What does this skill do, and when should you use it?
This skill runs top-level fusion search in an NVIDIA Video Search and Summarization deployment. It searches archived video, ingests video files or RTSP streams through the VSS agent backend, and deletes ingested sources. Cosmos Embed1 produces video-segment embeddings, which are matched through Elasticsearch. It requires a reachable VSS search profile, credentials, and command-line dependencies before use.
Checks whether the VSS agent and Elasticsearch are reachable; lists and verifies video sources in VIOS; uploads video files or registers RTSP streams through the VSS agent backend; starts RTVI-CV and RTVI-Embed processing and stores embeddings in Elasticsearch; sends natural-language queries to /generate; returns timestamped, relevance-ranked video segments with playback links; and deletes video files or RTSP sources through the agent backend.
- Security or operations teams searching warehouse archives for forklifts, people, or vehicles with natural-language queries.
- Investigators locating events such as a person entering a restricted area or a truck arriving at a gate within a time range.
- Video-platform administrators ingesting new files into the search system and waiting for embeddings before querying them.
- Camera operators registering RTSP streams so subsequent segments can enter the search index.
- System administrators removing a source while cleaning up both VIOS storage and its search embeddings.
How do you install this skill?
- Add explicit user confirmation, target-UUID validation, and a recoverable backup or rollback strategy before deletion.
- Do not expose NGC/API credentials in logs, command output, or errors; add requirements for video privacy, access control, and retention.
- Deployment and search depend on VSS, Docker, GPUs, Elasticsearch, NGC, and remote services; reachability should be verified on mainland-China or restricted networks.
- Reconcile the Apache-2.0 versus “Apache 2.0 OR MIT” metadata and add continuous tests, change history, and clear maintenance ownership.
- Shell / CLI
- Network access
- Local filesystem
curljqDockerVSS search profileNGC credentials
Install the NVIDIA skill collection with: npx skills add nvidia/skills. Install only this skill with: npx skills add nvidia/skills --skill vss-search-archive --yes. The README does not document additional manual installation steps.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/vss-search-archive" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- find all instances of forklifts
- show me people near the loading dock
- {"input_message":"find all instances of forklifts"}
Provide HOST_IP explicitly to the agent first. Ensure the VSS search profile is running and that sources have been ingested through the VSS agent backend. Then trigger requests such as “find all instances of forklifts” or “show me people near the loading dock”. REST example: curl -s -X POST http://${HOST_IP}:8000/generate -H "Content-Type: application/json" -d '{"input_message":"find all instances of forklifts"}' | jq . If a query names a source, verify it in VIOS first. If HOST_IP is missing, a source is absent, or results are unexpected, the workflow requires stopping and handling the issue before continuing.
What are this skill's strengths and limitations?
- Supports both archived video and RTSP stream sources.
- Provides agent-backed ingestion, stream registration, search, and deletion workflows.
- Supports attribute-only, embedding-only, and fused action-plus-attribute search behavior.
- Includes deployment checks, source verification, and troubleshooting guidance.
- The documentation labels video search an Alpha Feature and does not recommend it for production use.
- Requires a reachable VSS search profile, Elasticsearch, Docker, curl, jq, and NGC credentials.
- GPU memory, concurrency, and storage capacity depend on host hardware and the deployment profile.
- NGC-hosted models and NIMs may involve rate limits, GPU-memory requirements, and license restrictions.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| VSS Archive Video Search this page ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| VSS Video Auto-Calibration ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| VSS Video Analysis Reports ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| VSS Video Embedding Deployment ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| VSS Standalone Video Analytics API ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
Its scope differs from vss-ask-video, which is named for ad-hoc visual Q&A; this skill targets archive search. It does not provide live captioning, which belongs to vss-deploy-dense-captioning, or video summarization and reports, which belong to vss-summarize-video.
How did FollowSkills review this skill?
The skill requires an explicit HOST_IP and asks for confirmation before deployment. It documents data flow across video, RTSP, VIOS, RTVI, Elasticsearch, and NGC credentials, and avoids printing tokens. Deductions apply because search, ingestion, and deletion create external side effects; deletion lacks explicit second-party confirmation and a recovery path; video privacy and sensitive-data handling are unspecified; and least-privilege boundaries for credentials and remote services are incomplete.
The instructions provide concrete workflows, probes, source checks, some rollback language, and diagnostic feedback for common failures. Deductions apply because static review cannot reproduce the key paths, test coverage is very limited, and the benchmark contains only one evaluation task; service dependencies, API compatibility, and cross-component failure semantics still require runtime validation.
Use cases, source handling, search modes, and non-fit cases for Q&A, captioning, and summarization are clearly stated, with a hard source-existence check before searching. Deductions apply because there is no Chinese-language behavior or output example, and boundaries for network reachability, deployment geography, and non-NVIDIA environments are limited; core functionality depends on reachable VSS, NGC, and related services.
The documentation includes frontmatter, prerequisites, workflows, examples, limitations, troubleshooting, references, version 3.2.0, author attribution, and an Apache-2.0 declaration. Deductions apply because there is no clear changelog, named maintenance owner, or update procedure; skill-card metadata says “Apache 2.0 OR MIT,” conflicting with SKILL.md; and the benchmark records an author-format finding.
The skill covers natural-language fusion search, file/RTSP ingestion, cleanup, structured reporting, result verification, and optional screenshot inspection, giving it practical workflow value. Deductions apply because static review cannot confirm result correctness, the benchmark has only one task, and reported effectiveness is 62% for Codex and 68% for Claude; usable results still depend on deployment, indexing, and human visual review.
The selected path includes evaluation manifests, search scenarios, benchmark metrics, and troubleshooting commands, making some claims auditable. Deductions apply because there is no multi-source independent corroboration or continuous-integration coverage, the evaluation sample has one task, and static reading cannot establish actual runtime results; therefore the score remains at or below the static cap of 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →