haiku.rag
Agentic retrieval-augmented generation that answers questions about your own documents, with citations that quote the source passage, give its page and section, and show it highlighted on the original page. Runs on an embedded database, on local disk or object storage, with no database server to run. Works with any model provider Pydantic AI supports, including locally hosted open models. Built on LanceDB, Pydantic AI and Docling.
When a question needs computation rather than retrieval — counting across documents, comparing sections, aggregating figures — the model writes Python and runs it in a sandbox with the documents mounted as a virtual filesystem and search available from code.
It is evaluated end to end on five public datasets — OpenRAG Bench, T²-RAGBench, HotpotQA, FRAMES and MTRAG — scoring retrieval, answer accuracy and citation quality. Methodology and results are published in full.
