The Agent Skills Directory

[SAFE]: The skill consists of architectural documentation and code templates for building RAG systems. It utilizes well-known, trusted libraries and services (OpenAI, Anthropic, Cohere, LangChain).
[CREDENTIALS_UNSAFE]: The provided code snippets demonstrate secure secret management practices by retrieving the PostgreSQL connection string and Cohere API key from environment variables (os.getenv) rather than hardcoding credentials.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and provides mitigations for common RAG failure modes like hallucinations. It includes generation prompt templates with explicit refusal instructions ("If insufficient information: respond 'I don't have enough information to answer that.'") and citation requirements to reduce the risk of the model following instructions embedded in retrieved data.
Ingestion points: The ingest_documents function in SKILL.md loads content from external files using PyPDFLoader and TextLoader.
Boundary markers: The generation prompt uses explicit headers and rules to separate document context from instructions.
Capability inventory: The patterns include local file reading and network operations for embeddings and LLM generation.
Sanitization: The skill relies on prompt-based refusal logic and LLM-based evaluation metrics (RAGAS) rather than programmatic input sanitization.

rag-architecture