This project is a fully modular Multi-Modal Multi-Agent Retrieval-Augmented Generation (RAG) system capable of processing PDFs, HTMLs, images, and tables for answering questions using a pipeline of specialized agents:

• Text Agent + Image Agent: For generating insights from the retrieved contexts of the RAG system

• Generalize Agent would combine and generalize the answers from TextAgent and ImageAgent for each question

• Planning Agent would receive the query from the user and then separates it into several "tasks" or questions for retrieving many information from the RAG system.

• Merge Agent would combine all the responses from the Generalize Agent and merge them into a response which would answer the initial query from the user.

• Verfier Agent would score the combined answer of the Merge Agent and then telling if we need to query for more information by generating following up questions for continually retrieve information.

It supports local document extraction via Docling, embedding with SentenceTransformers, and multi-agent orchestration.

M3ARAG/
├── agents/                 # Modular agent logic.
├── pipeline/               # Pipeline and Chat launcher interface
├── data/                   # Storing the downloaded files
│   ├── store/              # Raw downloaded files (PDF, HTML, etc.)
│   ├── merge/              # Single processing location for indexing of RAG.
│   └── extract/            # Converted PDFs, extracted images/tables
├── RAG/                    # RAG system
├── config/                 # Config files for RAG, Agents and Prompt file
│   ├── agent_config.py     # Config for using Agents
│   ├── rag_config.py       # Config for using RAG
│   └── prompt.py           # Prompts Storage.
├── rag_text/               # RAG text captioning
├── rag_image/              # RAG image captioning
├── utils/                  # Helper utilities (e.g., process_documents)
├── test/                   # Testing places
├── main.py                 # Main entry point
├── chat_streamlit.py       # Main function for chatting via streamlit
├── README.md               # Main information about the repository
├── timeline.md             # Tasks and next tasks that we have done

git clone https://github.com/pdz1804/M3ARAG.git
cd M3ARAG

python -m venv myenv
# Or: conda create -n m3arag python=3.10

# Activate
source myenv/bin/activate  # macOS/Linux
myenv\Scripts\activate     # Windows

pip install -r requirements.txt

💡 If requirements.txt is missing, install manually like below:

pip install sentence-transformers langchain openai chromadb docling python-dotenv

1. Download Poppler for Windows:

   • Visit: https://github.com/oschwartz10612/poppler-windows/releases/
   • Download the latest .zip file under Assets (e.g., poppler-xx_xx_xx.zip).

2. Extract the zip to a location like C:\poppler.

3. Add Poppler to PATH:

   • Open Start > Environment Variables.
   • Under System Variables, find and select Path, click Edit.
   • Click New and add:

     C:\poppler\Library\bin
     

   • Click OK and restart your terminal.

4. Verify installation:

   where pdfinfo

   You should see:

   C:\poppler\Library\bin\pdfinfo.exe
   

brew install poppler

To verify:

which pdfinfo

sudo apt update
sudo apt install poppler-utils

To verify:

which pdfinfo

Copy .env.example and rename to .env, then fill in your keys:

OPENAI_API_KEY=pdz-...
GOOGLE_API_KEY=pdz-...

If you want to run RAG-flow individually without Agents or with Agents:

# Download data only (for local testing)
python main.py --download

# Ingest data only
python main.py --ingest

# Chatting 
python main.py --chat

# Small note: we can run --download --ingest --chat at once

# Run it on streamlit: by uploading docs or inputing urls
python main.py --app

This will:

• Download and store files from hardcoded URLs.
• Extract content using Docling.
• Index text via SentenceTransformers + Chroma.
• Start interactive agent-based chat loop.

Toggle to view Document Processing & Indexing

This diagram shows how documents are split into chunks and images, indexed via ChromaDB and stored on disk.

Toggle to view Multi-Modal Retrieval Pipeline

Illustration of text and image-based retrieval using sub-queries from the user question.

Toggle to view Agent-Oriented Workflow

Overview of how multiple specialized agents interact to process, merge, verify, and answer complex queries.

• ✅ PDF documents (.pdf)
• ✅ HTML, MD, PPTX, CSV, DOCX, TXT (converted to PDF)
• ✅ Extracted images (captioning + indexing coming soon)
• 🧪 Support for audio, .json, .xml being tested for later release

This project is licensed under the MIT License. See LICENSE for details.

Built by Nguyen Quang Phu (pdz1804) and Tieu Tri Bang

Reach out or open an issue for support or ideas.