchResult.label, "size-3 text-white")}
+                    {getIconForLabel(objectLabel, "size-3 text-white")}
+                    {Math.round(
+                      (searchResult.data.score ??
+                        searchResult.data.top_score ??
+                        searchResult.top_score) * 100,
+                    )}
+                    %
                   </Chip>
                 </div>
               </TooltipTrigger>
             </div>
             <TooltipPortal>
               <TooltipContent className="capitalize">
-                {[...new Set([searchResult.label])]
+                {[objectLabel]
                   .filter(
                     (item) => item !== undefined && !item.includes("-verified"),
                   )
@@ -103,18 +115,7 @@ export default function SearchThumbnail({
           </Tooltip>
         </div>
         <div className="rounded-t-l pointer-events-none absolute inset-x-0 top-0 z-10 h-[30%] w-full bg-gradient-to-b from-black/60 to-transparent"></div>
-        <div className="rounded-b-l pointer-events-none absolute inset-x-0 bottom-0 z-10 h-[20%] w-full bg-gradient-to-t from-black/60 to-transparent">
-          <div className="mx-3 flex h-full items-end justify-between pb-1 text-sm text-white">
-            {searchResult.end_time ? (
-              <TimeAgo time={searchResult.start_time * 1000} dense />
-            ) : (
-              <div>
-                <ActivityIndicator size={24} />
-              </div>
-            )}
-            {formattedDate}
-          </div>
-        </div>
+        <div className="rounded-b-l pointer-events-none absolute inset-x-0 bottom-0 z-10 flex h-[20%] items-end bg-gradient-to-t from-black/60 to-transparent"></div>
       </div>
     </div>
   );

web/src/components/card/SearchThumbnailFooter.tsx

@@ -0,0 +1,62 @@
+import TimeAgo from "../dynamic/TimeAgo";
+import useSWR from "swr";
+import { FrigateConfig } from "@/types/frigateConfig";
+import { useFormattedTimestamp } from "@/hooks/use-date-utils";
+import { SearchResult } from "@/types/search";
+import ActivityIndicator from "../indicators/activity-indicator";
+import SearchResultActions from "../menu/SearchResultActions";
+import { cn } from "@/lib/utils";
+
+type SearchThumbnailProps = {
+  searchResult: SearchResult;
+  columns: number;
+  findSimilar: () => void;
+  refreshResults: () => void;
+  showObjectLifecycle: () => void;
+};
+
+export default function SearchThumbnailFooter({
+  searchResult,
+  columns,
+  findSimilar,
+  refreshResults,
+  showObjectLifecycle,
+}: SearchThumbnailProps) {
+  const { data: config } = useSWR<FrigateConfig>("config");
+
+  // date
+  const formattedDate = useFormattedTimestamp(
+    searchResult.start_time,
+    config?.ui.time_format == "24hour" ? "%b %-d, %H:%M" : "%b %-d, %I:%M %p",
+    config?.ui.timezone,
+  );
+
+  return (
+    <div
+      className={cn(
+        "flex w-full flex-row items-center justify-between",
+        columns > 4 &&
+          "items-start sm:flex-col sm:gap-2 lg:flex-row lg:items-center lg:gap-1",
+      )}
+    >
+      <div className="flex flex-col items-start text-xs text-primary-variant">
+        {searchResult.end_time ? (
+          <TimeAgo time={searchResult.start_time * 1000} dense />
+        ) : (
+          <div>
+            <ActivityIndicator size={14} />
+          </div>
+        )}
+        {formattedDate}
+      </div>
+      <div className="flex flex-row items-center justify-end gap-6 md:gap-4">
+        <SearchResultActions
+          searchResult={searchResult}
+          findSimilar={findSimilar}
+          refreshResults={refreshResults}
+          showObjectLifecycle={showObjectLifecycle}
+        />
+      </div>
+    </div>
+  );
+}

web/src/components/dynamic/NewReviewData.tsx

@@ -36,6 +36,7 @@ export default function NewReviewData({
               : "invisible",
             "mx-auto mt-5 bg-gray-400 text-center text-white",
           )}
+          aria-label="View new review items"
           onClick={() => {
             pullLatestData();
             if (contentRef.current) {

web/src/components/filter/CalendarFilterButton.tsx

@@ -34,6 +34,7 @@ export default function CalendarFilterButton({
   const trigger = (
     <Button
       className="flex items-center gap-2"
+      aria-label="Select a date to filter by"
       variant={day == undefined ? "default" : "select"}
       size="sm"
     >
@@ -57,6 +58,7 @@ export default function CalendarFilterButton({
       <DropdownMenuSeparator />
       <div className="flex items-center justify-center p-2">
         <Button
+          aria-label="Reset"
           onClick={() => {
             updateSelectedDay(undefined);
           }}
@@ -99,6 +101,7 @@ export function CalendarRangeFilterButton({
   const trigger = (
     <Button
       className="flex items-center gap-2"
+      aria-label="Select a date to filter by"
       variant={range == undefined ? "default" : "select"}
       size="sm"
     >

web/src/components/filter/CameraGroupSelector.tsx

@@ -3,7 +3,7 @@ import { isDesktop, isMobile } from "react-device-detect";
 import useSWR from "swr";
 import { MdHome } from "react-icons/md";
 import { usePersistedOverlayState } from "@/hooks/use-overlay-state";
-import { Button } from "../ui/button";
+import { Button, buttonVariants } from "../ui/button";
 import { useCallback, useMemo, useState } from "react";
 import { Tooltip, TooltipContent, TooltipTrigger } from "../ui/tooltip";
 import { LuPencil, LuPlus } from "react-icons/lu";
@@ -141,6 +141,7 @@ export function CameraGroupSelector({ className }: CameraGroupSelectorProps) {
                     ? "bg-blue-900 bg-opacity-60 text-selected focus:bg-blue-900 focus:bg-opacity-60"
                     : "bg-secondary text-secondary-foreground focus:bg-secondary focus:text-secondary-foreground"
                 }
+                aria-label="All Cameras"
                 size="xs"
                 onClick={() => (group ? setGroup("default", true) : null)}
                 onMouseEnter={() => (isDesktop ? showTooltip("default") : null)}
@@ -165,6 +166,7 @@ export function CameraGroupSelector({ className }: CameraGroupSelectorProps) {
                         ? "bg-blue-900 bg-opacity-60 text-selected focus:bg-blue-900 focus:bg-opacity-60"
                         : "bg-secondary text-secondary-foreground"
                     }
+                    aria-label="Camera Group"
                     size="xs"
                     onClick={() => setGroup(name, group != "default")}
                     onMouseEnter={() => (isDesktop ? showTooltip(name) : null)}
@@ -191,6 +193,7 @@ export function CameraGroupSelector({ className }: CameraGroupSelectorProps) {
 
