RAG for a medical company: the technical and product challenges
summary
A RAG for the Vidal is built to let doctors query drug monographs and it uses Chainlit with a vector-backed pipeline over Qdrant and Langsmith. Naive 8,000-token chunking and no query reformulation yield many misses on a 100-question test and three-minute lookups; clever HTML-aware chunking plus reformulating to include the INN cuts retrieval errors and gives 88% okay answers and interactive highlighted citations, but latency (~54s) and limited evaluation keep it out of production.
AI-written summary · about
chapters (13)
- 0:00RAG for medical docs intro
- 2:54What is RAG
- 5:44Initial RAG performance
- 10:22Incorrect RAG answers
- 13:15Better chunking for retrieval
- 15:42Prompt engineering with INN
- 17:49Product owner’s prompt role
- 18:32Product risks for doctors
- 20:20Usable doctor interface
- 22:49Time saved measuring
- 24:59Next steps and limits
- 27:13Scaling open source
- 28:36LLM lifecycle conclusion
related talks
44mRAG for a Medical Company: Technical and Product Challenges AI & Data
36mDon't tell me RAG is easy
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
30mI Want to Build a RAG System, Now What?
02h 47mFrom naive to advanced RAG: the complete guide
32mBuilding Performant RAG Applications for Production
05mopen-rag-eval: RAG Evaluation without "golden" answers
11mRAG Evaluation Is Broken! Here's Why (And How to Fix It)