Get Started#
The Video Search and Summarization (VSS) sample application helps developers create a summary of long form video, search for the right video, and combine both search and summarization pipelines. This guide will help you set up, run, and modify the sample application on local and Edge AI systems.
This guide shows how to:
Set up the sample application: Use the setup script to quickly deploy the application in your environment.
Run different application modes: Deploy in summary-only, search-only, dual-UI, or unified mode.
Modify application parameters: Customize settings like inference models and deployment configurations to adapt the application to your specific requirements.
Deployment Modes#
The application supports four deployment modes. Each mode deploys only the services and UI(s) relevant to the selected functionality:
Mode |
Features |
UI Layout |
Description |
Command Option |
|---|---|---|---|---|
Summary |
Video summarization only |
Summary UI available at |
Summarize a given video with several tunable parameters. |
|
Search |
Video Search only |
Search UI available at |
Search for entities in a given video. Embedding used for search: Video frame embeddings |
|
Dual UI |
Video Summarization and Video Search |
Two separate UIs available at |
Embedding used for search: Video frame embeddings |
|
Unified UI |
Video Summarization and Modified Video Search |
A single unified UI available at |
Embedding used for search: Summarized content text embeddings |
|
Note
The video search in Unified UI mode is modified for creating the embeddings of video summary texts and searching over them, rather than creating and using video frame embeddings. Hence, this mode includes video summarization feature, as well, in the same UI.
Prerequisites#
Verify that your system meets the minimum requirements.
Install Docker tool: Installation Guide.
Install Docker Compose tool: Installation Guide.
Install Python programming language v3.11
Project Structure#
The repository is organized as follows:
sample-applications/video-search-and-summarization/
├── config/ # Runtime configs
│ ├── nginx/ # Nginx templates used by setup.sh + compose
│ │ ├── nginx.conf
│ │ ├── dual_ui.conf
│ │ └── singleton_ui.conf
│ └── rmq.conf # RabbitMQ configuration
├── docker/ # Docker Compose base and overlays
│ ├── compose.base.yaml
│ ├── compose.ui.yaml
│ ├── compose.summary.yaml
│ ├── compose.search.yaml
│ ├── compose.search.vdms.yaml # VDMS vector-DB backend overlay (default)
│ ├── compose.search.milvus.yaml # Milvus vector-DB backend overlay
│ ├── compose.ovms.yaml # OVMS backend: makes ovms-service a required dependency
│ ├── compose.vllm.yaml # vLLM CPU backend
│ ├── compose.vllm.xpu.yaml # vLLM Intel Arc Pro B-series GPU/XPU backend (experimental)
│ ├── compose.gpu_ovms.yaml
│ └── compose.metrics-manager.yaml
├── docs/
│ └── user-guide/ # User guides and tutorials
├── pipeline-manager/ # Orchestrates summarization and search pipelines
├── search-ms/ # Video search microservice
├── video-ingestion/ # Video ingestion and processing service
├── ui/
│ └── react/ # Frontend application
├── cli/ # Terminal UI and CLI workflows
├── scripts/ # Utility and helper scripts
├── data/ # Default watcher/input data directory
├── ov_models/ # Local model cache/artifacts
├── build.sh # Script for building application images
├── setup.sh # Main setup and deployment script
└── README.md
Set Required Environment Variables#
Before running the application, you need to set several environment variables:
Configure the registry: The application uses registry URL and tag to pull the required images.
export REGISTRY_URL=intel export TAG=2026.3.0-rc1
The model-download helper defaults to
intel/model-download:2026.3.0-rc1, independently ofREGISTRY_URLandTAG. SetMODEL_DOWNLOAD_IMAGEto override the image, orMODEL_DOWNLOAD_TAGto override its tag whenMODEL_DOWNLOAD_IMAGEis unset.Set required credentials for some services: Following variables must be set on your current shell before running the setup script:
# MinIO credentials (object storage) export MINIO_ROOT_USER=<your-minio-username> export MINIO_ROOT_PASSWORD=<your-minio-password> # PostgreSQL credentials (database) export POSTGRES_USER=<your-postgres-username> export POSTGRES_PASSWORD=<your-postgres-password> # RabbitMQ credentials (message broker) export RABBITMQ_USER=<your-rabbitmq-username> export RABBITMQ_PASSWORD=<your-rabbitmq-password>
Set environment variables for model selection:
You must set these environment variables on your current shell. Setting these variables is mandatory as they do not have any default values.
Mode-specific environment variables to set Models:
Variable
Mode
Purpose
VLM_MODEL_NAMESummary, Dual UI, Unified UI
VLM model for video captioning and summarization.
OD_MODEL_NAMESummary, Dual UI, Unified UI
YOLO model for object detection during video ingestion.
ASR_MODELSummary, Dual UI, Unified UI
Audio Analyzer transcription model. Only Whisper variants are supported:
whisper-{tiny,base,small,medium,large}.MULTIMODAL_EMBEDDING_MODELSearch, Dual UI
Multimodal model for generating video frame embeddings.
TEXT_EMBEDDING_MODELUnified UI
Text embedding model for generating summary text embeddings.
DATAPREP_DETECTION_MODELSearch, Dual UI
Hugging Face OpenVINO detection model used by
multimodal-dataprepfor object crops, for exampleOpenVINO/yolox_s-int8-ov(any object-detection model under huggingface.co/OpenVINO supported by OpenVINO Model API, such as YOLOX or YOLO11 exports).OpenVINO/yolox_tiny-fp16-ov(lightest),OpenVINO/yolox_l-int8-ov(higher accuracy) andOpenVINO/YOLO11n-fp16-ovare also verified on CPU, GPU and NPU. Only the selected model is downloaded on first use and cached in thedocker_dataprep-detection-modelsvolume. Each model keeps its own license (for example YOLOX is Apache-2.0, YOLO11 is AGPL-3.0); review the model card. No default; not needed whenENABLE_OBJECT_DETECTION=false. Unified UI mode embeds summary text only and always runs without object detection.DATAPREP_DETECTION_CLASSES(Optional) Search, Dual UI
Comma-separated class filter for DataPrep object crops, for example
person,car. Case-insensitive; spaces and hyphens become_(model labels look liketraffic_light). The tested OpenVINO detection models use the 80 COCO classes. Classes the model does not have prevent the detector from loading. Empty (default) keeps all classes.OVMS_LLM_MODEL_NAME(Optional) Summary, Dual UI, or Unified UI
LLM for OVMS-based final summary generation. When set and different from
VLM_MODEL_NAME, enables split-model mode (separate VLM for captioning, LLM for summarization).PM_AUDIO_USE_FULL_TRANSCRIPT_SUMMARY(Optional) Summary, Dual UI
Enables condensed transcript summary injection in the prompt to generate video summary.
