🏥 Medical Chatbot - NexgAI AI Engineering Challenge

🎯 Overview

An intelligent medical chatbot built using RAG (Retrieval-Augmented Generation) architecture that provides evidence-based medical answers from the MedMCQA dataset. The system combines semantic search with advanced AI to deliver accurate, contextual medical information.

✨ Key Features

• 🧠 182,822+ Medical Q&As from MedMCQA dataset
• 🔍 Semantic Vector Search with confidence scoring
• 🤖 AI-Powered Responses using Google Gemini 2.5 Flash
• 🌐 Beautiful Web Interface with real-time chat
• 📡 RESTful API for integration
• 🔄 LangGraph Flow for conversation management
• 📊 Confidence Indicators for response reliability

🏗️ Architecture

graph TD
    A[User Question] --> B[Medical Embedder]
    B --> C[Pinecone Vector Search]
    C --> D[Context Retrieval]
    D --> E[Confidence Check]
    E --> F{Confidence >= Threshold?}
    F -->|Yes| G[Google Gemini LLM]
    F -->|No| H[Fallback Response]
    G --> I[Structured JSON Response]
    H --> I
    I --> J[Web Interface / API]

Loading

🧩 Components

Component	Technology	Purpose
Embeddings	all-MiniLM-L6-v2	Convert text to 384-dim vectors
Vector DB	Pinecone	Semantic similarity search
LLM	Google Gemini 2.5 Flash	Generate medical responses
Flow	LangGraph	Conversation state management
API	FastAPI	Web interface & REST endpoints
Frontend	HTML/CSS/JS	Interactive chat interface

🚀 Quick Start

📋 Prerequisites

• Python 3.8+
• Google API Key (Get here)
• Pinecone API Key (Get here)
• MedMCQA Embeddings (see setup below)

⚡ Installation

1. Clone Repository

   git clone https://github.com/smit-faldu/Medical-Chatbot.git
   cd Medical-Chatbot

2. Create Virtual Environment

   python -m venv venv
   # Activate: venv\Scripts\activate (Windows) or source venv/bin/activate (macOS/Linux)

3. Install Requirements

   pip install -r requirements.txt

4. Set Environment Variables Create .env file:

   GOOGLE_API_KEY=your_google_api_key_here
   PINECONE_API_KEY=your_pinecone_api_key_here
   PINECONE_INDEX_NAME=medmcqa-embeddings
   PINECONE_ENVIRONMENT=us-east-1-aws

5. Run Notebook for Embeddings

   jupyter notebook pineconeembd.ipynb

   Or use Google Colab:

6. Run Server

   python main.py

🌐 Access Your Chatbot

Once running, access your chatbot at:

• 💬 Chat Interface: http://127.0.0.1:7860/ui
• 📚 API Documentation: http://127.0.0.1:7860/docs
• ❤️ Health Check: http://127.0.0.1:7860/health

🧪 Sample Questions

Try these medical questions:

• "What is hypertension and what causes it?"
• "What are the symptoms of diabetes mellitus?"
• "How does aspirin work as an antiplatelet agent?"
• "What is the difference between Type 1 and Type 2 diabetes?"
• "What are the side effects of ACE inhibitors?"

🎨 Design Choices & Justification

🔄 LangGraph Structure

Why LangGraph?

• ✅ State Management: Maintains conversation context and flow
• ✅ Conditional Logic: Handles confidence-based routing
• ✅ Error Handling: Graceful fallbacks for failed operations
• ✅ Modularity: Clean separation of concerns (retrieval → confidence → generation)
• ✅ Debugging: Clear state transitions for troubleshooting

Flow Design:

User Input → Retrieval → Confidence Check → LLM/Fallback → Response

🤖 LLM Choice: Google Gemini 2.5 Flash

Why Gemini 2.5 Flash?

• ✅ Medical Knowledge: Strong performance on medical queries
• ✅ JSON Output: Reliable structured response generation
• ✅ Speed: Fast inference for real-time chat
• ✅ Context Window: Large context for medical explanations
• ✅ Cost Effective: Good performance-to-cost ratio

🔍 Embedding Strategy: all-MiniLM-L6-v2

Why this model?

• ✅ Proven Performance: Excellent for medical Q&A similarity
• ✅ Efficiency: 384 dimensions - fast search, good accuracy
• ✅ Compatibility: Works well with MedMCQA dataset
• ✅ Resource Friendly: Runs efficiently on CPU

🗄️ Vector Database: Pinecone

Why Pinecone?

