The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the NeMo Retriever source code from NVIDIA's official GitHub repository and installs dependencies from the PyTorch wheel registry and PyPI registry.
[COMMAND_EXECUTION]: Executes shell commands to install system dependencies (such as ffmpeg and libreoffice via sudo), index document corpora, and query the local vector database. Utility scripts in the skill use subprocesses to invoke the retriever CLI for targeted extraction.
[PROMPT_INJECTION]: Processes untrusted external content (PDFs, images, and other media files) for indexing and retrieval. *Ingestion points: Processes files in user-specified directories via the 'retriever ingest' command. *Boundary markers: Instructions emphasize the use of citations and verbatim quoting, though no explicit delimiters for untrusted content are defined in the scripts. *Capability inventory: Utilizes Bash for indexing/querying and local file writes; library installation involves network access. *Sanitization: Content is processed through the NeMo Retriever pipeline; no additional sanitization of extracted text is implemented in the local skill scripts.

nemo-retriever