Common to all modes except
--search:# For VLM-based chunk captioning and video summarization with OVMS (CPU or GPU). # See "Recommended OVMS Models" for alternatives. vLLM needs a Hugging Face # checkpoint instead; see the vLLM options below. export VLM_MODEL_NAME="OpenVINO/Qwen3.5-4B-int8-ov" export VLM_TARGET_DEVICE="GPU" # Options: CPU, GPU, HETERO:GPU,CPU (NPU is not supported for the VLM) # (OPTIONAL) For OVMS split-model summarization, set a dedicated LLM model for final summary. # If this is not set, OVMS falls back to VLM_MODEL_NAME as the LLM model source. # OVMS uses shared mode only when model source, target device, and compression format all match. export OVMS_LLM_MODEL_NAME="OpenVINO/Qwen3-4B-int8-ov" # on NPU use OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov export LLM_TARGET_DEVICE="CPU" # Options: CPU, GPU, NPU, HETERO:GPU,CPU # When ENABLE_VLLM=true, vLLM is the only inference backend and setup.sh ignores OVMS_LLM_MODEL_NAME. # Object detection model used for Video Ingestion Service. Only generic # YOLO models supported by the model-download ultralytics hub are # supported (e.g. yolov8l, yolov8s). export OD_MODEL_NAME="yolov8l" # Audio Analyzer transcription model. Only Whisper variants are supported. export ASR_MODEL="whisper-base" # whisper-{tiny,base,small,medium,large}
Required in
--searchand--summary --searchmode:# Required for searching on video frame embeddings export MULTIMODAL_EMBEDDING_MODEL="CLIP/clip-vit-b-32" # Required Hugging Face OpenVINO object detection model for DataPrep (or export ENABLE_OBJECT_DETECTION=false) export DATAPREP_DETECTION_MODEL="OpenVINO/yolox_s-int8-ov" # Also verified on CPU/GPU/NPU: OpenVINO/yolox_tiny-fp16-ov (lightest), OpenVINO/yolox_l-int8-ov (more accurate), # OpenVINO/YOLO11n-fp16-ov (AGPL-3.0) # Optional: keep only some detected classes (COCO class names: https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/coco.yaml) # export DATAPREP_DETECTION_CLASSES="person,car"
Required in
--summary-and-searchmode:# Required for searching on video summary text embeddings export TEXT_EMBEDDING_MODEL="QwenText/qwen3-embedding-0.6b"
Note
Review the supported model list in supported-models before choosing model names.
Configure summarization to use audio transcript (Summary and Dual UI mode):
Used in
--summaryand--summary --searchmode:# (OPTIONAL) Default value is true. Users can override this per-video in the upload modal. export PM_AUDIO_USE_FULL_TRANSCRIPT_SUMMARY=false
Audio Transcript Summarization (
PM_AUDIO_USE_FULL_TRANSCRIPT_SUMMARY): When enabled (the default), the pipeline runs a separate LLM-based map-reduce summarization pass over the complete audio transcript before generating the final video summary. The condensed transcript summary is then injected into the video summary prompt via the%audio_summary%placeholder, giving the LLM a coherent, high-quality representation of spoken content rather than raw subtitle fragments. This significantly improves accuracy for dialogue-heavy or narration-heavy videos. When disabled, audio transcripts are only used at the chunk captioning level, each chunk’s VLM prompt includes its time-matched portion of the transcript but no audio content is included in the final map-reduce video summary.This environment variable sets the default value. Users can override it per-video using the “Use Audio in Summary” checkbox in the Audio Settings section of the video upload modal.
Configure Directory Watcher (Search and Dual UI mode):
For automated video ingestion into the search pipeline (available only in
--searchand--summary --searchmodes), you can use the directory watcher service:# Path to the directory to watch on the host system. Default: "edge-ai-libraries/sample-applications/video-search-and-summarization/data" export VS_WATCHER_DIR="/path/to/your/video/directory"
Directory Watcher: For complete setup instructions, configuration options, and usage details, see the Directory Watcher Service Guide. This service only works with
--searchand--summary --searchmodes.Configure live RTSP stream ingestion (Search and Dual UI mode):
Live camera ingestion is enabled by default in
--searchand--summary --searchmodes. Registered RTSP cameras are decoded continuously, embedded, and indexed for search. Manage them from the Live Streams button in the UI toolbar.It is not available in
--summarymode (there is no search index) or in--summary-and-searchmode, where the index holds text embeddings of summaries rather than video frame embeddings.# Set to FEATURE_OFF to hide the Live Streams button. export UI_LIVE_STREAMS_FEATURE=FEATURE_ON # Hours of live footage and embeddings to retain. 0 keeps everything, # which grows the index without bound. export LIVE_RETENTION_HOURS=24 # Maximum cameras ingesting at once. Registering more returns 503; # pause or delete a camera to free a slot. export LIVE_STREAM_MAX_CONCURRENT=8
Credentials: An RTSP URL may embed
user:password. DataPrep stores the credentialed URL only in its own registry, and returns a redacted URL (rtsp://camera-host:554/stream) in every API response, log line, and search result. The UI never persists the credentialed URL.In the Add Stream dialog, the RTSP URL is the only required field; every other field is optional and falls back to a service default. Each field carries an info icon explaining what it does and how it affects search results. Notable defaults:
Field
Default
Effect on search
Name
The redacted URL
Label shown in the camera list and with results
Tags
None
Lets you narrow a search to one camera or group
Frame interval
15
Lower indexes more moments; higher costs less
Enable object detection
On
Indexes detected objects separately, improving accuracy for small or crowded subjects
Start ingesting immediately
On
Clear it to register the camera paused
The object detection confidence threshold applies to all cameras and uploads. Set it with the
DETECTION_CONFIDENCEenvironment variable (default0.85) before runningsource setup.sh.Control the frame extraction interval (Search and Dual UI mode):
The DataPrep microservice samples frames from uploaded videos according to the
FRAME_INTERVALenvironment variable. Set this variable before runningsource setup.shto control how often frames are selected for processing.export FRAME_INTERVAL=15
In the example above, DataPrep processes every fifteenth frame: each selected frame (optionally after object detection) is converted into embeddings and stored in the vector database. Lower values improve recall at the cost of higher compute and storage usage, while higher values reduce processing load but may skip important frames. If you do not set this variable, the service falls back to its configured default.
