The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the ragas library from PyPI or directly from its official GitHub repository (https://github.com/explodinggradients/ragas). These are legitimate and standard sources for this evaluation framework.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to evaluate untrusted data (user inputs, assistant responses, and retrieved contexts) using LLM-based metrics. This introduces a surface for indirect prompt injection, as malicious content within the evaluated data could attempt to influence the evaluation result.
Ingestion points: Data is ingested via Hugging Face Dataset or pandas.DataFrame objects, often sourced from log files (e.g., logs/answers.jsonl, data/sql_cases.jsonl).
Boundary markers: The prompt templates for custom metrics (e.g., DiscreteMetric in SKILL.md) interpolate data directly without explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The skill uses ragas.evaluate and AsyncOpenAI to perform evaluations, which involve network calls to LLM providers.
Sanitization: No explicit sanitization or filtering of external content is described in the provided examples.

ragas-evaluation