What problem does it solve? Searching hours of archived surveillance or operational video manually is slow and error-prone. This Skill lets you query ingested video archives with plain natural language (e.g. "find all instances of forklifts") through the NVIDIA VSS agent backend, and handles ingesting video files or RTSP streams so they become searchable via Cosmos Embed1 embeddings stored in Elasticsearch. ## Core Features & Use Cases - Natural-language video search: Run fusion search (embedding + attribute reranking with an optional VLM critic) through the VSS agent's /generate endpoint and receive timestamped, similarity-scored clips formatted as a professional inspection report. - Agent-backed ingestion: Ingest video files via the three-step chunked-upload handshake or register RTSP streams so embeddings are generated; delete sources with cleanup of both VIOS storage and Elasticsearch indexes. - Guided troubleshooting: Built-in feedback loops for zero-hit results, missing sources, and backend failures, plus discovery modes (wide-net, high-precision, metadata-filtered) for tuning recall vs. precision. - Use Case: A safety officer asks "when did someone enter the restricted area?" — the Skill verifies the source is registered, runs the search, and returns a Video Search Results report with clips, similarity scores, and an optional screenshot verification step. ## Quick Start Ask the agent to find all instances of forklifts in the ingested warehouse video, providing the VSS agent backend host IP.
