NVIDIA AI Blueprint: Video Search and Summarization
A GPU-accelerated reference architecture for building video analytics agents that search, summarize and answer questions over live or recorded video using vision-language and language models plus retrieval-augmented generation.
NVIDIA AI Blueprint for Video Search and Summarization (VSS) is a reference architecture for building GPU-accelerated video AI agents that search, analyze, summarize and reason over live or recorded video using natural language. It combines vision-language models such as NVIDIA Cosmos, large language models such as NVIDIA Nemotron, retrieval-augmented generation and NVIDIA NIM microservices to support searching video streams or archives, summarizing hours of footage, answering visual questions with automatic report generation, and detecting and verifying real-time alerts.
It is organized into three processing areas: real-time video intelligence, which extracts features, embeddings and stream understanding published to a message broker; downstream analytics, which enriches that metadata into trajectories, incidents and verified alerts; and agentic and offline processing, which orchestrates tools for search, question answering, summarization and clip retrieval, including through the Model Context Protocol.
It powers NVIDIA's own build.nvidia.com experience for natural-language video agents and is written in Python.