           <Button
             className="bg-secondary text-muted-foreground"
+            aria-label="Add camera group"
             size="xs"
             onClick={() => setAddGroup(true)}
           >
@@ -355,6 +358,7 @@ function NewGroupDialog({
                         "size-6 rounded-md bg-secondary-foreground p-1 text-background",
                       isMobile && "text-secondary-foreground",
                     )}
+                    aria-label="Add camera group"
                     onClick={() => {
                       setEditState("add");
                     }}
@@ -518,7 +522,10 @@ export function CameraGroupRow({
             </AlertDialogDescription>
             <AlertDialogFooter>
               <AlertDialogCancel>Cancel</AlertDialogCancel>
-              <AlertDialogAction onClick={onDeleteGroup}>
+              <AlertDialogAction
+                className={buttonVariants({ variant: "destructive" })}
+                onClick={onDeleteGroup}
+              >
                 Delete
               </AlertDialogAction>
             </AlertDialogFooter>
@@ -533,10 +540,16 @@ export function CameraGroupRow({
               </DropdownMenuTrigger>
               <DropdownMenuPortal>
                 <DropdownMenuContent>
-                  <DropdownMenuItem onClick={onEditGroup}>
+                  <DropdownMenuItem
+                    aria-label="Edit group"
+                    onClick={onEditGroup}
+                  >
                     Edit
                   </DropdownMenuItem>
-                  <DropdownMenuItem onClick={() => setDeleteDialogOpen(true)}>
+                  <DropdownMenuItem
+                    aria-label="Delete group"
+                    onClick={() => setDeleteDialogOpen(true)}
+                  >
                     Delete
                   </DropdownMenuItem>
                 </DropdownMenuContent>
@@ -643,6 +656,11 @@ export function CameraGroupEdit({
 
       setIsLoading(true);
 
+      let renamingQuery = "";
+      if (editingGroup && editingGroup[0] !== values.name) {
+        renamingQuery = `camera_groups.${editingGroup[0]}&`;
+      }
+
       const order =
         editingGroup === undefined
           ? currentGroups.length + 1
@@ -655,9 +673,12 @@ export function CameraGroupEdit({
         .join("");
 
       axios
-        .put(`config/set?${orderQuery}&${iconQuery}${cameraQueries}`, {
-          requires_restart: 0,
-        })
+        .put(
+          `config/set?${renamingQuery}${orderQuery}&${iconQuery}${cameraQueries}`,
+          {
+            requires_restart: 0,
+          },
+        )
         .then((res) => {
           if (res.status === 200) {
             toast.success(`Camera group (${values.name}) has been saved.`, {
@@ -712,7 +733,6 @@ export function CameraGroupEdit({
                 <Input
                   className="text-md w-full border border-input bg-background p-2 hover:bg-accent hover:text-accent-foreground dark:[color-scheme:dark]"
                   placeholder="Enter a name..."
-                  disabled={editingGroup !== undefined}
                   {...field}
                 />
               </FormControl>
@@ -783,13 +803,19 @@ export function CameraGroupEdit({
         <Separator className="my-2 flex bg-secondary" />
 
         <div className="flex flex-row gap-2 py-5 md:pb-0">
-          <Button type="button" className="flex flex-1" onClick={onCancel}>
+          <Button
+            type="button"
+            className="flex flex-1"
+            aria-label="Cancel"
+            onClick={onCancel}
+          >
             Cancel
           </Button>
           <Button
             variant="select"
             disabled={isLoading}
             className="flex flex-1"
+            aria-label="Save"
             type="submit"
           >
             {isLoading ? (

web/src/components/filter/CamerasFilterButton.tsx

@@ -55,6 +55,7 @@ export function CamerasFilterButton({
   const trigger = (
     <Button
       className="flex items-center gap-2 capitalize"
+      aria-label="Cameras Filter"
       variant={selectedCameras?.length == undefined ? "default" : "select"}
       size="sm"
     >
@@ -69,6 +70,70 @@ export function CamerasFilterButton({
     </Button>
   );
   const content = (
+    <CamerasFilterContent
+      allCameras={allCameras}
+      groups={groups}
+      currentCameras={currentCameras}
+      setCurrentCameras={setCurrentCameras}
+      setOpen={setOpen}
+      updateCameraFilter={updateCameraFilter}
+    />
+  );
+
+  if (isMobile) {
+    return (
+      <Drawer
+        open={open}
+        onOpenChange={(open) => {
+          if (!open) {
+            setCurrentCameras(selectedCameras);
+          }
+
+          setOpen(open);
+        }}
+      >
+        <DrawerTrigger asChild>{trigger}</DrawerTrigger>
+        <DrawerContent className="max-h-[75dvh] overflow-hidden">
+          {content}
+        </DrawerContent>
+      </Drawer>
+    );
+  }
+
+  return (
+    <DropdownMenu
+      modal={false}
+      open={open}
+      onOpenChange={(open) => {
+        if (!open) {
+          setCurrentCameras(selectedCameras);
+        }
+        setOpen(open);
+      }}
+    >
+      <DropdownMenuTrigger asChild>{trigger}</DropdownMenuTrigger>
+      <DropdownMenuContent>{content}</DropdownMenuContent>
+    </DropdownMenu>
+  );
+}
+
+type CamerasFilterContentProps = {
+  allCameras: string[];
+  currentCameras: string[] | undefined;
+  groups: [string, CameraGroupConfig][];
+  setCurrentCameras: (cameras: string[] | undefined) => void;
+  setOpen: (open: boolean) => void;
+  updateCameraFilter: (cameras: string[] | undefined) => void;
+};
+export function CamerasFilterContent({
+  allCameras,
+  currentCameras,
+  groups,
+  setCurrentCameras,
+  setOpen,
+  updateCameraFilter,
+}: CamerasFilterContentProps) {
+  return (
     <>
       {isMobile && (
         <>
@@ -138,6 +203,7 @@ export function CamerasFilterButton({
       <DropdownMenuSeparator />
       <div className="flex items-center justify-evenly p-2">
         <Button
+          aria-label="Apply"
           variant="select"
           disabled={currentCameras?.length === 0}
           onClick={() => {
@@ -148,6 +214,7 @@ export function CamerasFilterButton({
           Apply
         </Button>
         <Button
+          aria-label="Reset"
           onClick={() => {
             setCurrentCameras(undefined);
             updateCameraFilter(undefined);
@@ -158,40 +225,4 @@ export function CamerasFilterButton({
       </div>
     </>
   );
-
-  if (isMobile) {
-    return (
-      <Drawer
-        open={open}
-        onOpenChange={(open) => {
-          if (!open) {
-            setCurrentCameras(selectedCameras);
-          }
-
-          setOpen(open);
-        }}
-      >
-        <DrawerTrigger asChild>{trigger}</DrawerTrigger>
-        <DrawerContent className="max-h-[75dvh] overflow-hidden">
-          {content}
-        </DrawerContent>
-      </Drawer>
-    );
-  }
-
-  return (
-    <DropdownMenu
-      modal={false}
-      open={open}
-      onOpenChange={(open) => {
-        if (!open) {
-          setCurrentCameras(selectedCameras);
-        }
-        setOpen(open);
-      }}
-    >
-      <DropdownMenuTrigger asChild>{trigger}</DropdownMenuTrigger>
-      <DropdownMenuContent>{content}</DropdownMenuContent>
-    </DropdownMenu>
-  );
 }

