vss-query-analytics

vss-query-analytics is an NVIDIA operational skill for reading video-analytics data from a deployed Video Search and Summarization (VSS) stack. It solves the problem of answering read-only analytics questions — incidents, alerts, metrics, occupancy, speeds, and sensor telemetry — by routing them through the VA-MCP server (MCP JSON-RPC on port 9901), which queries data stored in Elasticsearch. It emphasizes a two-step shell pattern that the agent must actually execute and relay, rather than guessing or describing results.

The skill is explicitly read-only and includes a strong scope guard: despite a deliberately broad trigger list, it must only be used for questions answerable by reading Elasticsearch via VA-MCP, and it defers ad-hoc VLM Q&A, narrative reports, archive search, and deploy/teardown actions to other dedicated skills. It probes the VA-MCP endpoint before querying and, if unreachable, asks the user which mode to deploy rather than acting on its own. Notably, it is prompt-injection aware: it instructs the agent to treat alert and analytics payloads as untrusted input that may contain attacker-controlled text, and never to auto-invoke deployment based on a use-case string in a payload — auto-deploy requires a trusted harness flag.

It targets developers and operators who need to query metrics and incidents from an operational VSS deployment. Prerequisites are a reachable VSS deployment, NGC credentials via environment variables for image pulls, and curl, jq, and Docker. The output is shell commands and API calls against a local analytics endpoint, with no destructive operations or exfiltration; NVSkills-Eval reported a 100% Security score.