The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external text corpora for synthetic data generation and training, which represents a potential attack surface.
Ingestion points: Processes user-provided text files via corpus_dir or external datasets from Hugging Face (hf://...) as described in references/embed.md and references/rerank.md.
Boundary markers: The instructions require the agent to ground actions in the current repository checkout and explicitly confirm data-governance policies before sending content to inference endpoints.
Capability inventory: Employs the Bash tool for repository inspection, dry-runs, and execution of the nemotron CLI.
Sanitization: Explicitly instructs the agent to verify that the user's data-governance policy permits sending corpus content to configured inference endpoints and to prefer private or air-gapped paths for sensitive data.
[DYNAMIC_EXECUTION]: The skill configuration specifies the use of trust_remote_code=true when loading specific model checkpoints.
Evidence: references/embed.md states, "Both profiles use approved NVIDIA checkpoint code with trust_remote_code=true."
Context: This setting is used to load official NVIDIA model architectures from trusted Hugging Face repositories (nvidia/Nemotron-3-Embed-1B-BF16 and nvidia/llama-nemotron-embed-1b-v2).
[EXTERNAL_DOWNLOADS]: The skill references and downloads datasets and container images from external registries.
Evidence: The skill fetches datasets from Hugging Face (hf://nvidia/...) and pulls Docker images from the NVIDIA GPU Cloud registry (nvcr.io/nim/nvidia/...).
Context: These downloads are targeted at official vendor-managed repositories and registries, conforming to standard operational practices for machine learning deployment.

nemotron-retrieval-recipes