web/src/components/filter/LogLevelFilter.tsx

@@ -17,7 +17,11 @@ export function LogLevelFilterButton({
   updateLabelFilter,
 }: LogLevelFilterButtonProps) {
   const trigger = (
-    <Button size="sm" className="flex items-center gap-2">
+    <Button
+      size="sm"
+      className="flex items-center gap-2"
+      aria-label="Filter log level"
+    >
       <FaFilter className="text-secondary-foreground" />
       <div className="hidden text-primary md:block">Filter</div>
     </Button>

web/src/components/filter/ReviewActionGroup.tsx

@@ -1,7 +1,7 @@
 import { FaCircleCheck } from "react-icons/fa6";
 import { useCallback, useState } from "react";
 import axios from "axios";
-import { Button } from "../ui/button";
+import { Button, buttonVariants } from "../ui/button";
 import { isDesktop } from "react-device-detect";
 import { FaCompactDisc } from "react-icons/fa";
 import { HiTrash } from "react-icons/hi";
@@ -79,7 +79,10 @@ export default function ReviewActionGroup({
           </AlertDialogDescription>
           <AlertDialogFooter>
             <AlertDialogCancel>Cancel</AlertDialogCancel>
-            <AlertDialogAction className="bg-destructive" onClick={onDelete}>
+            <AlertDialogAction
+              className={buttonVariants({ variant: "destructive" })}
+              onClick={onDelete}
+            >
               Delete
             </AlertDialogAction>
           </AlertDialogFooter>
@@ -101,6 +104,7 @@ export default function ReviewActionGroup({
           {selectedReviews.length == 1 && (
             <Button
               className="flex items-center gap-2 p-2"
+              aria-label="Export"
               size="sm"
               onClick={() => {
                 onExport(selectedReviews[0]);
@@ -113,6 +117,7 @@ export default function ReviewActionGroup({
           )}
           <Button
             className="flex items-center gap-2 p-2"
+            aria-label="Mark as reviewed"
             size="sm"
             onClick={onMarkAsReviewed}
           >
@@ -121,6 +126,7 @@ export default function ReviewActionGroup({
           </Button>
           <Button
             className="flex items-center gap-2 p-2"
+            aria-label="Delete"
             size="sm"
             onClick={handleDelete}
           >

web/src/components/filter/ReviewFilterGroup.tsx

@@ -241,6 +241,8 @@ export default function ReviewFilterGroup({
           mode="none"
           setMode={() => {}}
           setRange={() => {}}
+          showExportPreview={false}
+          setShowExportPreview={() => {}}
         />
       )}
     </div>
@@ -276,6 +278,7 @@ function ShowReviewFilter({
 
       <Button
         className="block duration-0 md:hidden"
+        aria-label="Show reviewed"
         variant={showReviewedSwitch ? "select" : "default"}
         size="sm"
         onClick={() =>
@@ -336,6 +339,7 @@ function GeneralFilterButton({
         selectedLabels?.length || selectedZones?.length ? "select" : "default"
       }
       className="flex items-center gap-2 capitalize"
+      aria-label="Filter"
     >
       <FaFilter
         className={`${selectedLabels?.length || selectedZones?.length ? "text-selected-foreground" : "text-secondary-foreground"}`}
@@ -536,6 +540,7 @@ export function GeneralFilterContent({
       <DropdownMenuSeparator />
       <div className="flex items-center justify-evenly p-2">
         <Button
+          aria-label="Apply"
           variant="select"
           onClick={() => {
             if (selectedLabels != currentLabels) {
@@ -552,6 +557,7 @@ export function GeneralFilterContent({
           Apply
         </Button>
         <Button
+          aria-label="Reset"
           onClick={() => {
             setCurrentLabels(undefined);
             setCurrentZones?.(undefined);
@@ -599,6 +605,7 @@ function ShowMotionOnlyButton({
         <Button
           size="sm"
           className="duration-0"
+          aria-label="Show Motion Only"
           variant={motionOnlyButton ? "select" : "default"}
           onClick={() => setMotionOnlyButton(!motionOnlyButton)}
         >

web/src/components/filter/SearchFilterGroup.tsx

@@ -1,5 +1,4 @@
 import { Button } from "../ui/button";
-import { Popover, PopoverContent, PopoverTrigger } from "../ui/popover";
 import useSWR from "swr";
 import { FrigateConfig } from "@/types/frigateConfig";
 import { useCallback, useEffect, useMemo, useState } from "react";
@@ -10,25 +9,19 @@ import { Switch } from "../ui/switch";
 import { Label } from "../ui/label";
 import FilterSwitch from "./FilterSwitch";
 import { FilterList } from "@/types/filter";
-import { CalendarRangeFilterButton } from "./CalendarFilterButton";
 import { CamerasFilterButton } from "./CamerasFilterButton";
 import {
   DEFAULT_SEARCH_FILTERS,
   SearchFilter,
   SearchFilters,
   SearchSource,
-  DEFAULT_TIME_RANGE_AFTER,
-  DEFAULT_TIME_RANGE_BEFORE,
 } from "@/types/search";
 import { DateRange } from "react-day-picker";
 import { cn } from "@/lib/utils";
-import SubFilterIcon from "../icons/SubFilterIcon";
-import { FaLocationDot } from "react-icons/fa6";
 import { MdLabel } from "react-icons/md";
-import SearchSourceIcon from "../icons/SearchSourceIcon";
 import PlatformAwareDialog from "../overlay/dialog/PlatformAwareDialog";
-import { FaArrowRight, FaClock } from "react-icons/fa";
-import { useFormattedHour } from "@/hooks/use-date-utils";
+import SearchFilterDialog from "../overlay/dialog/SearchFilterDialog";
+import { CalendarRangeFilterButton } from "./CalendarFilterButton";
 
 type SearchFilterGroupProps = {
   className: string;
@@ -79,8 +72,6 @@ export default function SearchFilterGroup({
     return [...labels].sort();
   }, [config, filterList, filter]);
 
