The characteristic failure of a retrieval system is answering from adjacent material and sounding identical to answering from the right document. Being able to enumerate what was indexed lets a caller check whether the question is even in scope before trusting the answer. The surface is deliberately two tools because the interesting part is the SDK behind it — this is what your knowledge base looks like once built, and the building is where the work is.
The server side of an SDK for building knowledge bases from arbitrary sources: the tools query the base in natural language and list what has been indexed into it.
- query asks the knowledge base a natural-language question rather than requiring a search syntax.
- list_sources returns the documents and URLs indexed into it, so what the base can actually answer from is visible rather than assumed.
OPENAI_API_KEY — retrieval runs against OpenAI, so queries bill to that account.
One command plus a key — npm install akyn-ai, then supply credentials