The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes the 'uv' package manager to synchronize project dependencies and install the benchmark driver in an editable mode. These operations are directed at the skill's own scripts and local repository structure. Given that the skill is authored by a trusted organization (NVIDIA), these downloads are considered standard environment setup and do not involve untrusted remote sources.
[COMMAND_EXECUTION]: The benchmarking process requires the execution of shell commands, including 'uv run' and 'curl', to orchestrate load tests and verify the status of RAG server collections. These tools are used for their intended primary purpose of performance measurement and service interaction within a developer-controlled environment.
[PROMPT_INJECTION]: The skill includes a synthetic query generation pipeline that processes responses from external LLM endpoints (via 'synthetic.llm_url'). This creates a surface for indirect prompt injection where malicious or malformed data from the LLM could influence the benchmark queries.
Ingestion points: The 'SyntheticQueryGenerator' in 'scripts/rag-perf/rag_perf/query.py' ingests content from external chat-completion endpoints.
Boundary markers: The skill uses strictly defined prompt templates in 'scripts/rag-perf/prompts/default_prompts.yaml' to enforce output discipline and minimize the risk of including unwanted text or reasoning leakage in the queries.
Capability inventory: The agent possesses capabilities to execute shell commands, perform network requests, and write results to the local filesystem.
Sanitization: The generator applies basic string stripping to remove numerical prefixes or numbering from the LLM output, though it primarily relies on prompt engineering to ensure query cleanliness.

rag-perf