Control duplicate-upload behavior in DataPrep (Search and Dual UI mode):
This setting controls whether DataPrep accepts repeated uploads of the same video content:
# Default true: allow content-identical re-uploads (historical behavior) # Set false to reject duplicate-content uploads with HTTP 409. export MM_DATAPREP_ALLOW_DUPLICATE_UPLOADS=true
Duplicate detection is content-based (SHA-256 of uploaded bytes), not filename-based. A renamed copy of the same video is still detected as duplicate content, while two different videos that share the same filename are treated as different uploads.
Enable ROI consolidation (Search and Dual UI mode):
ROI consolidation groups overlapping object detections into merged regions of interest (ROIs) before cropping for embeddings. Enable this feature and tune it with the following environment variables:
# Enable ROI consolidation (default: false) export ROI_CONSOLIDATION_ENABLED=true # IoU threshold for grouping ROIs (higher = stricter merging) export ROI_CONSOLIDATION_IOU_THRESHOLD=0.2 # Only merge ROIs with the same class label when true export ROI_CONSOLIDATION_CLASS_AWARE=false # Expand merged ROIs by a fraction of width/height export ROI_CONSOLIDATION_CONTEXT_SCALE=0.2
The IoU calculation follows the standard formula:
$$ IoU(A, B) = \frac{|A \cap B|}{|A \cup B|} $$
Note
Enabling ROI consolidation can improve search relevance by creating more meaningful regions for embedding, but it may also increase processing time.
(Optional) Metrics Manager (Search, Dual UI, and Unified UI modes):
Metrics Manager collects CPU, RAM, Intel® GPU, and Intel® NPU metrics and serves them to the UI over server-sent events. Multimodal DataPrep publishes each completed embedding pipeline’s embeddings-per-second value directly to Metrics Manager; Pipeline Manager is not part of this path.
# Disabled by default
export ENABLE_METRICS_MANAGER=false
# Enable live metrics
export ENABLE_METRICS_MANAGER=true
This integration requires the coordinated multimodal-dataprep image that
supports MM_DATAPREP_METRICS_MANAGER_URL. Metrics publishing is
non-blocking: video ingestion continues if Metrics Manager is unavailable.
GPU and NPU panels remain empty on hosts without those devices.
Tune Inference Concurrency (Summary and Dual UI mode):
Control how many concurrent inference requests the pipeline manager sends to OVMS or vLLM. These values affect throughput and resource utilization:
# Maximum concurrent VLM requests for chunk captioning (default: 4; 1 for vLLM) export PM_VLM_CONCURRENT=4 # Maximum concurrent LLM requests for final summarization (default: 2; 1 for vLLM) export PM_LLM_CONCURRENT=2
Note
For OVMS deployments, the defaults are the same on CPU, GPU, and NPU. The setup script sets OVMS
max_num_seqsto at leastPM_VLM_CONCURRENT + PM_LLM_CONCURRENT(minimum 16; override withOVMS_MAX_NUM_SEQS). vLLM deployments default both values to1. Limits are global across all videos, and when the VLM and LLM are the same model, LLM requests share the VLM limit. If a GPU runs out of memory, lower these values.Override OVMS Model Weight Compression Format (Summary and Dual UI mode):
When OVMS converts a model from a non-
OpenVINO/repo, the setup script auto-selects the weight compression format based on the target device (int8for CPU,int4for GPU/NPU). Pre-convertedOpenVINO/repos, such as the recommended models, keep the format in their name (for example-int8-ov), and these variables do not apply. You can override this auto-detection by setting these variables before running the setup script:# Override VLM model weight compression format (default: int8 for CPU, int4 for GPU) export VLM_COMPRESSION_WEIGHT_FORMAT=int4 # Override LLM model weight compression format (default: int8 for CPU, int4 for GPU/NPU) export LLM_COMPRESSION_WEIGHT_FORMAT=int4
Note
Lower precision formats like
int4reduce memory usage and can improve throughput, but may affect output quality. The default auto-detection (int8for CPU,int4for GPU/NPU) is recommended for most use cases.Configure Embedding Execution (Search and Dual UI mode):
The
multimodal-dataprepservice loads the embedding model in-process and generates embeddings during video search indexing. Enable OpenVINO™ optimization for the in-process embedding path:# Enable OpenVINO optimization for in-process embedding (default: true) # Automatically set to true when using GPU/NPU mode export SDK_USE_OPENVINO=true
Note
multimodal-dataprepembeds in-process (no separate embedding service is required for indexing). The standalonemultimodal-embedding-servingservice is used byvector-retrieverto embed queries at search time.Select devices per processing component (Search and Unified UI mode):
Each processing component picks its device independently. All default to
CPU; set any toGPUorNPUas needed.# Embedding in multimodal-dataprep (in-process) export DATAPREP_EMBEDDING_DEVICE=GPU # Object detection in multimodal-dataprep (CPU, GPU or NPU) export DATAPREP_DETECTION_DEVICE=CPU # Embedding in multimodal-embedding-serving (used by vector-retriever at query time) export MME_EMBEDDING_DEVICE=GPU
There is no “baseline” device, each component is configured directly, matching the Helm chart’s per-component model.
