Project 1: Appointment Prep Chatbot¶
Mentors: To be announced
Problem: Patients ask the same handful of pre-visit questions (what to bring, fasting rules, parking) over and over, and an ungrounded chatbot risks confidently making up an answer the clinic never gave.
Context: Grounded in a short FAQ doc you write yourself (5-10 Q&As; the point is demonstrating RAG grounding, not sourcing external data), using an open-source LLM with retrieval-augmented generation (RAG) over Chroma as the local vector store. No training.
Goals: Answer "what do I need to bring / can I eat before my visit" style questions grounded only in the clinic's own FAQ doc, and say "I don't know, call the office" when the answer isn't in it.
Deliverables: A chat widget (message list + input box) on a mock clinic page, backed by (1) an ingestion script that chunks the FAQ doc and embeds it into Chroma with an open sentence-transformers model, and (2) a chat endpoint that embeds the user question, retrieves top-k chunks, and stuffs them into the LLM prompt as context, tested to confirm it refuses questions outside the FAQ's scope.
Showcase: TBD
References:
- Development and Evaluation of a Retrieval-Augmented Generation Chatbot for Orthopedic and Trauma Surgery Patient Education (JMIR AI, 2025): a RAG chatbot built and scored on the same "does it stay grounded in the source doc" question this project's refusal test checks.
- Systematic Analysis of Retrieval-Augmented Generation-Based LLMs for Medical Chatbot Applications (open access): surveys the RAG-for-medical-chatbot design space this project implements a minimal version of.