RAG-Based Multi-Domain AI Chatbot with Cohere + Gemini

This project demonstrates how to build a scalable Retrieval-Augmented Generation (RAG) chatbot that can answer queries across multiple knowledge domains using vector embeddings and large language models.

---

🚀 What You’ll Learn

• How to use Cohere for embedding documents and queries
• Setting up ChromaDB for similarity search
• Integrating Google Gemini LLM for natural language responses
• Using FastAPI + React for a clean full-stack experience
• Optimizing with Redis and credit tracking
• Supporting multiple knowledge domains (e.g., Education, Legal, Healthcare)

---

🧱 Tech Stack Overview

• Backend: Python, FastAPI, Cohere, ChromaDB, Google Gemini API, Upstash Redis
• Frontend: React.js, Tailwind CSS
• Storage: Local file system + ChromaDB vector store
• Auth & Usage Control: Redis-based credit system

---

🧠 How It Works

1. Ingestion Phase

   • Documents are chunked and embedded using Cohere.
   • Embeddings are stored in ChromaDB under domain-specific collections.

2. Query Phase

   • User inputs a query through the React interface.
   • The query is embedded at runtime and matched against the selected domain's ChromaDB collection.
   • Retrieved contexts are combined into a prompt and passed to Google Gemini for answer generation.

3. Redis for Credit Control

   • Tracks per-user API usage with TTL (Time-to-Live) logic.
   • Helps manage rate limits and cost control.

---

📦 Sample Folder Structure

Backend_Server/
│
├── app/                  # Core logic
│   ├── main.py
│   ├── retriver.py
│   ├── embedder.py
│   ├── llm.py
│   └── utils/
│       └── credit_tracker.py
│
├── chroma_store/         # Chroma vector store
├── sampleData/           # Source docs in JSON
├── requirements.txt
├── .env
└── .gitignore

---

🛠️ Redis Credit Tracking Example

def track_credits(user_id: str) -> bool:
    key = f"user:{user_id}:credits"
    current = redis.get(key)

    if current is None:
        redis.set(key, DEFAULT_CREDITS - 1, ex=TTL_SECONDS)
        return True
    if int(current) <= 0:
        return False

    redis.decr(key)
    return True

---

💻 Frontend with React + Tailwind

• Clean interface for domain selection
• Query input box with real-time results
• Context citations for each answer
• Responsive layout with Tailwind CSS

---

✅ Prerequisites

• Python 3.9+
• Node.js 18+
• Redis (Upstash or local)
• ChromeDB (local file-based)
• API keys: GEMINI_API_KEY, COHERE_API_KEY, REDIS_URL, REDIS_TOKEN

---

🔧 Installation

# Clone the repository
git clone https://github.com/your-username/rag-chatbot.git
cd rag-chatbot/Backend_Server

# Create virtual environment
python -m venv venv
venv\Scripts\activate     # on Windows

# Install dependencies
pip install -r requirements.txt

---

⚙️ Configuration

Create a .env file in the root:

COHERE_API_KEY=your_key
GEMINI_API_KEY=your_key
REDIS_URL=https://...
REDIS_TOKEN=your_token

---

💡 Usage Guide

1. Start Backend Server

cd app
uvicorn main:app --reload

2. Start Frontend Server

cd Chat_UI
npm install
npm run dev

3. Interact via Browser

• Open http://localhost:5173
• Ask queries related to healthcare, education, law, etc.

---

🔒 Security Features

• ✅ Environment variables for all sensitive keys
• ✅ Credit-based usage control via Redis
• ✅ Token-level rate limiting (Upstash)
• ⛔ No hardcoded credentials in production

---

🎯 Performance Optimization

• ⚡ Efficient semantic search using ChromaDB
• 🔁 Batched embedding with Cohere
• 🔄 Redis caching to reduce API overuse
• 📦 Modular design to plug-and-play new domains

---

🔄 Future Enhancements

• 🧬 Add support for OpenAI/SentenceTransformers
• 🗃️ Optional Pinecone/Weaviate for scalable vector storage
• 🧑‍💻 User authentication and role-based access
• 🧠 Feedback learning loop for continuous improvement
• 📊 Admin dashboard for monitoring & analytics
• 📦 Docker + CI/CD for deployment pipeline

---

🙏 Acknowledgments

• Google Gemini API
• Cohere
• ChromaDB
• FastAPI
• Upstash Redis
• Open-source community & contributors

---

🎉 Conclusion

This RAG-based chatbot demonstrates the power of combining modern LLMs with vector databases for domain-specific knowledge retrieval. The integration of Cohere embeddings with Google Gemini provides accurate, contextual responses while maintaining scalability through Redis caching and ChromaDB's efficient vector operations.

Key Takeaways:

• ✅ RAG significantly improves response accuracy for specialized domains
• ✅ Vector embeddings enable semantic search across large knowledge bases
• ✅ Proper caching and rate limiting are crucial for production deployment
• ✅ Modular architecture allows easy expansion to new domains

Feel free to explore the code, contribute improvements, or adapt it for your own use cases!