VSS Video Summarizer
Creates timestamped narrative summaries of recorded videos through LVS, with a VLM fallback.
The skill narrowly scopes itself to summarizing one recorded video and explicitly excludes live captioning and report generation. The LVS path requires user confirmation of scenario, events, and optional objects, while deployment also requires confirmation; fallback behavior and backend disclosure are explicit. Nine points are deducted because video URLs are sent to LVS or VLM endpoints without documented handling of sensitive footage, access control, retention, or cross-network data flow. Authentication is only optional, and least-privilege, audit, and rollback controls are not specified.
Routing, readiness checks, warmup retries, model-ID validation, response parsing, and common error cases are documented with concrete commands. Eleven points are deducted because the workflow depends on VIOS, LVS, VLM, jq, curl, and other skills whose availability is not established here; failure reporting and HTTP error parsing remain limited, and static review cannot confirm executability.
The intended use, parameters, output format, exclusions, and LVS/VLM routing conditions are clearly defined, with separate handling for service-operations requests. Five points are deducted because Chinese-language support and mainland-China network reachability are not addressed. Core functionality depends on a reachable local deployment or VLM/LVS service, limiting environment fit.
The documentation has clear layering and includes purpose, prerequisites, limitations, troubleshooting, a reference map, version 3.2.1, an Apache-2.0 declaration, author, upstream link, examples, and evaluation manifests. Five points are deducted because the author format triggered a static validation finding; no explicit changelog, named maintenance responsibility, or stable update process is provided, and referenced skills and files must exist separately for the workflow to be complete.
The stated goal is a timestamped summary of one video, and the LVS response fields and fallback output have directly usable rendering rules. Nine points are deducted because usefulness depends on backend deployment, clip URL reachability, and model quality; no independently verifiable representative outputs are included, and the benchmark covers only one task, providing weak evidence of completeness or advantage over manual or alternative methods.
The skill file, evaluation manifests, benchmark report, and API examples provide auditable primary material, and the report records limited execution results. Six points are deducted because coverage is only one task and one attempt, with no verifiable CI workflow, comprehensive committed test suite, or real response samples; the benchmark claims cannot be independently reproduced from the supplied files, so the static score remains below the five-point ceiling.
- Before processing sensitive or private footage, verify clip-URL access control, transport boundaries, and retention policies for the LVS/VLM backends; the skill does not specify these safeguards.
- Execution depends on several services and companion skills whose availability is not established here; when LVS is unreachable, the workflow falls back to VLM with lower quality and no scenario/event targeting.
- The benchmark covers only one positive task and one attempt, and static validation found an author-format issue; this is insufficient evidence of broad stability.
- Chinese interaction and mainland-China network reachability are unspecified; core functionality may be unavailable when overseas or remote backends cannot be reached.
What does this skill do, and when should you use it?
VSS Video Summarizer is an NVIDIA Video Search and Summarization Blueprint skill for producing a polished summary of one recorded video clip. It prefers the HITL-gated LVS summarization microservice and falls back to direct VLM inference when LVS is unavailable. The LVS path can target a scenario and events, while the fallback uses one fixed general-purpose prompt. It is not intended for live RTSP captioning, report generation, or semantic archive search.
Obtains a temporary MP4 URL and timeline from VIOS; probes the VLM /v1/models and LVS /v1/ready endpoints with curl; collects scenario and events before calling POST /v1/summarize when LVS is ready; otherwise calls /v1/chat/completions with the fallback prompt; parses the backend response and renders a timestamped summary with the video name and duration.
- A video analyst needs a timestamped narrative summary of a recorded surveillance clip.
- An operations user wants one MP4 clip summarized for a specified scenario and set of events.
- A team running the VSS lvs profile wants targeted summaries rather than a generic VLM description.
- A user with an accessible RTVI VLM but no deployed LVS service needs a usable general video description.
What are this skill's strengths and limitations?
- Supports both a primary LVS workflow and a direct VLM fallback.
- The LVS path supports scenario, event, and optional object targeting.
- Documents readiness checks, response parsing, error handling, and presentation rules.
- Includes guidance for service warmup, empty responses, and Cosmos <think> blocks.
- Requires an accessible LVS or VLM/RT-VLM backend and a video URL reachable by that backend.
- The VLM fallback uses a fixed generic prompt, has lower quality, and cannot target scenarios or events.
- Makes one backend call per request with no parallel hedging or multi-pass summarization.
- Explicitly excludes live RTSP captioning, report generation, and archive semantic search.
How do you install this skill?
Install the skill from the NVIDIA catalog with: npx skills add nvidia/skills --skill vss-summarize-video --yes. The CLI determines the installation destination; the source does not specify a fixed folder path.
How do you use this skill?
After installation, ask a compatible Agent something like “Summarize this video.” and provide a video managed by VIOS. The Agent must obtain a temporary MP4 URL, check LVS and VLM availability, then follow the appropriate workflow. The LVS route requires scenario and event collection; if LVS is unavailable, the Agent uses the fixed VLM fallback prompt.
How does this skill compare with similar options?
Compared with vss-deploy-dense-captioning in the same VSS collection, this skill summarizes recorded video rather than captioning live RTSP streams. Compared with vss-generate-video-report, it produces a single-video summary rather than an incident or alert-window report. Compared with vss-search-archive, it processes a supplied clip rather than performing semantic search across an archive.