Medical Chatbot (FastAPI + React + RAG)

Medical Chatbot is a full-stack app that lets users upload PDF documents, build a FAISS vector index, and ask medical questions with Groq-powered responses.

Stack

• Backend: FastAPI, LangChain, FAISS
• Frontend: React + Vite
• Embeddings: FastEmbed (BAAI/bge-small-en-v1.5)
• LLM: Groq API
• Deployment: Railway (single Docker service)

Project Structure

• backend/ FastAPI app, ingestion, RAG chain
• frontend/ React app
• Dockerfile builds frontend and serves with backend
• docker-compose.yml local container orchestration

Local Run

1. Create backend environment file at backend/.env.
2. Install backend dependencies:

pip install -r backend/requirements.txt

3. Install frontend dependencies:

cd frontend
npm install
npm run build
cd ..

4. Start API:

uvicorn backend.main:app --host 0.0.0.0 --port 8000

5. Open:

• http://localhost:8000/
• Health: http://localhost:8000/api/health

Docker

docker build -t medical-chatbot .
docker run --env-file backend/.env -p 8000:8000 medical-chatbot

Railway Deployment

1. Login and initialize:

npx @railway/cli login
npx @railway/cli init --name Medical_chatbot

2. Create/link service and deploy:

npx @railway/cli service link backend
npx @railway/cli up --service backend

3. Set required variables:

• GROQ_API_KEYS (required, comma-separated keys supported)
• GROQ_MODEL (default: llama-3.1-8b-instant)
• DB_FAISS_BASE (recommended with volume: /data/vectorstore)
• EMBED_MODEL (default: BAAI/bge-small-en-v1.5)
• EMBED_BATCH_SIZE (recommended: 4 on small Railway plans)
• RAG_MAX_PDF_PAGES (recommended: 80 on small Railway plans)

4. Add persistent volume for vectorstore:

npx @railway/cli volume add

Mount path:

• /data

Then set:

npx @railway/cli variable set --service backend --environment production DB_FAISS_BASE=/data/vectorstore

Redeploy:

npx @railway/cli up --service backend

API Endpoints

• GET /api/health
• POST /api/upload
• GET /api/upload/status/{job_id}
• POST /api/ask
• POST /api/ask/stream
• GET /api/source/{doc_id}

Troubleshooting

• {"detail":"Not Found"} for /app/health:

  • Use /api/health instead.

• RAG not ready:

  • Upload at least one PDF first.
  • Ensure vectorstore exists in DB_FAISS_BASE.

• Frontend loads but blank/errors:

  • Ensure static assets are served from /assets.
  • Hard refresh browser after deploy.

• Railway OOM during ingest:

  • Reduce EMBED_BATCH_SIZE (e.g. 4).
  • Reduce RAG_MAX_PDF_PAGES (e.g. 80).
  • Upload smaller PDFs first.

Security

• Never commit real API keys.
• Rotate keys immediately if leaked.