The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of systems that ingest untrusted data from external sources and interpolate it into LLM prompts. This architecture is susceptible to indirect prompt injection if the source documents contain malicious instructions designed to subvert the agent's behavior.
Ingestion points: The skill uses DirectoryLoader to read all text files from the ./docs directory into the application context.
Boundary markers: The prompt templates use standard delimiters (e.g., Context:\n{context}) to separate data from instructions, which provides basic structural separation but does not prevent adversarial injection.
Capability inventory: The skill utilizes the OpenAI LLM to process retrieved context and generate responses, including citations and research summaries.
Sanitization: The provided implementation does not include logic for sanitizing, filtering, or validating the content of retrieved documents before they are passed to the LLM.

rag-implementation