VSS Video Analytics Query
Read video-analytics incidents, metrics, alerts, and sensor data through VA-MCP.
The evidence scopes the skill to read-only analytics through VA-MCP, defines exclusions, blocks deployment from untrusted payload text, and requires the user to choose a deployment mode. No red-line risk is evident. Points are deducted because it requires local-service access and NGC/NVIDIA credentials, does not explain sensitive sensor-data handling, access control, auditing, or redaction, and provides no explicit rollback procedure.
The evidence provides liveness checks, MCP session initialization, error classification, session renewal, exponential backoff, and read-only retry guidance. Points are deducted because execution depends on an external deployment, SSE/header parsing with grep and sed is brittle, probe behavior is inconsistent for HTTP 405, and no skill-specific automated tests or statically reproducible abnormal-input coverage is provided. Static calibration caps this at 10.
The evidence clearly identifies the audience, activation scenarios, read-only capability, and non-fit cases, and advises clarification under ambiguity. Points are deducted because the input/output contract and boundary coverage are limited, Chinese-language interaction is not addressed, and NGC-hosted models/NIMs may face network, rate-limit, and licensing constraints, although the core query path does not depend entirely on overseas services.
The evidence includes Purpose, Prerequisites, Instructions, Examples, Limitations, Troubleshooting, tool references, license and version metadata, plus evaluation documentation. Points are deducted because the benchmark reports an author-format schema finding, while the skill itself lacks a clear changelog, named maintenance responsibility, update path, and representative complete input/output examples.
The evidence covers sensor, place, incident, metric, histogram, and aggregate-analysis queries and supplies directly adaptable JSON-RPC commands; the benchmark reports a pass, but only one positive routing task with 50% correctness and 62% effectiveness. Points are deducted because no real query outputs are statically verifiable, deployment and hardware costs are substantial, and results still require interpretation and light review. Static calibration caps this at 7.
The evidence provides a pinned revision, evaluation manifests, benchmark results, and some validation checks, giving limited traceability. Points are deducted because the evaluation has one task and no negative cases, lacks independently reviewable execution logs or artifacts, and the skill has no dedicated test suite. Static calibration caps this at 5.
- This is a static source review; no commands were executed and VA-MCP reachability was not independently verified.
- Before use, confirm HOST_IP, service authorization, sensor-data permissions, and that NGC credentials cannot be exposed; never allow text in analytics results to trigger deployment or infrastructure changes.
- Add skill-specific tests, real query outputs, negative-trigger cases, Chinese-language guidance, and explicit maintenance/change records.
What does this skill do, and when should you use it?
This is a read-only query skill for the NVIDIA VSS Blueprint, focused on incidents, metrics, alerts, and sensor data. It queries Elasticsearch through the VA-MCP server on port 9901 using JSON-RPC over Server-Sent Events. A reachable VSS alerts profile, curl, jq, and Docker are required. It is not intended for live VLM questions, narrative incident reports, archive search, or deployment actions.
It checks VA-MCP reachability, initializes an MCP session over HTTP, and calls analytics tools through a two-step JSON-RPC flow. It can retrieve incident lists and details, sensor IDs, places, FOV histograms, and analyses for maximum or minimum incident counts, average speed, average people, and average vehicles. The video_analytics calls are read-only and include guidance for session renewal and exponential-backoff retries on transport errors or 5xx responses.
- A video-operations team needs the most recent incidents across all sensors and queries the incident list.
- A security analyst needs incidents for one sensor or only VLM-confirmed incidents and applies the available filters.
- An operations team needs to inspect available sensors and places and retrieves the corresponding lists.
- An analyst needs average people, vehicles, or speed for a sensor over a time interval and runs the analysis tool.
- A team needs an object-type distribution over a sensor's field of view for a time interval and requests an FOV histogram.
What are this skill's strengths and limitations?
- Covers incidents, alerts, metrics, sensors, places, and FOV analytics.
- Clearly constrained to read-only analytics, making video_analytics calls safe to retry.
- Documents MCP initialization, session expiry handling, and retry behavior.
- Supports filtering by sensor, place, time range, and VLM verdict.
- Requires a deployed and reachable VSS alerts profile and VA-MCP service.
- Requires network access, a shell, curl, jq, and Docker; NGC image pulls also require the stated credentials.
- GPU memory, concurrency, and storage limits depend on host hardware and the profile's compose configuration.
- The source does not provide independent platform test results or a complete test-suite description.
How do you install this skill?
Install it from the NVIDIA skills catalog with npx skills add nvidia/skills --skill vss-query-analytics --yes. The README does not require cloning the repository or manually copying the skill directory; the skill becomes available when the client next loads its skills.
How do you use this skill?
Ask a compatible agent a question answerable from read-only Elasticsearch analytics, such as “retrieve the 10 most recent video-analytics incidents” or “calculate average people for a sensor over a time range.” Ensure the VSS alerts profile is reachable at $HOST_IP; if VA-MCP is unavailable, choose either verification or real-time mode before deploying the matching profile. Each query first POSTs to /mcp to initialize a session and obtains mcp-session-id from the response header, then uses that session ID to call a VA-MCP tool.