How valencienseow/herai---codeblue works
Project type and purpose
• HerAI is a full-stack clinical decision support system for maternal health.
• Frontend: Next.js/TypeScript UI with Tailwind CSS, RAG (retrieval-augmented generation) endpoint.
• Backend: FastAPI Python service with ML risk-prediction models, SQLite DB, PDF CPG storage.
• ML: scikit-learn (.joblib) and Keras models.
Top-level components
• Frontend (herai-frontend/)
– Next.js App Router under src/app
– API routes (e.g., /api/rag) for local LLM or RAG calls
– React components, Tailwind styling
– localStorage mock or HTTP client (src/lib/api.ts, supabaseClient.js)
– Dockerized “llm-app” for LLM service (docker/llm-app)
• Backend (herai-backend/)
– main.py launches FastAPI, registers routes for patients, vitals, guidelines
– supabase_client.py (optional integration)
– cpg_helper.py for guideline lookup from PDF corpus (cpg_documents/)
– ml/ml_service.py wrapping ML models in models/
– SQLite database herai.db
• External/Infrastructure
– Docker container for LLM (llm-app)
– Potential Supabase or external API for production
– Vercel hosting frontend, Railway/Heroku for backend
Relationships & interactions
• User → Browser → Next.js UI
• Next.js fetches:
a) Mock/local patient data via localStorage or
b) HTTP to FastAPI endpoints at http://localhost:8000/patients, /vitals, /risk, /cpg
• FastAPI → SQLite for CRUD patient & vitals
• FastAPI → ml_service to predict risk → returns risk score & next-step recommendations
• FastAPI → cpg_helper to search embedded PDF guidelines → returns relevant CPG text
• Next.js → RAG API route (/api/rag) → Dockerized LLM service → generates narrative insights
• Optional: Frontend ↔ Supabase (auth, real DB) via supabaseClient.js
Architectural patterns & design principles
• Clear separation of concerns: UI, API layer, ML service, guideline search, persistence
• Modular ML encapsulation (ml_service.py + models folder)
• RAG microservice pattern for LLM (docker/llm-app)
• Component-based React UI
• Configurable backend client (supabase for production)
Detail
• Separation of static assets (public/ folder) vs. dynamic data
• Docker llm-app microservice for LLM/RAG
• Call out mock vs. real API integration path in src/lib/api.ts
• PDF guideline store as a filesystem mount under backend
• Environment/config layers (dotenv, next.config.ts, docker-compose env files)