Healthcare RAG Chatbot

Language models answer health questions fluently and, often, incorrectly. In any domain where a wrong answer carries real consequences, fluency without grounding is a liability. The requirement was an assistant that could hold a natural conversation while remaining anchored to a reviewed knowledge base — and that would decline to answer rather than improvise when the knowledge base had nothing relevant.

Source documents are chunked with overlap, embedded and stored in a vector index. Each user question is embedded and used to retrieve candidate passages, which are re-ranked for relevance before the top passages are passed to the model as the sole permitted context. Multi-turn questions are rewritten into standalone queries first, so follow-ups like "what about for children?" retrieve correctly. When retrieval confidence falls below threshold, the assistant says it doesn't know and recommends speaking to a professional instead of generating an answer.