Device note:
DATAPREP_EMBEDDING_DEVICEcontrols in-process embedding execution inmultimodal-dataprep.DATAPREP_DETECTION_DEVICEcontrols object detection inmultimodal-dataprep(CPU,GPUorNPU; check per-model NPU support on the OpenVINO supported models page and test your model —OpenVINO/yolox_s-int8-ov,OpenVINO/yolox_tiny-fp16-ov,OpenVINO/yolox_l-int8-ovandOpenVINO/YOLO11n-fp16-ovare verified on all three; on GPU, YOLOX FP16 variants run faster than INT8).MME_EMBEDDING_DEVICEcontrols embedding execution inmultimodal-embedding-serving, whichvector-retrieveruses to embed queries at search time.ENABLE_EMBEDDING_GPU=trueis a shortcut that setsDATAPREP_EMBEDDING_DEVICE=GPU. For NPU, set the explicit device variables.EXPERIMENTAL: Configure vLLM Intel Arc Pro B-series GPU/XPU Backend:
⚠️ Experimental Feature: Intel Arc Pro B-series GPU (XPU) support with vLLM is in early development stages and may have stability issues. Not recommended for production use.
Enable GPU-accelerated vLLM inference on Intel Arc Pro B-series GPUs (XPUs):
# Enable vLLM with Intel Arc Pro B-series GPU/XPU support (default: false) export ENABLE_VLLM_GPU=true # Set the VLM model for vLLM GPU inference export VLM_MODEL_NAME="Qwen/Qwen3-VL-4B-Instruct"
Additional vLLM GPU/XPU configuration options:
Variable
Description
Default
VLLM_XPU_IMAGEDocker image for vLLM XPU service
vllm/vllm-openai-xpu:v0.31.0VLLM_DTYPEModel precision
bfloat16VLLM_GPU_MEMGPU memory utilization (0-1)
0.8VLLM_MAX_MODEL_LENMaximum sequence length
32000VLLM_MAX_NUM_SEQSMaximum concurrent sequences
256VLLM_BLOCK_SIZEKV cache block size
128VLLM_MAX_NUM_BATCHED_TOKENSMaximum batched tokens
2048VLLM_TENSOR_PARALLEL_SIZETensor parallel size
1VLLM_MM_MAX_PIXELSMaximum pixels per frame (caps resolution, not frame count)
16777216VLLM_HOST_PORTHost port for vLLM service
8200System Requirements:
Intel Arc Pro B-series GPU (e.g., Intel Arc Pro B60, B65, B70)
Intel GPU drivers installed on host
User in
videoandrendergroupsMinimum 16GB GPU memory for the default
bfloat16configuration.
Tip
GPUs with less than 16 GB (for example Intel Arc B580, 12 GB): the default
bfloat16configuration will fail to start withNo available memory for the cache blocks. Use the FP8 checkpoint, which halves the weight footprint, and cap the image size:export VLM_MODEL_NAME="Qwen/Qwen3-VL-4B-Instruct-FP8" export VLLM_GPU_MEM=0.95 export VLLM_MM_MAX_PIXELS=401408 export VLLM_MAX_MODEL_LEN=24000
All four are needed together; omitting any one still fails to start on a 12 GB card at these settings. The FP8 checkpoint cuts the weights,
VLLM_GPU_MEMraises the share of the card vLLM may reserve,VLLM_MM_MAX_PIXELSshrinks the vision encoder, andVLLM_MAX_MODEL_LENlowers the KV cache needed per request.VLLM_MM_MAX_PIXELScaps the resolution of each frame, not how many frames you send: at401408a 1920x1080 frame is downscaled to 832x448, and the defaultPM_MULTI_FRAME_COUNT=12still works unchanged. See Troubleshooting for the sizing formula and how to derive these values for other cards.Behavior when enabled:
Automatically sets
VLLM_HOST=vllm-xpu-serviceDisables OVMS backend completely
Uses the same VLM model for captioning and summarization
Reduces concurrency to prevent GPU memory issues
Launches
vllm-xpu-servicewith GPU device access
Note
See the Run the Application section below for complete usage examples with
ENABLE_VLLM_GPU=true.
Work with Gated Models:
To run a GATED MODEL like Llama models, you will need to pass your huggingface token. You will need to request for an access to a specific model by going to the respective model page on Hugging Face website.
Go to https://huggingface.co/settings/tokens to get your token.
export GATED_MODEL=true
export HUGGINGFACE_TOKEN=<your_huggingface_token>
Once exported, run the setup script as mentioned here. Switch off the GATED_MODEL flag by running export GATED_MODEL=false, once you no longer use gated models. This avoids unnecessary authentication step during setup.
Application Overview#
The Video Search and Summarization application supports multiple deployment modes, each served behind a single nginx reverse proxy on one port. The mode determines which services and UI(s) are brought up.
Note
The application runs on port 12345 by default. You can change this by setting APP_HOST_PORT environment variable to another port number.
Mode |
Command Option |
UI Instances |
Default URL(s) |
|---|---|---|---|
Summary |
|
Single Summary UI |
|
Search |
|
Single Search UI |
|
Dual UI |
|
Separate Summary and Search UIs |
|
Unified UI |
|
Single unified UI (summary + search) |
|
Note
In --summary --search mode, visiting http://<host-ip>:12345/ redirects to the Video Summarization UI.
In modes, where Video Search is available (Search, Dual UI and Unified UI mode), the Vector DB index, the modality of input being used for creating embeddings and the embedding models would differ with modes.
Mode |
Vector-DB Index |
Search Modality |
Environment Variable Used |
|---|---|---|---|
Search |
|
Multimodal embeddings of video frames |
|
Dual UI |
|
Multimodal embeddings of video frames |
|
Unified UI |
|
Text embeddings of generated summaries |
|
Automated Video Ingestion: The Video Search pipeline includes an optional Directory Watcher service for automated video processing. See the Directory Watcher Service Guide for details.
