vss-summarize-video

VSS Summarize Video produces a single polished narrative summary of one recorded video clip, with timestamped events when NVIDIA's LVS summarization microservice is reachable, and a direct VLM NIM fallback when it is not. It solves the focused problem of turning a recorded clip into a scenario-aware summary, and it deliberately scopes itself narrowly: it is not for live RTSP captioning, incident or alert-window report generation, or semantic search across an archive, each of which is delegated to a sibling skill.

The skill drives the workflow by calling the summarization microservice or VLM endpoint directly with `curl` and `jq`, following reference-routed sections for the summarization API (endpoints such as `/v1/summarize`, health probes, request fields, and response shapes), deployment and service ops (the `lvs` profile, ports, env vars, backend and datastore selection across Elasticsearch/Neo4j/ArangoDB), environment variables, and debugging. A human-in-the-loop gate is central: before any summarize call the skill must collect `scenario`, `events`, and optional `objects_of_interest` from the user and wait for a response, applying generic defaults only on explicit opt-in. It also documents troubleshooting for warm-up 503s, empty results, Cosmos reasoning `<think>` blocks, and checking HTTP status rather than response bodies.

Target users are operators and developers running the NVIDIA Video Search and Summarization blueprint who need to summarize recorded clips reliably. It assumes network reachability to the endpoints, fetchable clip URLs, and `jq`/`curl` on the agent host, and it is explicit about limitations: the VLM fallback uses one fixed prompt with lower quality, remote endpoints often cannot reach localhost/private clip URLs, and each request makes a single backend call with no parallel hedging.