Chunking · hybrid retrieval · verifiable citations
Document Q&A fails in one specific way: confident answers you cannot source. It must answer and let you click back to the original text.
Documents are chunked semantically with page and paragraph numbers kept. Retrieval is keyword plus vector, recall first then rerank; generation must carry citation numbers, and the source snippet is listed under the answer.
Chunks overlap to avoid cut sentences; with no match it says not found instead of guessing; the index updates incrementally rather than rebuilding.
Download source ↓ (free with an account)