Deployment Options for Video Summarization#
Deployment Option |
Chunk-Wise Summary(1) Configuration |
Final Summary(2) Configuration |
Environment Variables to Set |
Recommended Models |
Recommended Usage Model |
|---|---|---|---|---|---|
OVMS shared-model CPU |
OVMS-hosted VLM on CPU |
Same OVMS-hosted VLM on CPU |
Default |
VLM: |
Default CPU-only summarization flow. |
OVMS shared-model GPU |
OVMS-hosted VLM on GPU |
Same OVMS-hosted VLM on GPU |
|
VLM: |
Single-model OVMS deployment with GPU acceleration. |
OVMS split-model CPU/CPU |
OVMS-hosted VLM on CPU |
OVMS-hosted LLM on CPU |
|
VLM: |
One OVMS instance hosts separate VLM and LLM models on CPU. |
OVMS split-model GPU/CPU |
OVMS-hosted VLM on GPU |
OVMS-hosted LLM on CPU |
|
VLM: |
Use GPU for captioning while keeping final summary on CPU; also supports same-source split by device. |
OVMS split-model CPU/GPU |
OVMS-hosted VLM on CPU |
OVMS-hosted LLM on GPU |
|
VLM: |
Use GPU for final summary while keeping captioning on CPU; also supports same-source split by device. |
OVMS split-model CPU/NPU |
OVMS-hosted VLM on CPU |
OVMS-hosted LLM on NPU |
|
VLM: |
Use NPU for the final-summary LLM while keeping captioning on CPU. The NPU LLM must be a text-only model; see NPU LLM limits. |
vLLM-only CPU |
vLLM-hosted VLM on CPU |
Same vLLM-hosted VLM on CPU |
|
VLM: |
All-vLLM mode for CPU-only deployments. |
vLLM-only GPU/XPU (EXPERIMENTAL) |
vLLM-hosted VLM on Intel Arc Pro B-series GPU |
Same vLLM-hosted VLM on Intel Arc Pro B-series GPU |
|
VLM: |
EXPERIMENTAL: All-vLLM mode with Intel Arc Pro B-series GPU/XPU acceleration. Early-stage feature. |
Note
Chunk-Wise Summary is a method of summarization where it breaks videos into chunks and then summarizes each chunk.
Final Summary is a method of summarization where it summarizes the whole video.
Mixed OVMS+vLLM deployments are not supported in the compose setup. Choose either OVMS-only, vLLM-CPU-only, or vLLM-GPU-only for summarization.
VLM_TARGET_DEVICEsupportsCPU,GPU, orHETERO:GPU,CPU.LLM_TARGET_DEVICEsupports the same values plusNPU.NPU Support: Only the OVMS final-summary LLM can run on NPU. The VLM is not supported on NPU, because NPU vision-language pipelines accept only one image per request while VSS sends several frames per chunk. The NPU LLM must be a text-only, NPU-optimized
-cw-model such asOpenVINO/Phi-3.5-mini-instruct-int4-cw-ov.OVMS mode selection is based on effective VLM/LLM settings: if model source, target device, and compression format are all identical, setup uses shared mode; otherwise it uses split mode.
OpenVINO/models are pre-converted, and the weight format is fixed by the repo name (for example-int8-ov). A same-source split, using the same model on two devices, works by device only. To also change the weight format, pick the other repo variant, such asOpenVINO/Qwen3.5-4B-int4-ov.** EXPERIMENTAL - Intel Arc Pro B-series GPU Support:** vLLM on Intel Arc Pro B-series GPUs/XPUs (
ENABLE_VLLM_GPU=true) is experimental and in early development stages. This feature provides GPU-accelerated inference for both VLM captioning and LLM summarization but may have stability issues and requires specific Intel Arc Pro B-series GPU hardware (e.g., B60, B65, B70). When enabled, it automatically disables OVMS and uses vLLM exclusively. Not recommended for production use.
Recommended OVMS Models#
All OVMS examples use pre-converted OpenVINO/ models, which are pulled as-is without a conversion step. vLLM deployments use Hugging Face checkpoints instead; see the vLLM rows above.
Role and device |
Primary example |
Other examples |
|---|---|---|
VLM on CPU or GPU ( |
|
|
LLM on CPU or GPU ( |
|
|
LLM on NPU ( |
|
|
VLM on NPU |
Not supported |
Run the VLM on CPU or GPU. |
Deployment Options for Video Search#
Deployment Option |
Embedding Pipeline Configuration |
Detection Configuration |
Environment Variables to Set |
Recommended Models |
Recommended Usage Model |
|---|---|---|---|---|---|
Shared CPU |
In-process embedding in |
Detection on CPU |
(defaults — no device overrides) |
Search/Dual: |
Default CPU-only search flow. |
Shared GPU |
In-process embedding in |
Detection on GPU |
|
Search/Dual: |
Run both detection and embedding execution on GPU. |
Shared NPU |
In-process embedding in |
Detection on NPU |
|
Search/Dual: NPU-supported multimodal model |
Run both detection and embedding execution on NPU. |
Split CPU/NPU |
In-process embedding in |
Detection on CPU |
|
Search/Dual: NPU-supported multimodal model |
Keep detection on CPU while offloading embedding to NPU. |
Split CPU/GPU |
In-process embedding in |
Detection on GPU |
|
Search/Dual: |
Offload object detection only while keeping embedding on CPU. |
Split CPU/NPU (Detection) |
In-process embedding in |
Detection on NPU |
|
Search/Dual: |
Offload object detection to NPU while keeping embedding on CPU. |
Note
These options apply to modes where search is enabled:
--search,--summary --search, and--summary-and-search.Each component device defaults to
CPUand is set independently — there is no “baseline” device.Embedding for indexing runs in-process in
multimodal-dataprepand followsDATAPREP_EMBEDDING_DEVICE; object detection followsDATAPREP_DETECTION_DEVICE.MME_EMBEDDING_DEVICEcontrolsmultimodal-embedding-serving, whichvideo-searchuses to embed queries at search time — set it independently of the indexing devices above.NPU Support: Verify model compatibility at the OpenVINO™ Supported Models page before selecting
NPU.ENABLE_EMBEDDING_GPU=trueis a shortcut that setsDATAPREP_EMBEDDING_DEVICE=GPU. For NPU, use explicit device variables.4K/8K videos: By default the shared-memory block size is sized for 1080p frames, so ingesting 4K/8K video can stall the DataPrep worker. Before ingesting high-resolution content, raise
SDK_VIDEO_SHM_BLOCK_SIZEas described in Troubleshooting.
