The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installation process described in references/install.md downloads the NeMo Retriever Library from the official NVIDIA GitHub repository and fetches PyTorch wheels from the official PyTorch wheel index.- [COMMAND_EXECUTION]: The scripts/filename_fast_path.py script utilizes subprocess.run to execute the local retriever command-line tool for PDF processing, while SKILL.md and reference guides provide instructions for running retriever query and retriever ingest commands.- [PRIVILEGE_ESCALATION]: Setup and installation instructions in references/install.md and references/setup.md include the use of sudo for installing host dependencies like libreoffice and ffmpeg.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted multimodal document collections (PDF, DOCX, images, audio, video) to support question-answering, which presents an indirect prompt injection surface where malicious instructions within the documents could attempt to manipulate the agent's synthesized responses.
Ingestion points: Document data enters the context via the retriever ingest and retriever query commands as outlined in SKILL.md and references/setup.md.
Boundary markers: The instructions in references/query.md lack specific delimiters or instructions for the agent to disregard instructions embedded within retrieved text chunks.
Capability inventory: The skill environment allows for shell command execution, Python script execution, and local file access.
Sanitization: There are no documented content sanitization or validation steps for the text retrieved from external documents.

nemo-retriever