video-search-and-summarization

Build GPU-accelerated video AI agents that search, analyze, summarize, and reason over live or recorded video using natural language.

NVIDIA AI Blueprint for Video Search and Summarization (VSS) combines vision-language models, RAG, and NVIDIA NIM microservices to deliver real-time video analytics, visual Q&A, alert verification, clip retrieval, and long-video summarization.

• Search video streams or archives using natural language queries
• Summarize hours of video
• Ask visual questions and automatically generate reports
• Detect and verify real-time alerts with VLMs

🚀 Try the Demo · ⚡ Quickstart · 📚 Documentation · 🏗️ Architecture · 📦 Latest Release

Table of Contents

• Overview
• Use Case / Problem Description
• Agent Workflows
• Software Components
• Target Audience
• Repository Structure Overview
• Documentation
• Prerequisites
• Hardware Requirements
• Quickstart Guide
• Contributing
• License

Overview

The NVIDIA Blueprint for Video Search and Summarization (VSS) provides a suite of reference architectures for building vision agents and AI-powered video analytics applications. Those architectures bring together accelerated vision microservices, vision language models (VLMs), and large language models (LLMs) so you can use them in existing applications, as standalone microservices, or as part of a larger vision agent.

VSS is organized into three areas of processing and analysis: real-time video intelligence (feature extraction, embeddings, and stream understanding with results published to a message broker), downstream analytics (enrichment of metadata into trajectories, incidents, and verified alerts), and agentic and offline processing (orchestrated tools for search, Q&A, summarization, and clip retrieval, including via the Model Context Protocol).

This repository implements the blueprint and powers the NVIDIA build experience for natural-language video agents—search, summarization, visual Q&A, and related workflows—backed by generative AI, VLMs and LLMs, and NVIDIA NIM microservices as configured in the stacks below.

Use Case / Problem Description

The NVIDIA AI Blueprint for Video Search and Summarization addresses the challenge of deploying visual agents capable of interacting with large volumes of video data, both stored and streamed. This can be used to create vision AI agents, that can be applied to a multitude of use cases such as monitoring smart spaces, warehouse automation, and SOP validation. This is important where quick and accurate video analysis can lead to better decision-making and enhanced operational efficiency.

Agent Workflows

We provide multiple reference Agent Workflows which demonstrate how the individual components can be leveraged by an agent:

Workflow	Description
Q&A and Report Generation (Quickstart)	Video retrieval, VLM-based Q&A, and report generation on short video clips
Alert Verification	Realtime processing of videos using perception (object detection, tracking) and behavior analytics to generate alerts, which are subsequently verified with VLM to reduce false positives
Real-Time Alerts	Continuous processing of video streams through VLM for anomaly detection
Video Search	Natural language search across video archives using video embeddings (alpha)
Long Video Summarization	Analysis and summarization of extended video recordings through chunking and aggregation of dense captions

Software Components

1. NIM microservices: Here are models used in this blueprint:

   • Cosmos3 Nano Reasoner
   • NVIDIA Nemotron 3.5 Lightning 30B A3B

2. Real-time video intelligence: The Real-Time Video Intelligence layer extracts rich visual features, semantic embeddings, and contextual understanding from video data in real-time, publishing results to a message broker for downstream analytics and agentic workflows. It provides three core microservices for processing video streams.

3. Downstream analytics: The Downstream Analytics layer processes and enriches the metadata streams generated by real-time video intelligence microservices, transforming raw detections into actionable insights and verified alerts.

4. Agent and offline processing: The top-level agent leverages the Model Context Protocol (MCP) to access video analytics data, incident records, and vision processing capabilities through a unified tool interface. It integrates multiple vision-based tools including video understanding with Vision Language Models (VLMs), semantic video search using embeddings, long video summarization for extended footage analysis, and video snapshot/clip retrieval.

Target Audience

This blueprint is designed for ease of setup with extensive configuration options, requiring technical expertise. It is intended for:

1. Video Analysts and IT Engineers: Professionals focused on analyzing video data and ensuring efficient processing and summarization. The blueprint offers 1-click deployment steps, easy-to-manage configurations, and plug-and-play models, making it accessible for early developers.

2. GenAI Developers / Machine Learning Engineers: Experts who need to customize the blueprint for specific use cases. This includes modifying the pipelines for unique datasets and fine-tuning LLMs as needed. For advanced users, the blueprint provides detailed configuration options and custom deployment possibilities, enabling extensive customization and optimization.

