Skip to content

README.md

Latest commit

 

History

History
151 lines (107 loc) · 4.4 KB

README.md

File metadata and controls

151 lines (107 loc) · 4.4 KB

Multi-Agent-RAG

A powerful document question-answering system powered by Docling and LangGraph, featuring a multi-agent workflow for accurate and verified answers.

Features

  • Multi-Agent Workflow: Research, verification, and relevance checking agents
  • Document Processing: Support for PDF, DOCX, TXT, and MD files
  • Hybrid Retrieval: Combines dense and sparse retrieval for better results
  • Answer Verification: Built-in fact-checking and verification
  • Modern Chat Interface: Powered by Chainlit
  • Session Management: Efficient document caching and state management

Quick Start

Prerequisites

  • Python 3.10+ (recommended)
  • uv (for fast dependency management)

Installation

  1. Clone the repository:

    git clone <repository-url>
cd Multi-Agent-RAG

  2. Install dependencies using uv:

    uv pip install -e .

    If you don’t have uv installed, get it with:

    pip install uv

Running the Chainlit Interface

Start the Chainlit chat interface:

chainlit run chainlit.py

The interface will be available at http://localhost:8000.

Usage

Document Upload

  1. Upload your documents (PDF, DOCX, TXT, MD) in the chat interface.
  2. The system will process and index them automatically.
  3. Documents are cached for efficient reuse.

Asking Questions

  1. Type your question in the chat.
  2. The system will:
    • Check document relevance
    • Research the answer using retrieved documents
    • Verify the answer for accuracy
    • Provide both the answer and verification report

Architecture

Multi-Agent Workflow

  1. Relevance Checker: Determines if documents can answer the question
  2. Research Agent: Generates initial answers from retrieved documents
  3. Verification Agent: Checks answer accuracy and provides verification report

Document Processing

  • File Handler: Processes various document formats
  • Chunking: Splits documents into manageable chunks
  • Embedding: Creates vector representations for retrieval

Retrieval System

  • Hybrid Retriever: Combines dense (vector) and sparse (BM25) retrieval
  • ChromaDB: Vector database for document storage
  • Ensemble: Merges results from multiple retrieval methods

Configuration

Settings are managed through the config/ directory:

  • constants.py: System constants and file type definitions
  • settings.py: Environment-specific settings

Supported File Types

  • PDF (.pdf)
  • Word documents (.docx)
  • Text files (.txt)
  • Markdown files (.md)

Development

Project Structure

Multi-Agent-RAG/
├── agent/                 # Multi-agent workflow
│   ├── workflow.py        # Main workflow orchestration
│   ├── research_agent.py  # Research agent implementation
│   ├── verification_agent.py # Verification agent
│   └── relevance_checker.py  # Relevance checking
├── retriever/             # Document retrieval system
│   ├── builder.py         # Retriever construction
│   └── file_handler.py    # Document processing
├── config/                # Configuration files
├── utils/                 # Utility functions
├── chainlit.py            # Chainlit interface (entrypoint)
├── pyproject.toml         # Project dependencies
└── uv.lock                # uv dependency lockfile

Adding New Agents

  1. Create a new agent class in the agent/ directory
  2. Implement the required interface methods
  3. Add the agent to the workflow in agent/workflow.py

Extending Document Support

  1. Add new file type to constants.ALLOWED_TYPES
  2. Implement processing logic in retriever/file_handler.py
  3. Update the interface validation

Troubleshooting

Common Issues

  1. Document Processing Errors: Ensure files are not corrupted and in supported formats
  2. Memory Issues: Large documents may require more memory allocation
  3. Retrieval Performance: Consider adjusting chunk sizes or retrieval parameters

Logging

The system uses structured logging. Check logs for detailed error information:

tail -f chainlit.log  # For Chainlit interface

Acknowledgments