MediBot

A health-information assistant built on a FAISS vector store and Mistral-7B. The whole design was shaped by one rule: in a medical context, a confident wrong answer is worse than no answer.

The problem

General chatbots will happily answer a health question whether or not they have grounds to. That’s dangerous. I wanted a system that only speaks when it has retrieved, citable sources — and admits it when it doesn’t.

Retrieval with a safety gate

flowchart TD
Q[User question] --> R[Retrieve from<br/>medical corpus]
R --> C{Relevant<br/>sources?}
C -->|yes| G[Generate grounded<br/>answer + citations]
C -->|no| S[Say I don't know<br/>suggest a professional]
G --> D[Add safety<br/>disclaimer]

The confidence check is the part that keeps it honest.

The lesson

It taught me to treat “I don’t know” as a feature, not a failure. Wiring the retrieval confidence directly into whether the model is even allowed to answer did more for trustworthiness than any amount of prompt tuning.

Stack

Python, FAISS, Mistral-7B, a curated medical corpus, source-citation rendering.