VSS Video Q&A
Answers visual questions by having an agent inspect a recorded video clip.
The skill restricts itself to fresh visual questions, requires VSS probing and sensor listing, and requires confirmation before upload. These provide scope control and partial data-flow transparency. Points are deducted because upload creates an external write, while authentication, privacy/sensitive-video handling, privilege isolation, and complete rollback guidance are absent; curl may transmit local video data to the service.
SKILL.md specifies probing, sensor checks, upload, /generate invocation, and agent-think removal, and defines stop-or-ask behavior when probing fails. Points are deducted for limited handling of non-2xx responses, timeout diagnostics, upload conflicts, malformed responses, and missing dependencies; BENCHMARK.md also records insufficient connection/error guidance. Static evidence cannot justify a score above 10.
Activation scenarios, exclusions, accepted clip identifiers, and VSS profile prerequisites are reasonably clear, distinguishing visual questions from metadata or existing-result questions. Points are deducted because Chinese interaction/localization is not documented and mainland-China network reachability is unverified; ambiguity handling for URLs, sensor IDs, and filename stems remains incomplete.
The material includes a name, description, Apache-2.0 license, version 3.2.0, product repository link, cross-references, output contract, and an evaluation report, with generally readable structure. Points are deducted for missing metadata.author, missing recommended Instructions and Examples sections, absent changelog and explicit maintenance/update path, and incomplete troubleshooting and limitation disclosure; repository license metadata is NOASSERTION, so governance evidence is not fully consistent despite the skill-level Apache-2.0 declaration.
The skill supplies a directly usable VSS agent /generate request pattern, sensor preflight, and final-text extraction procedure, so it plausibly completes ad-hoc visual Q&A for one clip. Points are deducted because the benchmark has only one task, correctness and effectiveness are both 50%, and static review cannot verify answer quality, timestamp accuracy, or coverage of video content; human review remains necessary.
Committed benchmark material, eval configuration, and reported security, correctness, discoverability, and efficiency signals provide limited auditable evidence. Points are deducted because coverage is one task with no negative cases, there is no skill-specific reproducible test suite, and no independent third-party corroboration; static evidence caps this at 5.
- Before execution, confirm that the VSS profile, VST, and video_understanding tool are available, and validate HOST_IP, sensor-id, and video path.
- Video upload is an external write; check for sensitive content, conflicts, and authorization before transmitting data.
- Do not treat VLM output as guaranteed fact, especially for safety judgments and timestamps; retain human review.
- Mainland-China network reachability is not demonstrated in the source material, and deployment dependencies may require locally reachable configuration.
What does this skill do, and when should you use it?
VSS Video Q&A calls the VSS Agent’s video_understanding tool for a fresh visual question about a recorded clip. It is intended for questions about people, objects, actions, colors, timing, safety, and other facts that require inspecting video frames. It requires a running VSS profile that serves video_understanding and a matching VST sensor. Version 3.2.0 is declared under Apache-2.0.
It probes the VSS Agent, lists VST sensors, and checks that the requested sensor or filename stem exists. If it is missing, it can upload the video after interactive confirmation and verify the sensor again. It then posts a request to the VSS Agent’s /generate endpoint containing the sensor and question, asks the agent to call video_understanding, reads the response value, removes any agent-think sections, and returns the remaining final answer.
- A video operator needs an answer about what happens in a specific recorded clip.
- A safety analyst needs to verify whether particular people, objects, actions, colors, or safety-relevant events appear in footage.
- A user has a coarse summary or report and needs a follow-up answer about visual content details.
- Existing search results, metadata, or prior tool output cannot answer a content question and the original pixels must be checked.
What are this skill's strengths and limitations?
- The workflow is clearly scoped to fresh, frame-based video questions.
- It specifies agent probing, sensor validation, upload confirmation, and response cleanup.
- It prevents using the skill when metadata, search results, or prior output already answers the question.
- The skill declares an Apache-2.0 license.
- It depends on a running VSS Agent, the video_understanding tool, and VST sensor services.
- The sensor-list check is mandatory before every /generate call.
- It is not intended for questions answerable from prior tool output, search hits, or metadata.
- The supplied source provides no platform matrix, performance measurements, or skill-specific test results.
How do you install this skill?
Install the specific skill with the documented skills CLI command:
npx skills add nvidia/skills --skill vss-ask-video --yes
The README says cloning the repository or manually copying skill folders is not required.
How do you use this skill?
After installation, ask the agent a fresh visual question tied to a specific sensor, for example: "Call video_understanding to answer whether anyone enters the restricted area in sensor warehouse_safety_0001."
Before the request:
- Confirm that a VSS profile serving video_understanding is running; base is recommended and lvs is also listed.
- List VST sensors and confirm the sensor ID or filename stem exists.
- If it does not exist, obtain interactive confirmation before uploading the video, then list sensors again.
- Ask the VSS Agent to answer through /generate.
The source does not document a complete deployment command or all upload semantics; those are delegated to vss-deploy-profile and vss-manage-video-io-storage.