MedQuery grounds every answer in a curated set of medical reference documents, reducing hallucination risk compared to a plain language-model chat.
Documents are chunked, embedded, and retrieved at query time, then passed to Llama 2 as context alongside the user's question.
Key contributions
- Built a document ingestion and chunking pipeline for medical references.
- Implemented vector retrieval with FAISS for low-latency lookups.
- Tuned prompts to keep answers grounded and cite source passages.
- Exposed the assistant through a simple FastAPI service.
The challenge
Long medical documents produced retrieval chunks that lost important context when split naively.
The solution
Switched to overlapping, section-aware chunking that kept headings attached to their content, which noticeably improved retrieval relevance.