The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install the NeMo Retriever Library from NVIDIA's official GitHub repository. It also retrieves pre-trained model weights from HuggingFace and package dependencies from official registries like PyTorch and PyPI.
[COMMAND_EXECUTION]: The skill operates primarily through Bash commands, invoking the retriever CLI for document processing and vector search. It utilizes helper Python scripts (filename_fast_path.py and grep_corpus.py) that safely manage subprocesses to handle file indexing and search operations.
[PRIVILEGE_ESCALATION]: The installation documentation (references/install.md) provides commands for installing system-level dependencies such as libreoffice and ffmpeg using sudo apt-get. This is a standard administrative requirement for enabling multimedia and Office document support.
[DYNAMIC_EXECUTION]: Instructions for querying and troubleshooting involve the use of Python one-liners (python -c "...") to parse and format JSON output from the retriever tools. This facilitates data handling within the CLI workflow.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process external data (PDFs, images, Office files, etc.). This introduces a potential surface for indirect prompt injection if those files contain adversarial content.
Ingestion points: User-provided document directories and media files (e.g., ./pdfs/, ./images/, ./media/).
Boundary markers: The instructions provide guidelines for synthesizing answers with citations but do not implement explicit boundary markers or delimiters for the retrieved text chunks themselves.
Capability inventory: The skill uses Bash, Write, and Read tools, allowing for extensive interaction with the local environment.
Sanitization: The skill does not explicitly describe automated sanitization or filtering of the retrieved content before it is processed by the agent.

nemo-retriever