RAG-Based Multi-Agent Chatbot

A customer support chatbot that uses RAG (Retrieval-Augmented Generation) with OpenAI and automatically routes queries to specialized AI agents.

Features

• Text Classification: Automatically classifies user queries and routes them to the appropriate agent
• RAG-Powered Responses: Uses vector database to retrieve relevant context from PDF documents
• Multi-Agent Architecture: Three specialized agents handle different types of queries
• Persistent Storage: ChromaDB stores document embeddings locally

Agents

Agent	Purpose	Knowledge Base
Product Usage Agent	How to use the product, cooking instructions, cleaning, maintenance	product_info database
Post-Sales Agent	Warranty, refunds, returns, order tracking	post_sales database
Complaint Agent	Customer complaints, issues, compensation requests	Both databases

Project Structure

text-classifier/
├── main.py              # CLI chatbot interface
├── text_classifier.py   # Query classification and agent routing
├── agents.py            # Specialized AI agents
├── vector_store.py      # PDF ingestion and ChromaDB vector store
├── documents/           # Place PDF files here
│   ├── product.pdf      # Product usage documentation
│   └── postsales.pdf    # Warranty and refund policies
├── chroma_db/           # Persistent vector database (auto-created)
├── pyproject.toml       # Project dependencies
└── .env                 # OpenAI API key

Installation

1. Clone the repository

   git clone https://github.com/yourusername/text-classifier.git
   cd text-classifier

2. Install dependencies using uv

   uv sync

3. Set up environment variables

   cp .env.example .env
   # Edit .env and add your OpenAI API key

4. Add your PDF documents

   • Place product documentation in documents/product.pdf
   • Place warranty/refund policies in documents/postsales.pdf

Usage

Run the chatbot

uv run python main.py

Commands

Command	Description
ingest product <path>	Ingest a PDF into the Product Info database
ingest postsales <path>	Ingest a PDF into the Post-Sales database
reset	Clear all databases and restart
quit or exit	Exit the chatbot

Reset Database

To clear all vector databases and start fresh:

You: reset
Databases cleared. Please restart the chatbot.

Then restart the chatbot to re-ingest your PDFs:

uv run python main.py

Example Conversation

You: How do I clean the air fryer?

[Product Usage Agent (Air Fryer)]
Bot: To clean your Philips Air Fryer, first unplug it and let it cool down...

You: What is the warranty period?

[Post-Sales Agent (Warranty & Refunds)]
Bot: The warranty period for your Philips Air Fryer is two (2) years from the original date of purchase...

You: My air fryer stopped working after 1 month!

[Complaint Resolution Agent]
Bot: I'm sorry to hear about this issue with your air fryer. Since your product is within the 2-year warranty period...

How It Works

1. PDF Ingestion: PDFs are chunked into smaller segments and embedded using OpenAI's text-embedding-3-small model
2. Query Classification: User queries are classified into one of three categories using GPT-4o-mini
3. Context Retrieval: Relevant document chunks are retrieved from the appropriate vector database
4. Response Generation: The specialized agent generates a response using the retrieved context

Configuration

Environment Variables

Variable	Description
OPENAI_API_KEY	Your OpenAI API key

Customization

• Modify agent prompts in agents.py to customize agent behavior
• Adjust classification categories in text_classifier.py
• Change chunking parameters in vector_store.py (chunk_size, overlap)

Dependencies

• openai - OpenAI API client
• chromadb - Vector database
• pypdf - PDF text extraction
• python-dotenv - Environment variable management

License

MIT License