Multi-Agent RAG System

A multi-agent Retrieval-Augmented Generation (RAG) system with Agent2Agent communication, built with LangGraph. The pipeline uses three specialized agents (Researcher, Summarizer, and Answer) that work in sequence to answer user questions from a document corpus.

Overview

• Researcher Agent: Retrieves relevant document chunks from a vector store (Chroma) using semantic search.
• Summarizer Agent: Summarizes the retrieved documents into a concise context.
• Answer Agent: Generates a final answer to the user's question using the summarized context.

Communication between agents follows a standardized Agent2Agent (A2A) protocol for request/response messages.

Architecture

User Query → [Retrieve] → [Summarize] → [Answer] → Final Answer
                ↓              ↓            ↓
           Chroma DB      Summary      LLM Answer

• State: Shared workflow state is defined in state.py (MultiAgentState).
• Workflow: Orchestrated in main.py via a LangGraph StateGraph with nodes: retrieve → summarize → answer.
• Agents: Implemented in agent.py with a common BaseAgent and A2A message handling from a2a_protocol.py.

Project Structure

multi_agent_rag/
├── README.md           # This file
├── main.py             # Entry point, workflow definition, RAG setup
├── agent.py            # ResearcherAgent, SummarizerAgent, AnswerAgent
├── state.py            # MultiAgentState (TypedDict)
├── a2a_protocol.py     # AgentMessage format and create_message()
├── documents/          # Source .txt documents for RAG
│   ├── artificial_intelligence.txt
│   ├── natural_language_processing.txt
│   ├── robotics.txt
│   ├── data_science.txt
│   ├── neural_networks.txt
│   ├── computer_vision.txt
│   └── machine_learning.txt
└── chroma_db/          # Persisted vector store (created at runtime)

Requirements

• Python 3.x
• Ollama running locally with the llama3 model (and optionally nomic-embed-text for embeddings)

From the project root, install dependencies:

pip install -r requirements.txt

Key dependencies: langgraph, langchain, langchain-ollama, chromadb, langchain-community, langchain-text-splitters, python-dotenv.

Setup

1. Install dependencies (see above).
2. Ensure Ollama is running and you have pulled:
   • llama3 (chat model)
   • nomic-embed-text (embeddings)
3. Optional: Add a .env in the project root if you need to configure environment variables.

Usage

From the project root (RAG_Learning):

python -m multi_agent_rag.main

Or from inside multi_agent_rag/:

python main.py

On startup, the system:

1. Loads .txt files from multi_agent_rag/documents/.
2. Splits them into chunks (500 chars, 50 overlap) and builds a Chroma vector store in multi_agent_rag/chroma_db/.
3. Starts an interactive loop.

Example:

Your question: What is machine learning?

Answer:
[... model's answer based on retrieved and summarized context ...]

Workflow Status: complete
Documents Retrieved: 5

Type quit, exit, or q to exit.

Agent2Agent Protocol

Messages between agents use the format in a2a_protocol.py:

• Fields: from_agent, to_agent, message_type (request/response/notification/error), task, data, timestamp, message_id, correlation_id.
• Helpers: create_message() builds a properly formatted AgentMessage.

Agents extend BaseAgent and implement process_message() to handle incoming A2A messages and use send_message() to reply.

Configuration

• Documents directory: Default is ./multi_agent_rag/documents. If missing, a sample document is created.
• Chroma persist path: ./multi_agent_rag/chroma_db.
• LLM: ChatOllama(model="llama3", temperature=0) in main.py.
• Embeddings: OllamaEmbeddings(model="nomic-embed-text").
• Retrieval: Top k=5 chunks per query; chunk size 500, overlap 50.

You can change these in main.py and agent.py as needed.

License

Part of the RAG Learning project. Use and modify as needed for learning and experimentation.