Build & Deploy an End-To-End Modular RAG-based Medical Assistant

Medical data requires strict privacy boundaries and grounded accuracy. A single generic chatbot cannot safely serve both a surgeon reviewing clinical reports and a patient checking basic health guidelines.

This project implements a full-stack, role-aware medical assistant that combines Retrieval-Augmented Generation (RAG) with Role-Based Access Control (RBAC).

System Architecture

The architecture enforces permission checks at every stage, from document ingestion to vector retrieval:

• Authentication & RBAC: MongoDB Atlas and Bcrypt manage user persistence and role definitions (Admin, Doctor, Nurse, Patient, and Others).

• Document Ingestion (Admin Only): Administrators upload clinical PDFs, which are chunked, embedded via Google AI Embeddings, and indexed in Pinecone with role-level metadata.

• Role-Filtered Retrieval: When a query is made, Pinecone filters chunks based on the user's role before returning context.

• Low-Latency Generation: Retrieved context is passed to Groq (LLaMA 3) to produce grounded, sub-second responses.

• Frontend: A Streamlit dashboard dynamically renders views according to authenticated privileges.

Core Tech Stack

Key Takeaway

By applying metadata filters during vector search and securing endpoints with FastAPI dependencies, the system prevents unauthorized data access and keeps LLM responses strictly grounded in verified medical records.