CureX Healthcare RAG Assistant

CureX is an agentic healthcare RAG application for asking evidence-grounded questions over medical documents. It combines document ingestion, local pgvector retrieval, cited responses, safety checks, JWT authentication, user-scoped memory, human review, and an evaluation dashboard.

The assistant is for educational and research workflows only. It is not a substitute for professional medical advice, diagnosis, or treatment.

Core Features

• Healthcare-focused chat assistant with streamed responses
• PDF, TXT, and Markdown document upload and indexing
• Local PostgreSQL + pgvector vector storage
• Gemini-based generation and embeddings
• User login with JWT authentication
• User-scoped chat sessions and memory isolation
• Source citations for retrieved document chunks
• Medical safety guardrails and faithfulness scoring
• Optional visible agent workflow toggle
• Human-in-the-loop answer review: helpful, unsupported, unsafe, needs review

Architecture

!CureX architecture diagram

Frontend React app
        |
        v
FastAPI backend
        |
        +-- Auth and user-scoped sessions
        +-- Document upload and ingestion
        +-- Chunking and metadata enrichment
        +-- pgvector retrieval
        +-- Agent routing, safety checks, and response generation
        +-- Feedback and evaluation APIs

Tech Stack

• Backend: FastAPI, LangGraph, LangChain, Pydantic
• Frontend: React, Vite, lucide-react
• Vector DB: PostgreSQL with pgvector
• Models: Gemini generation and Gemini embeddings
• Storage: SQLite for local auth, memory, feedback, safety logs, and query logs
• Evaluation: RAGAS-style metrics and faithfulness checks
• Tooling: uv, pytest, Docker Compose

Setup

Create the Python environment with uv:

python -m uv venv --python 3.11
.\.venv\Scripts\Activate.ps1
python -m uv sync

Install frontend dependencies:

cd frontend
npm install
cd ..

Create a local .env from the example:

Copy-Item .env.example .env

Set at least:

GOOGLE_API_KEY=your_google_api_key
GEMINI_API_KEY=your_google_api_key
POSTGRES_URL=postgresql://healthcare:healthcare@localhost:5432/healthcare_rag
VECTOR_BACKEND=pgvector

Start local infrastructure:

docker compose up -d

Run

Start the backend:

.\.venv\Scripts\python.exe -m uvicorn backend.api.main:app --reload

Start the frontend:

cd frontend
npm run dev

Open the Vite URL shown in the terminal, usually:

http://localhost:5173

Usage Flow

1. Create an account or sign in. 2. Start a new healthcare conversation. 3. Upload PDF, TXT, or Markdown medical documents. 4. Ask questions about the uploaded content or clinical research topics. 5. Inspect citations and optional agent workflow details. 6. Review answers as helpful, unsupported, unsafe, or needs review. 7. Open the evaluation dashboard to monitor quality signals.

API Highlights

• POST /auth/register
• POST /auth/login
• GET /auth/me
• POST /session/new
• GET /session/{session_id}/history
• POST /documents/upload
• POST /chat
• POST /feedback
• GET /feedback/summary
• GET /eval/metrics

Testing

Run the backend test suite:

.\.venv\Scripts\python.exe -m pytest tests -q

Build the frontend:

cd frontend
npm run build

MCP Tool Server

Run the local healthcare tool server:

.\.venv\Scripts\healthcare-mcp-server.exe

Available tools include drug interaction lookup, ICD-10 lookup, BMI calculation, and retrieval search over indexed healthcare chunks.