Repository Structure Overview

Directory	Description
services/agent/	Video search and summarization agent (Python). Contains packages/vss_agents/ (tools, agents, APIs, embeddings, evaluators, video analytics), scripts/, docker/, and 3rdparty/. See services/agent/README.md.
services/ui/	Frontend monorepo (Next.js, Turbo): apps/ (nv-metropolis-bp-vss-ui, vss-chat) and shared packages/. See services/ui/README.md.
services/rtvi/	Real-time video intelligence microservices: rt-vlm/ (dense captioning and incidents from live streams), rt-embed/ (video, image, and text embeddings), rt-cv/ (2D detection and tracking), and rt-cv-3d/ (RT-DETR plus Multi-View 3D Tracking for fused BEV output).
services/vios/	Video I/O Service: the ingest and storage layer. Connects to RTSP/ONVIF/Milestone cameras, VMS systems, and NVStreamer sources, manages sensor registration and recording, and serves live and recorded video plus a dashboard. See services/vios/README.md.
services/alert/	Configuration-driven alerts microservice: VLM-based alert verification, realtime alert generation, and on-demand clip verification. See services/alert/README.md.
services/video-summarization/	Long video summarization (LVS) microservice. Processes video with VLMs and returns timestamped captions, structured event detections, and aggregated summaries over a REST API. See services/video-summarization/README.md.
services/sdrc/	Sensor Distribution and Routing Controller: the stream-placement and traffic-routing layer that assigns live streams to backend workers and keeps that routing stable and recoverable. See services/sdrc/README.md.
services/configurators/	Configuration services: vss-configurator/ (unified sensor and hardware-profile configuration) and vss-rt-config-adaptor/ (writes DeepStream runtime configs from posted metadata).
services/analytics/	Downstream analytics services for processing real-time video intelligence metadata. Contains behavior analytics stream processing and REST APIs for querying analytics results.
services/analytics/behavior-analytics/	Python streaming pipeline for spatial AI analytics, incident detection, Smart City, warehouse, playback, and other behavior analytics applications. Includes app entry points, configs, Docker support, tests, and detailed guides. See services/analytics/behavior-analytics/README.md.
services/analytics/video-analytics-api/	Node.js and Express REST API service for VSS Video Analytics data. Exposes metrics, tracker, frames, behavior, clustering, events, sensor, config, alerts, and incidents endpoints backed by Elasticsearch. See services/analytics/video-analytics-api/README.md.
deploy/	Deployment configs, Docker Compose, and Helm charts: NIM model configs, developer profiles (dev-profile-base, dev-profile-search, dev-profile-alerts, dev-profile-lvs), foundational services, LVS, RTVI, VLM-as-verifier, VST, and root compose.yml. Also contains deploy/docker/scripts/ — the Brev launchable notebooks (base, warehouse, smart cities) and dev-profile / patch scripts.
tools/message-broker-consumers/	Multiprocessing Redis and Kafka consumers that decode VSS protobuf messages from streams/topics and export them as JSON Lines files for inspection, debugging, or offline processing. See tools/message-broker-consumers/README.md.
tools/sdg-postprocessing/	Dataset post-processing utilities for synthetic data generation workflows: semantic labeling helpers, raw data sanity checks, RGB/depth/video conversion, and ground-truth conversion for MTMC-compatible datasets. See tools/sdg-postprocessing/README.md.
tools/rtvi-cv-mv3dt-utils/	Offline utilities for generating MV3DT RTVI-CV configuration artifacts, including per-camera camInfo projection configs and MQTT publish/subscribe topology files for warehouse MV3DT deployments. See tools/rtvi-cv-mv3dt-utils/README.md.
tools/logstash-plugins/	Custom Logstash plugins for the ELK ingest pipeline. Includes input/redis-stream/ (logstash-input-redis_stream), a Java input plugin that reads from Redis Streams via consumer groups (XREADGROUP) with optional Protocol Buffer decoding. See tools/logstash-plugins/input/redis-stream/README.md.
skills/	agentskills.io-compatible agent skills for VSS: self-contained skill directories with SKILL.md frontmatter, grouped by category (deployment/, operations/, tools/, benchmarking/) plus the top-level vss-build-vision-ai. Covers deploy and usage of search, summarization, alerts, VIOS, RT-VLM, LVS, and other related workflows—see the catalog and install notes in skills/README.md.
libs/analytics/spatialai-data-utils/	Spatial AI Data Utils (SDU): NVSchema / ground-truth / calibration / Sparse4D loaders, camera calibration + grouping (BEV group-origin / per-group fan-out), 3D↔2D geometry, multi-cam 3D-bbox visualization, detection (mAP) + tracking (HOTA, CLEAR, identity, count) evaluators, NVSchema result converters, and video↔frame utilities. See libs/analytics/spatialai-data-utils/README.md.
libs/vss/	The vss command-line interface (cli/) and its shared core (core/) — the host-side entry point to a deployed stack, used by the skills above. See libs/vss/cli/README.md.
libs/nvschema/	Protobuf schema shared by the services that publish and consume VSS metadata. See libs/nvschema/README.md.
.openclaw/, .hermes/	Self-contained agent harnesses that package the harness-neutral skills/ tree for OpenClaw and Hermes. See .openclaw/README.md and .hermes/README.md.

