The problem
Teams answering security questionnaires, contracts and reports need to find exact answers in long PDFs, and they can’t trust AI answers that might be invented.
What I built
A question-answering service: upload a PDF or JSON document plus a list of questions, and get answers with page-level citations. When the document doesn’t support an answer, it returns “Not found” instead of guessing.
How it works
Combines meaning-based (vector) search with exact keyword (BM25) search, so precise terms and acronyms aren’t missed
The model returns passage IDs, and the server checks them before building any citation, so a citation can’t be made up
Answers several questions at once, with timeouts, request-ID logging and token-usage tracking
52 automated tests with mocked LLM calls; Docker-ready, running as a non-root user
Stack: Python, FastAPI, LangChain, Qdrant, OpenAI, React/TypeScript
0 likes · 0 views · Nov 15, 2026
A RAG service in Python that answers questions over PDF and JSON documents with page-level citations, and returns “Not found” instead of making things up.