RAG + MCP

MCP can expose a retrieval system as reusable tools or resources. The vector database and indexing pipeline stay behind the server; clients see a stable search interface.

Example tool surface

• search_docs(query, filters)
• get_document(document_id)
• list_collections()

Preserve provenance

Return source IDs, titles, URLs, snippets, timestamps, and relevance metadata. The agent should be able to cite and inspect the evidence it uses.

Authorize retrieval

Vector search can leak data just as easily as SQL. Apply access filters before retrieval, not after results reach the model.