Documentation

For detailed instructions and additional information about this blueprint, please refer to the official documentation.

Prerequisites

Obtain API Key

• NVIDIA AI Enterprise developer licence required to local host NVIDIA NIM.
• API catalog keys:
  • NVIDIA API catalog or NGC (steps to generate key)

Container Image Availability

Development Container Notice: Images tagged develop-* or nightly-* are pre-release artifacts intended for development and testing only. They may change without notice and are provided “AS IS.” Official release images are published through NVIDIA NGC. For production deployments, use only a versioned release image from NGC that has completed NVIDIA’s release review.

Hardware Requirements

The platform requirement can vary depending on the configuration and deployment topology used for VSS and dependencies like VLM, LLM, etc. For a list of validated GPU topologies and what configuration to use, see the GPU requirements.

Quickstart Guide

Launchable Deployment

Ideal for: Quickly getting started with your own videos without worrying about hardware and software requirements.

Follow the steps from the documentation and notebook in deploy/docker/scripts directory to complete all pre-requisites and deploy the blueprint using Brev Launchable in a 2xRTX PRO 6000 SE AWS instance.

• deploy/docker/scripts/deploy_vss_launchable.ipynb: This notebook is tailored specifically for the AWS CSP which uses Ephemeral storage.

Docker Compose Deployment

Ideal for: Deploying a VSS agent on your own hardware or bare metal cloud instance.

System Requirements

• OS:
  • x86 hosts: Ubuntu 24.04
  • DGX Station (GB300): DGX OS 7.6.0
  • DGX-SPARK: DGX OS 7.5.0
  • IGX-THOR (T5000 and T7000): IGX-SW 2.0 Production (LTS), including IGX OS 2.0 (Ubuntu 24.04)
  • Jetson AGX-THOR: JetPack 7.2 (Jetson Linux/L4T r39.2)
  • Jetson AGX Orin and Jetson Orin NX: public JetPack 7.2.2 release
• NVIDIA Driver:
  • 595.58.03 (x86 hosts with Ubuntu 24.04)
  • 595.58.03 (DGX Station GB300)
  • 580.173.02 (DGX-SPARK)
  • 580.00, bundled with IGX-SW 2.0 (IGX-THOR)
  • 595.78 (Jetson AGX Orin, AGX-THOR, and Jetson Orin NX)
• NVIDIA Container Toolkit: 1.17.8+
• Docker Engine: 28.3.3 <= Docker Engine < 29.5.0
• Docker Compose: v2.39.1+
• NGC CLI: 4.10.0+

Docker upper bound: Docker Engine 29.5.0+ may fail pulling NGC-hosted images. Use Docker Engine 28.3.3 or another supported version below 29.5.0.

Please refer to Prerequisites section here for installation details.

Contributing

See CONTRIBUTING.md for the development workflow, branch naming convention, and PR guidelines.

License

This project's licensing is documented across the following files:

• LICENSE — Apache-2.0 for the repository. Elasticsearch, Kibana, and Redis configuration files remain dual-licensed as described there.
• LICENSE-3rd-party.txt — index of per-service third-party attribution files, including services/ui/LICENSE-3rd-party.txt for the UI image.
• LICENSE.DATA — NVIDIA Asset License.