Vector Database Backend (VDMS or Milvus)#
The search path is vector-database agnostic. All vector similarity search is delegated to the standalone vector-retriever microservice — which is part of the search stack for every backend — while video-search keeps only query orchestration and frame-to-video aggregation. Select the backend with the VECTORDB_BACKEND environment variable before running setup.sh (object storage stays on MinIO for both):
|
Vector database |
Query path |
|---|---|---|
|
VDMS |
|
|
Milvus (standalone) |
|
# Default — VDMS backend (no variable needed)
source setup.sh --search
# Milvus backend
export VECTORDB_BACKEND=milvus
source setup.sh --search
Note
For both backends,
video-searchdelegates the raw vector search to the always-onvector-retrievermicroservice and never talks to a vector database directly.VECTORDB_BACKENDselects the backend-bakedvector-retrieverimage (vector-retriever-${VECTORDB_BACKEND}, via theRETRIEVER_BACKENDbuild arg). The frame-to-video aggregation still runs invideo-search, so search results are identical in shape across backends.setup.shapplies exactly one vector-DB overlay, so only the selected backend’s containers start.VECTORDB_BACKEND=vdms(default) addsdocker/compose.search.vdms.yaml, which startsvdms-vector-db;VECTORDB_BACKEND=milvusaddsdocker/compose.search.milvus.yaml, which starts a Milvus standalone stack (milvus-etcd+milvus-standalone) and no VDMS container.multimodal-dataprepwrites embeddings to the selected backend and the matchingvector-retrieverimage reads them back.The similarity metric/index (
VDB_METRIC_TYPE=IP,VDB_INDEX_TYPE=FLAT) is shared betweenmultimodal-dataprepandvector-retrieverand must match on both — the defaults already align.
Important
Important — clean the data when switching backends: embeddings live only in the vector database, but uploaded videos and their metadata live in MinIO and the Pipeline Manager’s PostgreSQL database, which are shared across backends. Switching VECTORDB_BACKEND without clearing them leaves the previously ingested videos visible in the UI while the new backend holds no embeddings for them, so search returns nothing for those videos. Always wipe the user data first:
source setup.sh --clean-data
export VECTORDB_BACKEND=milvus # or vdms
source setup.sh --search
--clean-data stops the stack and removes the MinIO, PostgreSQL, dataprep, VDMS and Milvus data volumes. Model-cache volumes are preserved, so no models are re-downloaded. Alternatively, re-ingest every video after the switch instead of cleaning.
Using Edge Microvisor Toolkit#
If you are running the VSS application on an OS image built with Edge Microvisor Toolkit (EMT), an Azure Linux-based build pipeline for Intel® platforms, the deployment approach depends on the EMT flavor. Refer to the detailed documentation for EMT-D and EMT-S for full details.
EMT-D (Mutable)#
EMT-D is a mutable image that supports standard package management. You can run the VSS setup.sh script directly on the node after installing the required dependencies.
Install the mesa-libGL package (required by the Audio Analyzer service):
sudo dnf install mesa-libGL
# Or using TDNF:
sudo tdnf install mesa-libGL
Install additional tools such as git and wget using the same package manager. Once dependencies are in place, proceed with running the application normally.
EMT-S (Immutable)#
EMT-S is an immutable OS image, standard package managers such as apt are not available, and the VSS setup.sh script cannot be run directly on the EMT-S node (doing so will fail with sudo: apt: command not found). Use one of the following approaches:
Option 1 (USB provisioning): While preparing the USB drive, copy the required Docker images under
/opt/user-appson the image, then flash and deploy the Edge node.Option 2 (Remote copy): On a Ubuntu development system, pull/build all required Docker images and prepare the project directory. Copy the entire directory to the EMT-S node without modifications and deploy from there. This approach has been verified to successfully bring up all VSS containers.
When packages must be installed on EMT-S (for example, mesa-libGL), use the installroot method:
sudo env no_proxy="localhost,127.0.0.1" dnf --installroot=/opt/user-apps/tools/ -y install mesa-libGL
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/user-apps/tools/usr/lib/
The same method applies to any other required packages (for example, git, wget). Refer to the EMT-S documentation for further details.
For GPU deployments on EMT-S host : If the host does not include the host OpenCL runtime required by OpenVINO™ GPU discover, the script uses a conservative integrated-GPU OVMS cache size and continues; inference still runs inside the GPU-enabled OVMS container. Set
OVMS_CACHE_SIZE_GBbefore sourcingsetup.shif you need an explicit cache size.
Run the Application#
Follow these steps to run the application:
Clone the repository and navigate to the project directory:
git clone https://github.com/open-edge-platform/edge-ai-libraries.git edge-ai-libraries -b release-2026.3.0 cd edge-ai-libraries/sample-applications/video-search-and-summarization
Run the setup script with the desired deployment mode:
First, bring down any running deployment before starting:
source setup.sh --stop # or, `source setup.sh --down`
Important
You should always run the above command before changing modes (for example: changing from –summary mode to –search mode).
Clean-up Tip: If you encounter issues or want to completely reset the application data, use
source setup.sh --clean-datato stop all containers and remove all Docker volumes including user data. This provides a fresh start for troubleshooting.Refresh cached images: Docker reuses a local copy of an image tag such as
latesteven when a newer image is published. Add--pull-always(alias--force-pull) to any deployment mode, or setPULL_ALWAYS=true, to check the registry for every image before starting. This applies Composeup --pull alwaysand also refreshes the model-download image:source setup.sh --summary --pull-always # or, with Make make deploy MODE=summary PULL_ALWAYS=true
Images that exist only locally, such as those from
make build, are kept when the registry has no matching image. A registry image with the same name and tag replaces a locally built one, so use a distinctTAGfor local builds. The option is ignored forconfig,--stop,--clean-dataand--setenv.Summary mode: Setup only video summarization:
source setup.sh --summary
Search mode: Setup only video search:
source setup.sh --search
Dual UI mode: Setup both video summarization and video search application with separate UIs:
source setup.sh --summary --search # or, `source setup.sh --search --summary`
When the script finishes, it prints the URLs for the both summary and search UI.
Unified UI mode: Single UI containing video summarization and modified video search:
source setup.sh --summary-and-search # or, `source setup.sh --search-and-summary`
Live metrics: Metrics Manager is disabled by default. Enable it with:
ENABLE_METRICS_MANAGER=true source setup.sh --search # or ENABLE_METRICS_MANAGER=true source setup.sh --summary --search # or ENABLE_METRICS_MANAGER=true source setup.sh --summary-and-search
Directory Watcher: For automated video ingestion into the Search pipeline, see the Directory Watcher Service Guide.
