nemotron-retrieval-recipes

Nemotron Retrieval Recipes is an operational assistant skill for working with NVIDIA's public Nemotron `embed` and `rerank` retrieval recipes inside a source checkout or installed package. It exists to route a user's retrieval failure to the correct recipe family, reconcile documentation against the actively-changing checkout (CLI, configs, containers, and output paths), avoid accidental long-running launches, protect secrets, and return concrete preview, dry-run, and execution commands. The skill treats the current checkout as authoritative over its own bundled references and memory, and reports a blocker rather than guessing when a required recipe directory, CLI command, config, or env profile is missing.

The skill maps the full pipeline across stages: Stage 0 synthetic data generation (SDG), Stage 1 prep (hard-negative mining and BEIR eval data), Stage 2 contrastive fine-tuning, Stage 3 evaluation, Stage 4 export (ONNX/TensorRT for the Llama profile), and Stage 5 deployment (vLLM or NIM backends). It provides two model profiles (`default` for Nemotron-3-Embed and `llama` for llama-nemotron-embed), references for embedding vs. reranking selection, remote/batch execution via `env.toml` profiles, and evaluation tooling. It emphasizes dry-run previews and dotlist config overrides rather than editing recipe defaults.

Target users are ML engineers and applied researchers tuning first-stage (bi-encoder) or second-stage (cross-encoder) retrieval models who diagnose issues like low Recall@k or poor top-rank ordering. It is best suited to teams running Nemotron retrieval fine-tuning end to end on GPU infrastructure, and explicitly scopes out generic retrieval theory, vector-database selection, and non-recipe Docker/Slurm/NIM troubleshooting.