-  const { data: allSubLabels } = useSWR(["sub_labels", { split_joined: 1 }]);
-
   const allZones = useMemo<string[]>(() => {
     if (filterList?.zones) {
       return filterList.zones;
@@ -159,6 +150,15 @@ export default function SearchFilterGroup({
           }}
         />
       )}
+      {filters.includes("general") && (
+        <GeneralFilterButton
+          allLabels={filterValues.labels}
+          selectedLabels={filter?.labels}
+          updateLabelFilter={(newLabels) => {
+            onUpdateFilter({ ...filter, labels: newLabels });
+          }}
+        />
+      )}
       {filters.includes("date") && (
         <CalendarRangeFilterButton
           range={
@@ -173,54 +173,12 @@ export default function SearchFilterGroup({
           updateSelectedRange={onUpdateSelectedRange}
         />
       )}
-      {filters.includes("time") && (
-        <TimeRangeFilterButton
-          config={config}
-          timeRange={filter?.time_range}
-          updateTimeRange={(time_range) =>
-            onUpdateFilter({ ...filter, time_range })
-          }
-        />
-      )}
-      {filters.includes("zone") && allZones.length > 0 && (
-        <ZoneFilterButton
-          allZones={filterValues.zones}
-          selectedZones={filter?.zones}
-          updateZoneFilter={(newZones) =>
-            onUpdateFilter({ ...filter, zones: newZones })
-          }
-        />
-      )}
-      {filters.includes("general") && (
-        <GeneralFilterButton
-          allLabels={filterValues.labels}
-          selectedLabels={filter?.labels}
-          updateLabelFilter={(newLabels) => {
-            onUpdateFilter({ ...filter, labels: newLabels });
-          }}
-        />
-      )}
-      {filters.includes("sub") && (
-        <SubFilterButton
-          allSubLabels={allSubLabels}
-          selectedSubLabels={filter?.sub_labels}
-          updateSubLabelFilter={(newSubLabels) =>
-            onUpdateFilter({ ...filter, sub_labels: newSubLabels })
-          }
-        />
-      )}
-      {config?.semantic_search?.enabled &&
-        filters.includes("source") &&
-        !filter?.search_type?.includes("similarity") && (
-          <SearchTypeButton
-            selectedSearchSources={
-              filter?.search_type ?? ["thumbnail", "description"]
-            }
-            updateSearchSourceFilter={(newSearchSource) =>
-              onUpdateFilter({ ...filter, search_type: newSearchSource })
-            }
-          />
-        )}
+      <SearchFilterDialog
+        config={config}
+        filter={filter}
+        filterValues={filterValues}
+        onUpdateFilter={onUpdateFilter}
+      />
     </div>
   );
 }
@@ -269,6 +227,7 @@ function GeneralFilterButton({
       size="sm"
       variant={selectedLabels?.length ? "select" : "default"}
       className="flex items-center gap-2 capitalize"
+      aria-label="Labels"
     >
       <MdLabel
         className={`${selectedLabels?.length ? "text-selected-foreground" : "text-secondary-foreground"}`}
@@ -295,7 +254,11 @@ function GeneralFilterButton({
     <PlatformAwareDialog
       trigger={trigger}
       content={content}
-      contentClassName={isDesktop ? "" : "max-h-[75dvh] overflow-hidden p-4"}
+      contentClassName={
+        isDesktop
+          ? "scrollbar-container h-auto max-h-[80dvh] overflow-y-auto"
+          : "max-h-[75dvh] overflow-hidden p-4"
+      }
       open={open}
       onOpenChange={(open) => {
         if (!open) {
@@ -326,7 +289,7 @@ export function GeneralFilterContent({
 }: GeneralFilterContentProps) {
   return (
     <>
-      <div className="scrollbar-container h-auto max-h-[80dvh] overflow-y-auto overflow-x-hidden">
+      <div className="overflow-x-hidden">
         <div className="mb-5 mt-2.5 flex items-center justify-between">
           <Label
             className="mx-2 cursor-pointer text-primary"
@@ -374,6 +337,7 @@ export function GeneralFilterContent({
       <DropdownMenuSeparator />
       <div className="flex items-center justify-evenly p-2">
         <Button
+          aria-label="Apply"
           variant="select"
           onClick={() => {
             if (selectedLabels != currentLabels) {
@@ -386,6 +350,7 @@ export function GeneralFilterContent({
           Apply
         </Button>
         <Button
+          aria-label="Reset"
           onClick={() => {
             setCurrentLabels(undefined);
             updateLabelFilter(undefined);
@@ -397,681 +362,3 @@ export function GeneralFilterContent({
     </>
   );
 }
-
-type TimeRangeFilterButtonProps = {
-  config?: FrigateConfig;
-  timeRange?: string;
-  updateTimeRange: (range: string | undefined) => void;
-};
-function TimeRangeFilterButton({
-  config,
-  timeRange,
-  updateTimeRange,
-}: TimeRangeFilterButtonProps) {
-  const [open, setOpen] = useState(false);
-  const [startOpen, setStartOpen] = useState(false);
-  const [endOpen, setEndOpen] = useState(false);
-
-  const [afterHour, beforeHour] = useMemo(() => {
-    if (!timeRange || !timeRange.includes(",")) {
-      return [DEFAULT_TIME_RANGE_AFTER, DEFAULT_TIME_RANGE_BEFORE];
-    }
-
-    return timeRange.split(",");
-  }, [timeRange]);
-
-  const [selectedAfterHour, setSelectedAfterHour] = useState(afterHour);
-  const [selectedBeforeHour, setSelectedBeforeHour] = useState(beforeHour);
-
-  // format based on locale
-
-  const formattedAfter = useFormattedHour(config, afterHour);
-  const formattedBefore = useFormattedHour(config, beforeHour);
-  const formattedSelectedAfter = useFormattedHour(config, selectedAfterHour);
-  const formattedSelectedBefore = useFormattedHour(config, selectedBeforeHour);
-
-  useEffect(() => {
-    setSelectedAfterHour(afterHour);
-    setSelectedBeforeHour(beforeHour);
-    // only refresh when state changes
-    // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [timeRange]);
-
-  const trigger = (
-    <Button
-      size="sm"
-      variant={timeRange ? "select" : "default"}
-      className="flex items-center gap-2 capitalize"
-    >
-      <FaClock
-        className={`${timeRange ? "text-selected-foreground" : "text-secondary-foreground"}`}
-      />
-      <div
-        className={`${timeRange ? "text-selected-foreground" : "text-primary"}`}
-      >
-        {timeRange ? `${formattedAfter} - ${formattedBefore}` : "All Times"}
-      </div>
-    </Button>
-  );
-  const content = (
-    <div className="scrollbar-container h-auto max-h-[80dvh] overflow-y-auto overflow-x-hidden">
-      <div className="my-5 flex flex-row items-center justify-center gap-2">
-        <Popover
-          open={startOpen}
-          onOpenChange={(open) => {
-            if (!open) {
-              setStartOpen(false);
-            }
-          }}
-        >
-          <PopoverTrigger asChild>
-            <Button
-              className={`text-primary ${isDesktop ? "" : "text-xs"} `}
-              variant={startOpen ? "select" : "default"}
-              size="sm"
-              onClick={() => {
-                setStartOpen(true);
-                setEndOpen(false);
-              }}
-            >
-              {formattedSelectedAfter}
-            </Button>
-          </PopoverTrigger>
-          <PopoverContent className="flex flex-row items-center justify-center">
-            <input
-              className="text-md mx-4 w-full border border-input bg-background p-1 text-secondary-foreground hover:bg-accent hover:text-accent-foreground dark:[color-scheme:dark]"
-              id="startTime"
-              type="time"
-              value={selectedAfterHour}
-              step="60"
-              onChange={(e) => {
-                const clock = e.target.value;
-                const [hour, minute, _] = clock.split(":");
-                setSelectedAfterHour(`${hour}:${minute}`);
-              }}
-            />
-          </PopoverContent>
-        </Popover>
-        <FaArrowRight className="size-4 text-primary" />
-        <Popover
-          open={endOpen}
-          onOpenChange={(open) => {
-            if (!open) {
-              setEndOpen(false);
-            }
-          }}
-        >
-          <PopoverTrigger asChild>
-            <Button
-              className={`text-primary ${isDesktop ? "" : "text-xs"}`}
-              variant={endOpen ? "select" : "default"}
-              size="sm"