To run Video Summarization with OVMS using a dedicated LLM for final summary:
# Note: If OVMS_LLM_MODEL_NAME is not set, setup falls back to VLM_MODEL_NAME for final summary model source. # Shared vs split mode is then decided from effective model/device/compression equality. # For Summary mode OVMS_LLM_MODEL_NAME="OpenVINO/Qwen3-4B-int8-ov" source setup.sh --summary # For Dual UI mode OVMS_LLM_MODEL_NAME="OpenVINO/Qwen3-4B-int8-ov" source setup.sh --summary --search # For Unified UI mode OVMS_LLM_MODEL_NAME="OpenVINO/Qwen3-4B-int8-ov" source setup.sh --summary-and-search
Use vLLM as the only inference backend (CPU):
ENABLE_VLLM=true source setup.sh --summary # for Summary mode ENABLE_VLLM=true source setup.sh --summary --search # for Dual UI mode ENABLE_VLLM=true source setup.sh --summary-and-search # for Unified UI mode
Note
The vLLM configuration has been tested on Intel® Xeon® 6 processors.
Review docker/compose.vllm.yaml to understand the VLLM engine and environment variables exposed. Modify it as per your use case. Refer to the vLLM Engine Arguments documentation and vLLM Environment Variables documentation for more details.
** EXPERIMENTAL: Use vLLM with Intel Arc Pro B-series GPU/XPU acceleration:**
⚠️ Experimental Feature - Intel Arc Pro B-series GPU Support: Intel Arc Pro B-series GPU/XPU support with vLLM is currently experimental and in early stages. This implementation provides foundational infrastructure for Intel Arc Pro B-series enablement and may require further tuning and optimization. Performance characteristics on Intel Arc Pro B-series hardware are still being evaluated.
System Requirements:
Intel Arc Pro B-series GPU hardware (e.g., B60, B65, B70)
Intel GPU drivers properly installed on the host system
User must be a member of the
videoandrendergroupsAccess to
/dev/dridevicesMinimum 16GB GPU memory for the default
bfloat16configuration. See Configure vLLM Intel Arc Pro B-series GPU/XPU Backend for the settings to use on cards with less memory.
Verify GPU access before enabling:
# Check if /dev/dri devices exist ls -la /dev/dri/ # Verify user group memberships groups | grep -E "video|render"
Usage:
ENABLE_VLLM_GPU=true source setup.sh --summary # for Summary mode ENABLE_VLLM_GPU=true source setup.sh --summary --search # for Dual UI mode ENABLE_VLLM_GPU=true source setup.sh --summary-and-search # for Unified UI mode
Optional: Tune vLLM GPU parameters:
export VLLM_GPU_MEM=0.8 # GPU memory utilization (default: 0.8) export VLLM_MAX_MODEL_LEN=32000 # Maximum sequence length (default: 32000) export VLLM_MAX_NUM_SEQS=256 # Maximum concurrent sequences (default: 256) export VLLM_DTYPE=bfloat16 # Model precision (default: bfloat16) ENABLE_VLLM_GPU=true source setup.sh --summary
Note
When
ENABLE_VLLM_GPU=true, the application automatically setsVLLM_HOST=vllm-xpu-serviceand disables OVMS completelyThe same VLM model is used for both chunk captioning and final summarization (split-model mode is not supported)
Default concurrency is reduced to
PM_VLM_CONCURRENT=1andPM_LLM_CONCURRENT=1to prevent GPU memory pressureReview docker/compose.vllm.xpu.yaml to understand the vLLM XPU configuration. Refer to the vLLM Engine Arguments documentation for parameter details.
For troubleshooting, check container logs:
docker logs vllm-xpu-service
(Optional) Verify the resolved environment variables and setup configurations:
# To just set environment variables without starting containers source setup.sh --setenv # To see the fully resolved compose configuration (defaults to Dual UI mode) source setup.sh config # To see resolved config for a specific mode source setup.sh --summary config # for Summary mode source setup.sh --search config # for Search Mode source setup.sh --summary --search config # for Dual UI Mode source setup.sh --search-and-summary config # for Unified UI Mode # To see resolved configurations for OVMS split-model summarization without starting containers. # (for other modes, combine --summary with --search option or replace all options with --summary-and-search) OVMS_LLM_MODEL_NAME="OpenVINO/Qwen3-4B-int8-ov" source setup.sh --summary config # To see resolved configurations for summarization services with vLLM enabled without starting containers. # (for other modes, combine --summary with --search option or replace all options with --summary-and-search) ENABLE_VLLM=true source setup.sh --summary config
Use GPU/NPU Acceleration#
Note
Offloading models to different devices (e.g., VLM on GPU and LLM on NPU) is only supported with the OVMS backend. The vLLM backend runs a single model on a single device.
⚠️ NPU Support is Experimental: Only the final-summary LLM can run on NPU, and it must be a text-only, NPU-optimized -cw- model. The VLM is not supported on NPU: NPU vision-language pipelines accept only one image per request, while VSS sends several frames per chunk. Run the VLM on CPU or GPU.
Note
To bring down a running deployment before re-running with different options, run:
source setup.sh --stop # or, `source setup.sh --down`
Use GPU acceleration for VLM inference#
# for Summary mode
VLM_TARGET_DEVICE=GPU source setup.sh --summary
# for Dual UI mode
VLM_TARGET_DEVICE=GPU source setup.sh --summary --search
# for Unified UI mode
VLM_TARGET_DEVICE=GPU source setup.sh --summary-and-search
Use GPU acceleration for the OVMS final-summary LLM#
# for Summary mode
LLM_TARGET_DEVICE=GPU OVMS_LLM_MODEL_NAME=OpenVINO/Qwen3-4B-int8-ov source setup.sh --summary
# for Dual UI mode
LLM_TARGET_DEVICE=GPU OVMS_LLM_MODEL_NAME=OpenVINO/Qwen3-4B-int8-ov source setup.sh --summary --search
# for Unified UI mode
LLM_TARGET_DEVICE=GPU OVMS_LLM_MODEL_NAME=OpenVINO/Qwen3-4B-int8-ov source setup.sh --summary-and-search
Note
Search-indexing embedding runs in-process in multimodal-dataprep and follows DATAPREP_EMBEDDING_DEVICE. MME_EMBEDDING_DEVICE controls multimodal-embedding-serving, which video-search uses for query-time embeddings. ENABLE_EMBEDDING_GPU=true is a shortcut that sets DATAPREP_EMBEDDING_DEVICE=GPU.
