‹ All work
Live on Google Cloud Run

MediQA Bot

Medical Q&A that checks its own evidence.

Open live demoGitHub
MediQA Bot home screen: a sidebar with New question and Topics, and the prompt Ask about a health topic with three suggested questions.

The problem

A medical chatbot that answers confidently from weak evidence is worse than no chatbot. I wanted a system that notices when its retrieval is poor, tries again, and shows exactly what it used.

How it works

  1. QuestionA clinician asks a question in the React Progressive Web App.
  2. RetrieveDense PubMedBERT vectors and BM25 sparse vectors are searched together in Qdrant and fused with Reciprocal Rank Fusion.
  3. GradeA LangGraph node judges whether the passages answer the question.
  4. CorrectWeak evidence triggers a query rewrite and a second retrieval.
  5. AnswerThe reply cites its sources; Langfuse and PostgreSQL record the trace and cost.

Decisions

Hybrid retrieval

Dense vectors capture meaning. Sparse BM25 vectors catch exact clinical terms. Fusing both beats either alone.

ONNX embeddings

The embedding model is exported to ONNX and published, so the production image does not need PyTorch.

Cloud Run, not Spaces

Hugging Face Spaces needs a paid plan for Docker apps, so version 2 runs on Google Cloud Run.

Inspectable by design

Langfuse traces, PostgreSQL cost analytics and a retrieval-inspection view show what was fetched and why.

Built with

PythonFastAPIPostgreSQLRAGHybrid retrievalLangGraphLangfuseGCPDockerReact