• ✅ Scalability: Handles 182K+ vectors efficiently
• ✅ Speed: Sub-second similarity search
• ✅ Reliability: Managed service with high uptime
• ✅ Metadata: Rich filtering and metadata support
• ✅ Integration: Excellent Python SDK

🎯 RAG Implementation Techniques

1. Confidence-Based Routing

   if confidence >= threshold:
       return llm_response
   else:
       return fallback_response

2. Context Formatting

   • Structured medical context from MedMCQA
   • Question + Options + Explanation + Subject
   • Optimized for LLM understanding

3. Response Validation

   • JSON schema validation
   • Fallback parsing for malformed responses
   • Error handling with graceful degradation

4. Semantic Search Optimization

   • Normalized embeddings for cosine similarity
   • Configurable similarity thresholds
   • Multi-document context aggregation

📁 Project Structure

nexgAI-medical-chatbot/
├── src/                          # Source code
│   ├── embedder.py              # Medical text embedder
│   ├── pinecone_retriever.py    # Vector search & retrieval
│   ├── llm.py                   # Google Gemini integration
│   ├── flow.py                  # LangGraph conversation flow
│   └── api.py                   # FastAPI web application
├── static/                       # Web interface
│   └── index.html               # Chat UI
├── data/                        # Data directory (optional)
├── pineconeembd.ipynb          # Embedding creation notebook
├── main.py                     # Application entry point
├── start.py                    # Quick start script
├── test_system.py              # System testing
├── check_pinecone.py           # Pinecone verification
├── requirements.txt            # Python dependencies
├── .env                        # Environment variables
└── README.md                   # This file

🧪 Sample Questions

Try these medical questions:

1. "What is hypertension and what causes it?"
2. "What are the symptoms of diabetes mellitus?"
3. "How does aspirin work as an antiplatelet agent?"
4. "What is the difference between Type 1 and Type 2 diabetes?"
5. "What are the side effects of ACE inhibitors?"
6. "What causes myocardial infarction?"
7. "How is pneumonia diagnosed?"
8. "What is the mechanism of action of beta-blockers?"

📊 Performance Metrics

Metric	Value
Database Size	182,822 medical Q&As
Response Time	~2-5 seconds
Search Accuracy	High confidence (>0.7) for medical queries
Embedding Dimension	384 (optimized for speed)
Concurrent Users	Supports multiple simultaneous chats

🛠️ API Usage

Chat Endpoint

curl -X POST "http://127.0.0.1:8000/chat" \
     -H "Content-Type: application/json" \
     -d '{"question": "What is hypertension?"}'

Response:

{
  "question": "What is hypertension?",
  "answer": "Hypertension is persistently elevated blood pressure...",
  "explanation": "Detailed medical explanation...",
  "key_points": ["High blood pressure", "Cardiovascular risk", "Treatment options"],
  "subject": "Cardiology",
  "confidence": 0.85,
  "source": "MedMCQA",
  "is_fallback": false
}

🔧 Configuration

Environment Variables

Variable	Description	Default
GOOGLE_API_KEY	Google Gemini API key	Required
PINECONE_API_KEY	Pinecone API key	Required
PINECONE_INDEX_NAME	Pinecone index name	medmcqa-embeddings
PINECONE_ENVIRONMENT	Pinecone environment	us-east-1-aws
PINECONE_MODEL	Embedding model name	all-MiniLM-L6-v2

Confidence Threshold

Adjust the confidence threshold in src/flow.py:

# Higher = more strict, Lower = more permissive
confidence_threshold = 0.3  # Default: 0.3

🐛 Troubleshooting

Common Issues

1. "Index not found" error

   # Check your Pinecone indexes
   python check_pinecone.py

2. "Model not found" error

   # Verify embedding model
   python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"

3. "API key invalid" error

   • Check your .env file
   • Verify API keys are correct
   • Ensure no extra spaces in keys

4. Low confidence responses

   • Lower the confidence threshold
   • Check if your question is medical-related
   • Verify Pinecone index has data

Debug Mode

Run with debug logging:

python main.py --debug

🚀 Deployment

Docker Deployment (Recommended)

We provide optimized multi-stage Docker builds for production deployment:

Quick Start with Docker

# Linux/macOS
chmod +x deploy.sh
./deploy.sh deploy prod

# Windows PowerShell
.\deploy.ps1 deploy prod

Manual Docker Compose

# Development
docker-compose up -d --build

# Production (with Nginx reverse proxy)
docker-compose -f docker-compose.prod.yml up -d --build

Features:

• ✅ Multi-stage build: Optimized image size (~500MB)
• ✅ Security: Non-root user, minimal attack surface
• ✅ Performance: Pre-cached models and dependencies
• ✅ Monitoring: Health checks and logging
• ✅ Scaling: Nginx load balancer ready

📖 See DEPLOYMENT.md for complete Docker deployment guide

Local Production

# Install production server
pip install gunicorn

# Run with Gunicorn
gunicorn src.api:app -w 4 -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:8000

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Run tests: python test_system.py
5. Submit a pull request

📄 License

This project is licensed under the MIT License.

🙏 Acknowledgments

• MedMCQA Dataset: Medical question-answer pairs
• Sentence Transformers: Embedding models
• Pinecone: Vector database platform
• Google: Gemini AI model
• LangChain: LangGraph framework

📞 Support

• 🧪 Run Tests: python test_system.py
• 📊 Check Health: http://127.0.0.1:8000/health
• 📚 API Docs: http://127.0.0.1:8000/docs
• 🔍 Debug: Check console logs for errors

---

🏥 Medical Chatbot - Providing Evidence-Based Medical Information
Built with ❤️ for the NexgAI AI Engineering Challenge