The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill frequently references the requirement of an NVIDIA_API_KEY for RAGAS evaluation but provides comprehensive guidance on credential hygiene. It explicitly instructs users to avoid hardcoding keys, suggest using environment variables or secret managers, and warns against leaking keys into shell history or version control.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided datasets consisting of a corpus/ directory and a train.json file. This represents a known attack surface for indirect prompt injection, as malicious content within the dataset could target the RAG system under evaluation. However, this capability is the primary functional purpose of the tool, and the skill includes validation steps for the input data. Ingestion points: Reads dataset files from the filesystem. Boundary markers: Provides Python-based validation checks to ensure train.json matches a specific JSON array schema. Capability inventory: The driver script executes network requests to RAG server/ingestor endpoints and writes result artifacts to the filesystem. Sanitization: Relies on documented schema constraints and standard JSON parsing.
[METADATA_POISONING]: A minor documentation inconsistency exists where the YAML frontmatter specifies version 2.0 while the skill-card metadata refers to version 2.6.0. This is categorized as metadata poisoning but is assessed as a benign documentation error rather than a deceptive practice.

rag-eval