-              onClick={() => {
-                setEndOpen(true);
-                setStartOpen(false);
-              }}
-            >
-              {formattedSelectedBefore}
-            </Button>
-          </PopoverTrigger>
-          <PopoverContent className="flex flex-col items-center">
-            <input
-              className="text-md mx-4 w-full border border-input bg-background p-1 text-secondary-foreground hover:bg-accent hover:text-accent-foreground dark:[color-scheme:dark]"
-              id="startTime"
-              type="time"
-              value={
-                selectedBeforeHour == "24:00" ? "23:59" : selectedBeforeHour
-              }
-              step="60"
-              onChange={(e) => {
-                const clock = e.target.value;
-                const [hour, minute, _] = clock.split(":");
-                setSelectedBeforeHour(`${hour}:${minute}`);
-              }}
-            />
-          </PopoverContent>
-        </Popover>
-      </div>
-      <DropdownMenuSeparator />
-      <div className="flex items-center justify-evenly p-2">
-        <Button
-          variant="select"
-          onClick={() => {
-            if (
-              selectedAfterHour == DEFAULT_TIME_RANGE_AFTER &&
-              selectedBeforeHour == DEFAULT_TIME_RANGE_BEFORE
-            ) {
-              updateTimeRange(undefined);
-            } else {
-              updateTimeRange(`${selectedAfterHour},${selectedBeforeHour}`);
-            }
-
-            setOpen(false);
-          }}
-        >
-          Apply
-        </Button>
-        <Button
-          onClick={() => {
-            setSelectedAfterHour(DEFAULT_TIME_RANGE_AFTER);
-            setSelectedBeforeHour(DEFAULT_TIME_RANGE_BEFORE);
-            updateTimeRange(undefined);
-          }}
-        >
-          Reset
-        </Button>
-      </div>
-    </div>
-  );
-
-  return (
-    <PlatformAwareDialog
-      trigger={trigger}
-      content={content}
-      open={open}
-      onOpenChange={(open) => {
-        setOpen(open);
-      }}
-    />
-  );
-}
-
-type ZoneFilterButtonProps = {
-  allZones: string[];
-  selectedZones?: string[];
-  updateZoneFilter: (zones: string[] | undefined) => void;
-};
-function ZoneFilterButton({
-  allZones,
-  selectedZones,
-  updateZoneFilter,
-}: ZoneFilterButtonProps) {
-  const [open, setOpen] = useState(false);
-
-  const [currentZones, setCurrentZones] = useState<string[] | undefined>(
-    selectedZones,
-  );
-
-  const buttonText = useMemo(() => {
-    if (isMobile) {
-      return "Zones";
-    }
-
-    if (!selectedZones || selectedZones.length == 0) {
-      return "All Zones";
-    }
-
-    if (selectedZones.length == 1) {
-      return selectedZones[0];
-    }
-
-    return `${selectedZones.length} Zones`;
-  }, [selectedZones]);
-
-  // ui
-
-  useEffect(() => {
-    setCurrentZones(selectedZones);
-    // only refresh when state changes
-    // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [selectedZones]);
-
-  const trigger = (
-    <Button
-      size="sm"
-      variant={selectedZones?.length ? "select" : "default"}
-      className="flex items-center gap-2 capitalize"
-    >
-      <FaLocationDot
-        className={`${selectedZones?.length ? "text-selected-foreground" : "text-secondary-foreground"}`}
-      />
-      <div
-        className={`${selectedZones?.length ? "text-selected-foreground" : "text-primary"}`}
-      >
-        {buttonText}
-      </div>
-    </Button>
-  );
-  const content = (
-    <ZoneFilterContent
-      allZones={allZones}
-      selectedZones={selectedZones}
-      currentZones={currentZones}
-      setCurrentZones={setCurrentZones}
-      updateZoneFilter={updateZoneFilter}
-      onClose={() => setOpen(false)}
-    />
-  );
-
-  return (
-    <PlatformAwareDialog
-      trigger={trigger}
-      content={content}
-      open={open}
-      onOpenChange={(open) => {
-        if (!open) {
-          setCurrentZones(selectedZones);
-        }
-
-        setOpen(open);
-      }}
-    />
-  );
-}
-
-type ZoneFilterContentProps = {
-  allZones?: string[];
-  selectedZones?: string[];
-  currentZones?: string[];
-  updateZoneFilter?: (zones: string[] | undefined) => void;
-  setCurrentZones?: (zones: string[] | undefined) => void;
-  onClose: () => void;
-};
-export function ZoneFilterContent({
-  allZones,
-  selectedZones,
-  currentZones,
-  updateZoneFilter,
-  setCurrentZones,
-  onClose,
-}: ZoneFilterContentProps) {
-  return (
-    <>
-      <div className="scrollbar-container h-auto max-h-[80dvh] overflow-y-auto overflow-x-hidden">
-        {allZones && setCurrentZones && (
-          <>
-            {isDesktop && <DropdownMenuSeparator />}
-            <div className="mb-5 mt-2.5 flex items-center justify-between">
-              <Label
-                className="mx-2 cursor-pointer text-primary"
-                htmlFor="allZones"
-              >
-                All Zones
-              </Label>
-              <Switch
-                className="ml-1"
-                id="allZones"
-                checked={currentZones == undefined}
-                onCheckedChange={(isChecked) => {
-                  if (isChecked) {
-                    setCurrentZones(undefined);
-                  }
-                }}
-              />
-            </div>
-            <div className="my-2.5 flex flex-col gap-2.5">
-              {allZones.map((item) => (
-                <FilterSwitch
-                  key={item}
-                  label={item.replaceAll("_", " ")}
-                  isChecked={currentZones?.includes(item) ?? false}
-                  onCheckedChange={(isChecked) => {
-                    if (isChecked) {
-                      const updatedZones = currentZones
-                        ? [...currentZones]
-                        : [];
-
-                      updatedZones.push(item);
-                      setCurrentZones(updatedZones);
-                    } else {
-                      const updatedZones = currentZones
-                        ? [...currentZones]
-                        : [];
-
-                      // can not deselect the last item
-                      if (updatedZones.length > 1) {
-                        updatedZones.splice(updatedZones.indexOf(item), 1);
-                        setCurrentZones(updatedZones);
-                      }
-                    }
-                  }}
-                />
-              ))}
-            </div>
-          </>
-        )}
-      </div>
-      {isDesktop && <DropdownMenuSeparator />}
-      <div className="flex items-center justify-evenly p-2">
-        <Button
-          variant="select"
-          onClick={() => {
-            if (updateZoneFilter && selectedZones != currentZones) {
-              updateZoneFilter(currentZones);
-            }
-
-            onClose();
-          }}
-        >
-          Apply
-        </Button>
-        <Button
-          onClick={() => {
-            setCurrentZones?.(undefined);
-            updateZoneFilter?.(undefined);
-          }}
-        >
-          Reset
-        </Button>
-      </div>
-    </>
-  );
-}
-
-type SubFilterButtonProps = {
-  allSubLabels: string[];
-  selectedSubLabels: string[] | undefined;
-  updateSubLabelFilter: (labels: string[] | undefined) => void;
-};
-function SubFilterButton({
-  allSubLabels,
-  selectedSubLabels,
-  updateSubLabelFilter,
-}: SubFilterButtonProps) {
-  const [open, setOpen] = useState(false);
-  const [currentSubLabels, setCurrentSubLabels] = useState<
-    string[] | undefined
-  >(selectedSubLabels);
-
-  const buttonText = useMemo(() => {
-    if (isMobile) {
-      return "Sub Labels";
-    }
-
-    if (!selectedSubLabels || selectedSubLabels.length == 0) {
-      return "All Sub Labels";
-    }
-
-    if (selectedSubLabels.length == 1) {
-      return selectedSubLabels[0];
-    }
-
-    return `${selectedSubLabels.length} Sub Labels`;
-  }, [selectedSubLabels]);
-
-  const trigger = (
-    <Button
-      size="sm"
-      variant={selectedSubLabels?.length ? "select" : "default"}
-      className="flex items-center gap-2 capitalize"
-    >
-      <SubFilterIcon
-        className={`${selectedSubLabels?.length || selectedSubLabels?.length ? "text-selected-foreground" : "text-secondary-foreground"}`}
