The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes the implementation of RAG systems that ingest data from external sources, which introduces a potential surface for indirect prompt injection if the ingested content is malicious.
Ingestion points: Documents are loaded from a local directory (./docs) using LangChain's DirectoryLoader in SKILL.md.
Boundary markers: Prompt templates utilize {context} placeholders to delineate external data. Some templates include instructions like "If you cannot answer based on the context, say 'I don't have enough information'" which serves as a basic instruction-following constraint.
Capability inventory: The skill uses RetrievalQA chains combined with OpenAI LLMs to process and respond to queries based on retrieved document content.
Sanitization: No explicit sanitization, escaping, or validation of the retrieved content is demonstrated in the code snippets before it is interpolated into the prompt.

rag-implementation