Nemotron Retrieval Recipes
Plan, debug, tune, evaluate, export, and deploy Nemotron embedding and reranking recipes.
The skill scopes Bash to repository inspection, help, dry-runs, and user-approved execution; requires data-governance confirmation before hosted SDG, protects NVIDIA_API_KEY/NGC_API_KEY, forbids broad environment dumps, and asks before deleting artifacts. Data flows and external API, GPU, Docker, Slurm, and NIM effects are disclosed. Four points are deducted for missing systematic privilege isolation, a general rollback procedure, and more granular command-level confirmation.
It documents stage contracts, config paths, profile invariants, dry-runs, failure localization, and abnormal-input checks, while making the current checkout authoritative. Static files do not prove the commands run in the assessed checkout, and several command, dependency-version, and deployment details require runtime validation. With no execution evidence, the score is capped and reduced to 8/20.
Invocation conditions, embed/rerank routing, failure modes, non-fit cases, inputs, outputs, and stage selection are reasonably clear, including data-governance and credential concerns relevant to Chinese users. Core SDG, model-download, NGC/Docker, hosted-inference, and Hugging Face paths may depend on mainland-China network reachability, credentials, and GPUs, with no China-specific fallback documented; 3 points are deducted.
The skill provides version 0.2.0, author, Apache-2.0 licensing, layered references, examples, limitations, output guidance, and maintenance/evaluation notes; the repository README describes synchronization and update mechanisms. It lacks a clear changelog, explicit maintainer/update owner for this skill, locked dependency guidance, and a fuller FAQ/troubleshooting path. The skill-card output format also does not perfectly match the SKILL.md's conditional format; 4 points are deducted.
The skill can route retrieval failures to the appropriate recipe family and provide stage commands, preview-versus-execution boundaries, metric interpretation, governance checks, and deployment diagnostics. BENCHMARK.md reports 40 tasks and with-skill results. Because the evaluation dataset is unavailable and the results are repository-authored claims not executed in this review, representative directly usable outputs and stable marginal benefit over manual work cannot be confirmed; the static score is 6/15.
The fixed revision, source paths, stage commands, test commands, evaluation fixtures, and benchmark metrics provide useful audit trails. However, the benchmark dataset is unavailable, and there is no independently reproducible CI/test result or corroborating external evidence in the supplied files. Verification is therefore partial at 4/10.
- This is a source-only review; no command, test, dry-run, or deployment was executed.
- The evaluation dataset cited by BENCHMARK.md was not supplied, so its results are limited repository-authored evidence.
- In mainland-China environments, hosted SDG, model downloads, NGC/Docker, Hugging Face, and remote-cluster paths may be unreachable or require additional compliance and network arrangements.
- Before Stage 0, confirm that corpus transfer to the configured inference endpoint is permitted; never expose credentials in commands, logs, or prompts.
What does this skill do, and when should you use it?
Nemotron Retrieval Recipes is for NVIDIA Nemotron’s public embedding and reranking retrieval workflows. It maps retrieval failure modes to embedding or reranker recipes and reconciles commands, configurations, and paths with the current checkout. The skill covers data generation, training, evaluation, export, and deployment while prioritizing previews, configuration checks, and explicit approval for long-running work. It is best suited to engineering teams that already have the relevant repository, data, credentials, and compute resources.
Reads recipe, CLI, configuration, and reference files from the current checkout; uses uv run nemotron embed or uv run nemotron rerank help and dry-run commands to inspect the active interface; selects embedding or reranking based on candidate recall and top-rank ordering problems; recommends dotlist overrides, input paths, artifact roots, and stage ranges; and, with explicit approval, provides local, remote, export, or deployment commands plus validation signals and output paths.
- A retrieval team whose relevant documents are missing from the candidate set and needs an embedding-tuning path.
- A search team that retrieves the right documents but ranks them poorly near the top and needs reranker guidance.
- An engineer who wants to inspect model profiles, stage parameters, and output directories before running a recipe.
- A retrieval team comparing a base model with a tuned model on a fixed evaluation set.
- An operations engineer checking GPU, Docker, container, and credential requirements before export or deployment.
What are this skill's strengths and limitations?
- Covers planning and troubleshooting across Nemotron embedding and reranking recipe stages.
- Treats the current checkout as authoritative, reducing the risk of stale commands, flags, and paths.
- Provides low-cost validation through help commands, dry runs, configuration review, and evaluation comparisons.
- Defines concrete boundaries for data governance, secret handling, GPUs, Docker, and remote execution.
- Does not provide datasets, checkpoints, credentials, GPU capacity, Docker images, or NIM services.
- Depends on the Nemotron recipe repository and its current file layout; missing files are blockers rather than cases for guessed commands.
- Some stages require an NVIDIA API key, CUDA/NVIDIA drivers, Docker, or specific containers.
- The supplied skill text provides no independent test results or platform-coverage evidence.
How do you install this skill?
Install the skill with the default skills CLI: npx skills add nvidia/skills --skill nemotron-retrieval-recipes --yes. Use --agent codex or another supported client option to select an installation target. The README does not document a separate manual-copy requirement for this skill; after installation, the agent uses it when it loads skills and encounters a relevant task.
How do you use this skill?
Invoke it with $nemotron-retrieval-recipes. For example, ask: “Recall is decent, but nDCG is poor and the right passage is around rank 40. Should I tune embed or rerank?” Confirm the recipe checkout, data paths, model profile, stage range, artifact root, execution mode, GPU resources, and configured credentials. Start with help or a dry run such as uv run nemotron rerank run -c default -d --from prep --to eval; do not launch API, GPU, Docker, Slurm, NIM, or other long-running work unless explicitly requested.