-      />
-      <div
-        className={`${selectedSubLabels?.length ? "text-selected-foreground" : "text-primary"}`}
-      >
-        {buttonText}
-      </div>
-    </Button>
-  );
-  const content = (
-    <SubFilterContent
-      allSubLabels={allSubLabels}
-      selectedSubLabels={selectedSubLabels}
-      currentSubLabels={currentSubLabels}
-      setCurrentSubLabels={setCurrentSubLabels}
-      updateSubLabelFilter={updateSubLabelFilter}
-      onClose={() => setOpen(false)}
-    />
-  );
-
-  return (
-    <PlatformAwareDialog
-      trigger={trigger}
-      content={content}
-      open={open}
-      onOpenChange={(open) => {
-        if (!open) {
-          setCurrentSubLabels(selectedSubLabels);
-        }
-
-        setOpen(open);
-      }}
-    />
-  );
-}
-
-type SubFilterContentProps = {
-  allSubLabels: string[];
-  selectedSubLabels: string[] | undefined;
-  currentSubLabels: string[] | undefined;
-  updateSubLabelFilter: (labels: string[] | undefined) => void;
-  setCurrentSubLabels: (labels: string[] | undefined) => void;
-  onClose: () => void;
-};
-export function SubFilterContent({
-  allSubLabels,
-  selectedSubLabels,
-  currentSubLabels,
-  updateSubLabelFilter,
-  setCurrentSubLabels,
-  onClose,
-}: SubFilterContentProps) {
-  return (
-    <>
-      <div className="scrollbar-container h-auto max-h-[80dvh] overflow-y-auto overflow-x-hidden">
-        <div className="mb-5 mt-2.5 flex items-center justify-between">
-          <Label
-            className="mx-2 cursor-pointer text-primary"
-            htmlFor="allLabels"
-          >
-            All Sub Labels
-          </Label>
-          <Switch
-            className="ml-1"
-            id="allLabels"
-            checked={currentSubLabels == undefined}
-            onCheckedChange={(isChecked) => {
-              if (isChecked) {
-                setCurrentSubLabels(undefined);
-              }
-            }}
-          />
-        </div>
-        <div className="my-2.5 flex flex-col gap-2.5">
-          {allSubLabels.map((item) => (
-            <FilterSwitch
-              key={item}
-              label={item.replaceAll("_", " ")}
-              isChecked={currentSubLabels?.includes(item) ?? false}
-              onCheckedChange={(isChecked) => {
-                if (isChecked) {
-                  const updatedLabels = currentSubLabels
-                    ? [...currentSubLabels]
-                    : [];
-
-                  updatedLabels.push(item);
-                  setCurrentSubLabels(updatedLabels);
-                } else {
-                  const updatedLabels = currentSubLabels
-                    ? [...currentSubLabels]
-                    : [];
-
-                  // can not deselect the last item
-                  if (updatedLabels.length > 1) {
-                    updatedLabels.splice(updatedLabels.indexOf(item), 1);
-                    setCurrentSubLabels(updatedLabels);
-                  }
-                }
-              }}
-            />
-          ))}
-        </div>
-      </div>
-      {isDesktop && <DropdownMenuSeparator />}
-      <div className="flex items-center justify-evenly p-2">
-        <Button
-          variant="select"
-          onClick={() => {
-            if (selectedSubLabels != currentSubLabels) {
-              updateSubLabelFilter(currentSubLabels);
-            }
-
-            onClose();
-          }}
-        >
-          Apply
-        </Button>
-        <Button
-          onClick={() => {
-            updateSubLabelFilter(undefined);
-          }}
-        >
-          Reset
-        </Button>
-      </div>
-    </>
-  );
-}
-
-type SearchTypeButtonProps = {
-  selectedSearchSources: SearchSource[] | undefined;
-  updateSearchSourceFilter: (sources: SearchSource[] | undefined) => void;
-};
-function SearchTypeButton({
-  selectedSearchSources,
-  updateSearchSourceFilter,
-}: SearchTypeButtonProps) {
-  const [open, setOpen] = useState(false);
-
-  const buttonText = useMemo(() => {
-    if (isMobile) {
-      return "Sources";
-    }
-
-    if (
-      !selectedSearchSources ||
-      selectedSearchSources.length == 0 ||
-      selectedSearchSources.length == 2
-    ) {
-      return "All Search Sources";
-    }
-
-    if (selectedSearchSources.length == 1) {
-      return selectedSearchSources[0];
-    }
-
-    return `${selectedSearchSources.length} Search Sources`;
-  }, [selectedSearchSources]);
-
-  const trigger = (
-    <Button
-      size="sm"
-      variant={selectedSearchSources?.length != 2 ? "select" : "default"}
-      className="flex items-center gap-2 capitalize"
-    >
-      <SearchSourceIcon
-        className={`${selectedSearchSources?.length != 2 ? "text-selected-foreground" : "text-secondary-foreground"}`}
-      />
-      <div
-        className={`${selectedSearchSources?.length != 2 ? "text-selected-foreground" : "text-primary"}`}
-      >
-        {buttonText}
-      </div>
-    </Button>
-  );
-  const content = (
-    <SearchTypeContent
-      selectedSearchSources={selectedSearchSources}
-      updateSearchSourceFilter={updateSearchSourceFilter}
-      onClose={() => setOpen(false)}
-    />
-  );
-
-  return (
-    <PlatformAwareDialog
-      trigger={trigger}
-      content={content}
-      open={open}
-      onOpenChange={setOpen}
-    />
-  );
-}
-
-type SearchTypeContentProps = {
-  selectedSearchSources: SearchSource[] | undefined;
-  updateSearchSourceFilter: (sources: SearchSource[] | undefined) => void;
-  onClose: () => void;
-};
-export function SearchTypeContent({
-  selectedSearchSources,
-  updateSearchSourceFilter,
-  onClose,
-}: SearchTypeContentProps) {
-  const [currentSearchSources, setCurrentSearchSources] = useState<
-    SearchSource[] | undefined
-  >(selectedSearchSources);
-
-  return (
-    <>
-      <div className="scrollbar-container h-auto max-h-[80dvh] overflow-y-auto overflow-x-hidden">
-        <div className="my-2.5 flex flex-col gap-2.5">
-          <FilterSwitch
-            label="Thumbnail Image"
-            isChecked={currentSearchSources?.includes("thumbnail") ?? false}
-            onCheckedChange={(isChecked) => {
-              const updatedSources = currentSearchSources
-                ? [...currentSearchSources]
-                : [];
-
-              if (isChecked) {
-                updatedSources.push("thumbnail");
-                setCurrentSearchSources(updatedSources);
-              } else {
-                if (updatedSources.length > 1) {
-                  const index = updatedSources.indexOf("thumbnail");
-                  if (index !== -1) updatedSources.splice(index, 1);
-                  setCurrentSearchSources(updatedSources);
-                }
-              }
-            }}
-          />
-          <FilterSwitch
-            label="Description"
-            isChecked={currentSearchSources?.includes("description") ?? false}
-            onCheckedChange={(isChecked) => {
-              const updatedSources = currentSearchSources
-                ? [...currentSearchSources]
-                : [];
-
-              if (isChecked) {
-                updatedSources.push("description");
-                setCurrentSearchSources(updatedSources);
-              } else {
-                if (updatedSources.length > 1) {
-                  const index = updatedSources.indexOf("description");
-                  if (index !== -1) updatedSources.splice(index, 1);
-                  setCurrentSearchSources(updatedSources);
-                }
-              }
-            }}
-          />
-        </div>
-        {isDesktop && <DropdownMenuSeparator />}
-        <div className="flex items-center justify-evenly p-2">
-          <Button
-            variant="select"
-            onClick={() => {
-              if (selectedSearchSources != currentSearchSources) {
-                updateSearchSourceFilter(currentSearchSources);
-              }
-
-              onClose();
-            }}
-          >
-            Apply
-          </Button>
-          <Button
-            onClick={() => {
-              updateSearchSourceFilter(undefined);
-              setCurrentSearchSources(["thumbnail", "description"]);
-            }}
-          >
-            Reset
-          </Button>
-        </div>
-      </div>
-    </>
-  );
-}

