Math Routing Agent 🧠

This project is an Agentic-RAG (Retrieval-Augmented Generation) system designed to function as an AI-powered mathematics professor. It intelligently answers mathematical questions by either retrieving information from a dedicated knowledge base or searching the web. The agent includes AI guardrails for safety and a human-in-the-loop mechanism to learn from user feedback.

---

🔍 Features

• Intelligent Routing: The agent dynamically decides whether to use its internal knowledge base or perform a web search to find the best answer for a given question.

• AI Guardrails: An AI Gateway moderates all inputs and outputs, ensuring questions are math-related and that answers are safe and on-topic.

• Vector Knowledge Base: Uses Qdrant to store and retrieve existing math problems and solutions, acting as the agent's long-term memory.

• Live Web Search: Leverages the Tavily AI API to find up-to-date information online for questions not covered in the knowledge base.

• Human-in-the-Loop Feedback: Users can provide feedback on answers, which the agent uses to refine its response, enabling a self-correction and learning loop.

• Structured Context (MCP): Implements the Model Context Protocol (MCP) by passing a structured state object through the entire workflow, ensuring every component has full context.

---

🧰 Tech Stack

• Backend: Python, FastAPI
• Frontend: React
• AI Agent Framework: LangGraph
• LLM: Google Gemini 1.5 Pro
• Vector Database: Qdrant
• Web Search Tool: Tavily AI
• Embeddings: Hugging Face Sentence Transformers

---

📁 Project Structure

math-routing-agent/
├── backend/
│   ├── app/
│   │   ├── agent.py         # The core LangGraph agent workflow
│   │   ├── feedback.py      # Human-in-the-loop refinement logic
│   │   ├── guardrails.py    # AI Gateway for moderation
│   │   ├── main.py          # FastAPI server and API endpoints
│   │   └── schemas.py       # Pydantic data models (for MCP)
│   ├── data/
│   │   └── math_dataset.json # The knowledge base source file
│   ├── scripts/
│   │   └── setup_knowledge_base.py # Script to create the VectorDB
│   ├── run_benchmark.py     # Script for JEE Bench evaluation
│   └── .env                 # API keys and environment variables
└── frontend/
    ├── src/
    │   └── App.js           # Main React component for the UI
    └── ...                  # Other standard React files

---

🚀 Setup and Installation

✅ Prerequisites

• Python 3.8+
• Node.js and npm

---

🖥️ Backend Setup

Navigate to the backend directory:

cd backend

Create and activate a virtual environment:

python -m venv venv
.�env\Scripts�ctivate

Install Python dependencies:

pip install -r requirements.txt

Set up API Keys:

1. Create a file named .env in the backend directory.
2. Add your API keys:

TAVILY_API_KEY="your_tavily_api_key"
GOOGLE_API_KEY="your_google_api_key"

Build the Knowledge Base:

1. Place your math_dataset.json file in the backend/data/ folder.
2. Run the setup script to populate the Qdrant database:

cd scripts
python setup_knowledge_base.py
cd ..

---

🌐 Frontend Setup

Navigate to the frontend directory:

cd frontend

Install Node.js dependencies:

npm install

---

▶️ How to Run the Application

First and foremost, we need to setup the knowledge base

Go to backend/scripts and run the file "setup_knowledge_base.py" to build the vector db

🔧 Start the Backend Server:

In a terminal, navigate to the backend directory and run:

uvicorn app.main:app --reload

The server will be running at:
http://localhost:8000

---

💻 Start the Frontend Application:

In a separate terminal, navigate to the frontend directory and run:

npm start

Your browser will automatically open to:
http://localhost:3000, where you can interact with the agent.

---