The Agent Skills Directory

[COMMAND_EXECUTION]: The references/install.md file contains instructions for system-level package installation using sudo apt-get install for dependencies such as libreoffice and ffmpeg. These tools are necessary for the skill's primary function of processing Office and multimedia files. Additionally, scripts/filename_fast_path.py uses subprocess.run to invoke the retriever binary, which is a functional requirement for localized PDF processing within the venv.\n- [EXTERNAL_DOWNLOADS]: The skill's installation workflow in references/install.md involves cloning the official NVIDIA NeMo-Retriever repository from GitHub and downloading Torch wheels from the official PyTorch registry. These sources are considered trusted and well-known for the technical requirements of the skill.\n- [PROMPT_INJECTION]: The skill processes user-provided multimodal documents (PDFs, images, etc.) and interpolates the extracted text into the agent's context, presenting an indirect prompt injection surface.\n
Ingestion points: The retriever ingest CLI command processes directories containing user documents (SKILL.md, references/cli/ingest.md).\n
Boundary markers: The instructions mandate citations and passage inspection but do not explicitly require the agent to use robust delimiters (e.g., XML tags) to isolate document text from instructions.\n
Capability inventory: The agent utilizes the retriever CLI for searching and a targeted text-extraction tool for retrieving page elements.\n
Sanitization: There is no mention of explicit sanitization or filtering logic for the content extracted from processed documents before it is added to the AI's prompt.

nemo-retriever