web/src/components/filter/ZoneMaskFilter.tsx

@@ -21,6 +21,7 @@ export function ZoneMaskFilterButton({
       size="sm"
       variant={selectedZoneMask?.length ? "select" : "default"}
       className="flex items-center gap-2 capitalize"
+      aria-label="Filter by zone mask"
     >
       <FaFilter
         className={`${selectedZoneMask?.length ? "text-selected-foreground" : "text-secondary-foreground"}`}

web/src/components/graph/CombinedStorageGraph.tsx

@@ -216,7 +216,7 @@ export function CombinedStorageGraph({
                     </Popover>
                   )}
                 </TableCell>
-                <TableCell>{getUnitSize(item.usage)}</TableCell>
+                <TableCell>{getUnitSize(item.usage ?? 0)}</TableCell>
                 <TableCell>{item.data[0].toFixed(2)}%</TableCell>
                 <TableCell>
                   {item.name === "Unused"

web/src/components/icons/IconPicker.tsx

@@ -66,7 +66,10 @@ export default function IconPicker({
       >
         <PopoverTrigger asChild>
           {!selectedIcon?.name || !selectedIcon?.Icon ? (
-            <Button className="mt-2 w-full text-muted-foreground">
+            <Button
+              className="mt-2 w-full text-muted-foreground"
+              aria-label="Select an icon"
+            >
               Select an icon
             </Button>
           ) : (

web/src/components/input/DeleteSearchDialog.tsx

@@ -8,6 +8,7 @@ import {
   AlertDialogHeader,
   AlertDialogTitle,
 } from "@/components/ui/alert-dialog";
+import { buttonVariants } from "../ui/button";
 
 type DeleteSearchDialogProps = {
   isOpen: boolean;
@@ -35,7 +36,7 @@ export function DeleteSearchDialog({
           <AlertDialogCancel onClick={onClose}>Cancel</AlertDialogCancel>
           <AlertDialogAction
             onClick={onConfirm}
-            className="bg-destructive text-white"
+            className={buttonVariants({ variant: "destructive" })}
           >
             Delete
           </AlertDialogAction>

web/src/components/input/InputWithTags.tsx

@@ -2,11 +2,11 @@ import React, { useState, useRef, useEffect, useCallback } from "react";
 import {
   LuX,
   LuFilter,
-  LuImage,
   LuChevronDown,
   LuChevronUp,
   LuTrash2,
   LuStar,
+  LuSearch,
 } from "react-icons/lu";
 import {
   FilterType,
@@ -43,6 +43,7 @@ import {
 import { toast } from "sonner";
 import useSWR from "swr";
 import { FrigateConfig } from "@/types/frigateConfig";
+import { MdImageSearch } from "react-icons/md";
 
