Summary
A private set of documents becomes answerable.
Everything the loop needs is in one server: put documents in, search them, and ask questions that get answered from what came back. It is small enough to point at a handful of documents for a single project, and clearing the store is one call when that project ends.
What it is
An MCP server that runs a retrieval-augmented generation loop: documents are chunked and vectorised into a store, retrieved by cosine similarity, and answered against. The model work runs on an OpenAI key you supply.
What you get
- Documents added with metadata, chunked and vectorised automatically
- Similarity search over the store, with the number of results you want
- Questions answered against the retrieved context rather than from the model alone
- Statistics on what the knowledge base currently holds
- The whole store cleared when you want to start again
Requirements
Node.js, a clone of the repository, and an OpenAI API key in the environment. The RAG agent is initialised once before the other tools will do anything.