Medical Q&A RAG Assistant

Business problem

Medical-question assistants need to ground answers in source text, provide citations, and refuse or qualify answers when context is insufficient.

Dataset and context

The app is inference-only on Hugging Face. Heavy indexing and RAGAS evaluation are handled outside the Space. The README includes a medical disclaimer and notes source-document rights considerations.

Methodology

• -Chunked source material with overlap, embedded it, and stored vectors in ChromaDB.
• -Retrieved top-k context passages for each question and generated answers with citations.
• -Added fallback behavior for insufficient context instead of forcing unsupported medical answers.
• -Evaluated 20 questions with RAGAS metrics for answer relevancy, context precision, context recall, and faithfulness.
• -Kept the deployed Space lightweight by separating inference from index-building and evaluation workflows.

Technical approach

The system uses a retrieval-first architecture: user question, vector search, context assembly, answer generation, citations, and safety fallback. The project is framed as an educational decision-support demo, not medical advice or a clinical tool.

Key insights

• -Answer relevancy is 0.9644 across the 20-question RAGAS sample.
• -Context precision is 0.9278, showing retrieved context is usually relevant.
• -Context recall is 0.6067, which points to retrieval coverage as a production-improvement area.
• -Faithfulness is 0.5735, so stronger grounding checks would be needed before any high-stakes use.

Business recommendations

• -Use the demo to show applied AI workflow skill, not as a medical product claim.
• -For production, prioritize source governance, evaluation traces, safety review, and human oversight.
• -Improve retrieval coverage before expanding the answer scope.
• -Track unanswered or fallback questions because they reveal corpus gaps.

Results and impact

This is an educational source-grounded RAG portfolio project. It is not medical advice, not a diagnostic tool, and not intended for clinical deployment.

What I would improve in production

• -Add stricter citation verification and answer-grounding checks.
• -Expand the evaluation set and keep per-question traces for auditability.
• -Add source-document versioning and rights review.
• -Monitor latency, cost, fallback rate, and hallucination-risk examples.

Recruiter takeaway

This project is useful for applied AI roles because it shows retrieval, citations, safety fallback, evaluation, and deployment restraint rather than just a chatbot wrapper.