To offload only specific DataPrep components during search (for example, keep embedding on CPU and move detection to GPU):
DATAPREP_EMBEDDING_DEVICE=CPU DATAPREP_DETECTION_DEVICE=GPU source setup.sh --search
To verify the configuration and resolved environment variables without running the application:
Use NPU acceleration for the final-summary LLM (split-model mode)#
# for Summary mode
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh --summary
# for Dual UI mode
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh --summary --search
# for Unified UI mode
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh --summary-and-search
Important
The LLM on NPU must be a text-only model, so set OVMS_LLM_MODEL_NAME explicitly. The setup script stops with an error when the model serving the LLM role on NPU is a vision-language model, because raising the NPU prompt limit breaks vision-language models on NPU and the 1024-token default is too small for a summary.
For pre-converted OpenVINO/* repos on NPU, use an NPU-optimized channel-wise repo whose name contains -cw-, such as OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov or OpenVINO/Qwen3-8B-int4-cw-ov. Other variants, such as -int4-gq-ov, cannot be downloaded for NPU, and -cw- repos cannot be downloaded for CPU or GPU. The setup script stops with an error before the download in either case.
An NPU LLM is compiled for a fixed prompt length. The setup script writes the limits below into the model’s graph.pbtxt and sizes Pipeline Manager’s summarization prompts to fit:
Variable |
Default |
Effect |
|---|---|---|
|
|
Maximum prompt tokens. Longer prompts are rejected by OVMS. |
|
|
Response tokens reserved beyond the prompt. Output past |
|
|
Character budget for each summarization prompt; assumes at least 2 characters per token. |
|
|
Final-summary completion cap. |
Explicit PM_LLM_MAX_CONTEXT_LENGTH or PM_SUMMARIZATION_MAX_COMPLETION_TOKENS values are kept, with a warning when they exceed the NPU limits. Long videos are summarized in more map-reduce rounds. A single audio transcript longer than the character budget is still sent in one request and fails on NPU; disable audio or keep the LLM on GPU/CPU for long, speech-heavy videos. The NPU reserves memory for the full prompt limit at load time, so larger OVMS_NPU_MAX_PROMPT_LEN values increase system memory use, model load time, and the processing time of long prompts. Measured with Phi-3.5-mini-instruct-int4-cw-ov: about 7 GB at 4096, 13 GB at 8192, and 20 GB at 16384. On systems with 16 GB of memory or less, set OVMS_NPU_MAX_PROMPT_LEN=4096.
Use GPU acceleration for search-indexing embedding#
# for Search mode
ENABLE_EMBEDDING_GPU=true source setup.sh --search
# for Dual UI mode
ENABLE_EMBEDDING_GPU=true source setup.sh --summary --search
# to run the in-process embedding on NPU instead
DATAPREP_EMBEDDING_DEVICE=NPU source setup.sh --search
Verify the configuration and resolved environment variables#
These commands help to validate the deployment configuration without actually deploying the application.
# For VLM inference on GPU
VLM_TARGET_DEVICE=GPU source setup.sh config --summary # for Summary mode
VLM_TARGET_DEVICE=GPU source setup.sh config --summary --search # for Dual UI mode
VLM_TARGET_DEVICE=GPU source setup.sh config --summary-and-search # for Unified UI mode
# For LLM on NPU (split-model mode)
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh config --summary # for Summary mode
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh config --summary --search # for Dual UI mode
LLM_TARGET_DEVICE=NPU OVMS_LLM_MODEL_NAME=OpenVINO/Phi-3.5-mini-instruct-int4-cw-ov source setup.sh config --summary-and-search # for Unified UI mode
# For search-indexing embedding on GPU
ENABLE_EMBEDDING_GPU=true source setup.sh config --search # for Search mode
ENABLE_EMBEDDING_GPU=true source setup.sh config --search --summary # for Dual UI mode
# For search-indexing embedding on NPU
DATAPREP_EMBEDDING_DEVICE=NPU source setup.sh config --search
# For component-specific offload in DataPrep
DATAPREP_EMBEDDING_DEVICE=CPU DATAPREP_DETECTION_DEVICE=GPU source setup.sh --search config
Tip
VLM_TARGET_DEVICE and LLM_TARGET_DEVICE support values: CPU (default), GPU, or HETERO:GPU,CPU for heterogeneous execution with fallback. LLM_TARGET_DEVICE also supports NPU; the VLM is not supported on NPU.
Note
Avoid setting the ENABLE_VLM_GPU, ENABLE_OVMS_LLM_SUMMARY_GPU, or ENABLE_EMBEDDING_GPU flags explicitly on the shell using export, because you need to switch these flags off as well, to return to the CPU configuration.
Access the Application#
After successfully starting the application, access the application UI on following URLs based on chosen mode:
--summary mode#
UI |
URL |
|---|---|
Video Summarization |
|
--search mode#
UI |
URL |
|---|---|
Video Search |
|
--summary --search mode#
UI |
URL |
|---|---|
Video Summarization |
|
Video Search |
|
Visiting the root URL http://<host-ip>:12345/ redirects to the Video Summarization UI.
--summary-and-search mode#
UI |
URL |
|---|---|
Unified Summary/Search |
|
Customizing Application Port#
The port where we access the application is customizable by setting the
APP_HOST_PORTenvironment variable (default12345).
Monitoring OVMS Metrics#
When running in summary mode with OVMS, Prometheus-compatible metrics are available at http://<host-ip>:12345/ovms/metrics. These metrics provide insights into inference performance:
curl http://localhost:12345/ovms/metrics
Key metrics include ovms_requests_success, ovms_inference_time_us, and ovms_current_requests. See Deploy with Helm - Monitoring and Metrics for the full metrics list.
CLI Usage#
Refer to CLI Usage for details on using the application from a text user interface (terminal-based UI).
Running in Kubernetes Cluster#
Refer to Deploy with Helm for the details. Ensure the prerequisites mentioned on this page are addressed before proceeding to deploy with Helm chart.
Advanced Setup Options#
For alternative ways to set up the sample application, see How to Build from Source
Supporting Resources#
Connect AI Agents via MCP#
Once VSS is running, you can connect AI agents to it using the included MCP server. The MCP server exposes VSS capabilities as typed tools and read-only resources for any MCP-compatible client.
See the MCP Server guide for setup instructions.