 type InputWithTagsProps = {
   inputFocused: boolean;
@@ -161,8 +162,12 @@ export default function InputWithTags({
         .map((word) => word.trim())
         .lastIndexOf(words.filter((word) => word.trim() !== "").pop() || "");
       const currentWord = words[lastNonEmptyWordIndex];
+      if (words.at(-1) === "") {
+        return current_suggestions;
+      }
+
       return current_suggestions.filter((suggestion) =>
-        suggestion.toLowerCase().includes(currentWord.toLowerCase()),
+        suggestion.toLowerCase().startsWith(currentWord),
       );
     },
     [inputValue, suggestions, currentFilterType],
@@ -196,10 +201,13 @@ export default function InputWithTags({
         allSuggestions[type as FilterType]?.includes(value) ||
         type == "before" ||
         type == "after" ||
-        type == "time_range"
+        type == "time_range" ||
+        type == "min_score" ||
+        type == "max_score"
       ) {
         const newFilters = { ...filters };
         let timestamp = 0;
+        let score = 0;
 
         switch (type) {
           case "before":
@@ -239,6 +247,40 @@ export default function InputWithTags({
               newFilters[type] = timestamp / 1000;
             }
             break;
+          case "min_score":
+          case "max_score":
+            score = parseInt(value);
+            if (score >= 0) {
+              // Check for conflicts between min_score and max_score
+              if (
+                type === "min_score" &&
+                filters.max_score !== undefined &&
+                score > filters.max_score * 100
+              ) {
+                toast.error(
+                  "The 'min_score' must be less than or equal to the 'max_score'.",
+                  {
+                    position: "top-center",
+                  },
+                );
+                return;
+              }
+              if (
+                type === "max_score" &&
+                filters.min_score !== undefined &&
+                score < filters.min_score * 100
+              ) {
+                toast.error(
+                  "The 'max_score' must be greater than or equal to the 'min_score'.",
+                  {
+                    position: "top-center",
+                  },
+                );
+                return;
+              }
+              newFilters[type] = score / 100;
+            }
+            break;
           case "time_range":
             newFilters[type] = value;
             break;
@@ -254,6 +296,14 @@ export default function InputWithTags({
               );
             }
             break;
+          case "has_snapshot":
+            if (!newFilters.has_snapshot) newFilters.has_snapshot = undefined;
+            newFilters.has_snapshot = value == "yes" ? 1 : 0;
+            break;
+          case "has_clip":
+            if (!newFilters.has_clip) newFilters.has_clip = undefined;
+            newFilters.has_clip = value == "yes" ? 1 : 0;
+            break;
           case "event_id":
             newFilters.event_id = value;
             break;
@@ -297,6 +347,10 @@ export default function InputWithTags({
       } - ${
         config?.ui.time_format === "24hour" ? endTime : convertTo12Hour(endTime)
       }`;
+    } else if (filterType === "min_score" || filterType === "max_score") {
+      return Math.round(Number(filterValues) * 100).toString() + "%";
+    } else if (filterType === "has_clip" || filterType === "has_snapshot") {
+      return filterValues ? "Yes" : "No";
     } else {
       return filterValues as string;
     }
@@ -315,7 +369,11 @@ export default function InputWithTags({
           isValidTimeRange(
             trimmedValue.replace("-", ","),
             config?.ui.time_format,
-          ))
+          )) ||
+        ((filterType === "min_score" || filterType === "max_score") &&
+          !isNaN(Number(trimmedValue)) &&
+          Number(trimmedValue) >= 50 &&
+          Number(trimmedValue) <= 100)
       ) {
         createFilter(
           filterType,
@@ -397,6 +455,11 @@ export default function InputWithTags({
     setIsSimilaritySearch(false);
   }, [setFilters, resetSuggestions, setSearch, setInputFocused]);
 
+  const handleClearSimilarity = useCallback(() => {
+    removeFilter("event_id", filters.event_id!);
+    removeFilter("search_type", "similarity");
+  }, [removeFilter, filters]);
+
   const handleInputBlur = useCallback(
     (e: React.FocusEvent) => {
       if (
@@ -504,7 +567,7 @@ export default function InputWithTags({
             onFocus={handleInputFocus}
             onBlur={handleInputBlur}
             onKeyDown={handleInputKeyDown}
-            className="text-md h-9 pr-24"
+            className="text-md h-9 pr-32"
             placeholder="Search..."
           />
           <div className="absolute right-3 top-0 flex h-full flex-row items-center justify-center gap-5">
@@ -539,7 +602,7 @@ export default function InputWithTags({
             {isSimilaritySearch && (
               <Tooltip>
                 <TooltipTrigger className="cursor-default">
-                  <LuImage
+                  <MdImageSearch
                     aria-label="Similarity search active"
                     className="size-4 text-selected"
                   />
@@ -631,14 +694,26 @@ export default function InputWithTags({
             inputFocused ? "visible" : "hidden",
           )}
         >
-          {(Object.keys(filters).length > 0 || isSimilaritySearch) && (
+          {!currentFilterType && inputValue && (
+            <CommandGroup heading="Search">
+              <CommandItem
+                className="cursor-pointer"
+                onSelect={() => handleSearch(inputValue)}
+              >
+                <LuSearch className="mr-2 h-4 w-4" />
+                Search for "{inputValue}"
+              </CommandItem>
+            </CommandGroup>
+          )}
+          {(Object.keys(filters).filter((key) => key !== "query").length > 0 ||
+            isSimilaritySearch) && (
             <CommandGroup heading="Active Filters">
               <div className="my-2 flex flex-wrap gap-2 px-2">
                 {isSimilaritySearch && (
                   <span className="inline-flex items-center whitespace-nowrap rounded-full bg-blue-100 px-2 py-0.5 text-sm text-blue-800">
                     Similarity Search
                     <button
-                      onClick={handleClearInput}
+                      onClick={handleClearSimilarity}
                       className="ml-1 focus:outline-none"
                       aria-label="Clear similarity search"
                     >

web/src/components/input/SaveSearchDialog.tsx

@@ -59,11 +59,14 @@ export function SaveSearchDialog({
           placeholder="Enter a name for your search"
         />
         <DialogFooter>
-          <Button onClick={onClose}>Cancel</Button>
+          <Button aria-label="Cancel" onClick={onClose}>
+            Cancel
+          </Button>
           <Button
             onClick={handleSave}
             variant="select"
             className="mb-2 md:mb-0"
+            aria-label="Save this search"
           >
             Save
           </Button>