What does this skill do, and when should you use it?
rag-blueprint is an operations-focused Agent Skill for NVIDIA RAG Blueprint. It covers deployment, configuration, feature management, troubleshooting, shutdown, and teardown across Docker Compose, Kubernetes/Helm, and library deployments. The skill routes requests to the relevant playbook and checks services, configuration, GPUs, and health state before and after changes. It is best suited to teams already working with the NVIDIA RAG Blueprint repository and its supported runtime environments.
Routes requests to deployment, configuration, troubleshooting, or shutdown workflows; checks GPU, Docker, Kubernetes, Helm, processes, ports, disk, and system information; reads repository documentation, configuration files, and referenced playbooks; manages Agentic RAG, VLM, guardrails, query rewriting, ingestion, search, models, reasoning, summarization, observability, and related features; applies Docker Compose, Helm, or Python configuration changes; and verifies service status and health endpoints after changes.
- A platform engineer needs to deploy and start NVIDIA RAG Blueprint with Docker Compose.
- A Kubernetes operator needs to enable VLM, guardrails, query rewriting, or observability.
- An operations engineer needs to investigate unhealthy RAG services or failed restarts.
- A RAG developer needs to change models, vector databases, ingestion, retrieval, or generation settings.
- An administrator needs to stop or tear down a deployment while confirming destructive cleanup.
How do you install this skill?
- Do not perform shutdown, teardown, data deletion, or restart operations without explicit confirmation, current-state checks, and a recovery plan.
- NGC_API_KEY and APP_LLM_APIKEY may be indirectly exposed through environment, process, or container inspection; never print complete secrets and use least privilege.
- NVIDIA-hosted mode depends on integrate.api.nvidia.com, whose reachability should be separately validated from mainland-China networks; self-hosted mode still requires NVIDIA GPUs, images, and NGC access.
- Before publication, complete and validate the referenced deploy, troubleshoot, and shutdown playbooks, address the benchmarked over-triggering and duplication findings, and rerun evaluation.
- Shell / CLI
- Network access
- Local filesystem
NVIDIA RAG Blueprint repository checkoutDocker/ComposeKubernetes/HelmPython 3.11+NVIDIA GPU tooling
Install the named skill with the repository README's supported skills CLI command:
npx skills add nvidia/skills --skill rag-blueprint --yes
The CLI prompts for the installation destination. Cloning the repository or manually copying the skill directory is not required.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/rag-blueprint" ~/.claude/skills/
rm -rf "$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:
- RAG is unhealthy
After the skill is loaded by the agent, use a concrete request such as:
“Deploy RAG”
“Enable VLM”
“RAG is unhealthy”
“Stop RAG”
For configuration requests, the skill first expects a running deployment, then selects and follows the relevant reference file.
What are this skill's strengths and limitations?
- Supports Docker Compose, Kubernetes/Helm, and Python library deployment modes.
- Defines intent routing, environment detection, configuration cross-checks, and post-change verification.
- Covers deployment, feature configuration, troubleshooting, shutdown, and teardown across many RAG subsystems.
- Includes source-repository metadata, API schema paths, tags, languages, and framework information.
- It is explicitly scoped to the NVIDIA RAG Blueprint repository and is not general guidance for arbitrary RAG systems.
- Operational changes depend on a working Docker, Kubernetes/Helm, or Python environment.
- Self-hosted NIM services require NVIDIA GPU tooling, and secrets such as NGC_API_KEY must come from the user's environment.
- The supplied source does not provide an independent test suite or performance results for this skill.
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 |
|---|---|---|---|---|
| NVIDIA RAG Blueprint Operations this page ✓ NVIDIA · Official | 39 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| NVIDIA AI-Q Deployment ✓ NVIDIA · Official | 57 · 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 Multi-Camera 3D Detection and Tracking ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO NVIDIA GPU Host Setup ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
The skill supports Docker Compose, Kubernetes/Helm, and Python library deployment modes, but the supplied source does not provide a direct comparison with other RAG operations skills.
How did FollowSkills review this skill?
The skill scopes itself to the specified NVIDIA RAG Blueprint repository and requires confirmation or user-supplied input for deletion, API keys, and platform choices. It permits inspection of processes, filesystem metadata, container state, and network state, and may use an external NVIDIA API, but it does not adequately disclose sensitive data flows, credential-exposure risks, rollback procedures, or all external effects, so points are deducted.
Intent routing, configuration flow, health checks, and hardware restrictions are reasonably specific for the happy path. However, key playbooks and many referenced files are absent from the supplied evidence; the allowed tools do not include actual editing, restart, or deployment commands even though the instructions direct the agent to perform those changes; abnormal-input handling and failure feedback are limited. The score therefore stays within the static-review ceiling.
Use cases, deployment modes, and feature keywords are relatively clear. However, the description covers essentially any RAG action; the benchmark records that it is overly long, omits WHEN-to-use guidance, and lacks negative triggers, creating over-triggering risk. Boundaries for non-Docker/Helm/library environments, missing repositories, Chinese interaction, and mainland-China reachability are under-specified. NVIDIA-hosted mode depends on integrate.api.nvidia.com, limiting environmental fit.
SKILL.md provides a version, purpose, prerequisites, routing table, progressive configuration sections, examples, limitations, and hardware notes, and declares Apache-2.0. Repository materials provide maintenance, mirroring, and CI signals. However, license metadata is NOASSERTION, the repository describes combined Apache-2.0 and CC-BY-4.0 licensing, skill-specific maintenance responsibility is not explicit, and there is no skill-specific changelog, FAQ, or complete installation guide. The benchmark also reports duplicate content and unexpected files, so points are deducted.
The skill covers deployment, configuration, troubleshooting, shutdown, and API management, with routing tables and verification commands that could help complete operational tasks. Static evidence does not show directly usable representative outputs; BENCHMARK.md explicitly reports an overall FAIL and no Tier 3 live results. Because key playbooks are not supplied, core coverage and comparative benefit cannot be verified beyond a plausible happy path.
The evidence includes a pinned repository revision, OpenAPI schema paths, a benchmark file, evaluation scenarios, and real CI workflow references, making some claims traceable. However, the benchmark lacks Tier 3 results and actual execution signals, and no committed test suite covering the skill's key paths is shown. Conclusions therefore remain primarily static and do not justify a higher score.
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 →