Massive Multimodal Open RAG & Extraction

mmore is an open-source, end-to-end pipeline to ingest, process, index, and retrieve knowledge from heterogeneous files: PDFs, Office docs, spreadsheets, emails, images, audio, video, and web pages. It standardizes content into a unified multimodal format, supports distributed CPU/GPU processing, and provides hybrid dense+sparse retrieval with an integrated RAG service (CLI, APIs).

👉 Read the paper for more details (arXiv): mmore: Massive Multimodal Open RAG & Extraction

Documentation

👉 Read the full documentation here: mmore Documentation.

:bulb: Quickstart

Installation

:whale: Prefer Docker? Skip the steps below and pull a pre-built multi-platform image directly from GHCR, with CPU and GPU variants:

docker pull ghcr.io/EPFLiGHT/mmore:edge-gpu   # GPU (CUDA 12.6)
docker pull ghcr.io/EPFLiGHT/mmore:edge-cpu   # CPU-only

See docker/ubuntu/README.md for build instructions and additional base OS variants (Arch Linux, openSUSE Leap).

(Step 0 – Install system dependencies)

Our package requires system dependencies. This snippet will take care of installing them for Linux!

sudo apt update
sudo apt install -y ffmpeg libsm6 libxext6 libnss3 \
  libxi6 libxrandr2 libxcomposite1 libxcursor1 libxdamage1 \
  libxfixes3 libxrender1 libasound2 libatk1.0-0 libgtk-3-0 libreoffice \
  libpango-1.0-0 libpangoft2-1.0-0 weasyprint

:warning: On Ubuntu 24.04, replace libasound2 with libasound2t64. You may also need to add the repository for Ubuntu 20.04 focal to have access to a few of the sources (e.g. create /etc/apt/sources.list.d/mmore.list with the contents deb http://cz.archive.ubuntu.com/ubuntu focal main universe).

For MacOS, use instead:

brew update
brew install ffmpeg gtk+3 pango cairo \
  gobject-introspection libffi pkg-config libx11 libxi \
  libxrandr libxcomposite libxcursor libxdamage libxext \
  libxrender atk libreoffice weasyprint

If weasyprint fails to find GTK or Cairo, also run:

brew install cairo pango gdk-pixbuf libffi
uv pip install weasyprint

You can also run mmore on Windows by following our Windows setup notes.

Step 1 – Install mmore

Dependencies are split by pipeline stage. Install only what you need:

Full install (CPU):

uv pip install "mmore[all,cpu]"

Full install (GPU — CUDA 12.6):

uv pip install "mmore[all,cu126]"

Partial install example (processing only):

uv pip install "mmore[process,cpu]"

:warning: This package requires many big dependencies, so it is recommended to install with uv to handle pip installations. Check our tutorial on uv.

:warning: Check the instructions for contributors directly in the documentation for_devs.md.

Interactive TUI

Prefer a guided experience over editing YAML by hand? Install the tui extra and launch the interactive Terminal UI:

uv pip install "mmore[tui]"
mmore tui

From the launcher you can:

• run any stage (process / postprocess / index / rag / chat) interactively,
• chain the full pipeline (process → postprocess → index → chat),
• generate stage YAML configs through a guided wizard,
• pick from existing example configs without leaving the terminal.

Minimal Example

You can use our predefined CLI commands to execute parts of the pipeline. Note that you might need to prepend python -m to the command if the package does not properly create bash aliases.

# Run processing
python -m mmore process --config-file examples/process/config.yaml
python -m mmore postprocess --config-file examples/postprocessor/config.yaml --input-data examples/process/outputs/merged/merged_results.jsonl

# Run indexer
python -m mmore index --config-file examples/index/config.yaml --documents-path examples/postprocessor/outputs/merged/results.jsonl

# Run RAG
python -m mmore rag --config-file examples/rag/config.yaml

You can also use our package in python code as shown here:

from mmore.process.processors.pdf_processor import PDFProcessor
from mmore.process.processors.base import ProcessorConfig
from mmore.type import MultimodalSample

pdf_file_paths = ["/path/to/examples/sample_data/pdf/calendar.pdf"] #write here the full path, not a relative path
out_file = "/path/to/examples/process/outputs/example.jsonl"

pdf_processor_config = ProcessorConfig(custom_config={"output_path": "examples/process/outputs"})
pdf_processor = PDFProcessor(config=pdf_processor_config)
result_pdf = pdf_processor.process_batch(pdf_file_paths, False, 1) # args: file_paths, fast mode (True/False), num_workers

MultimodalSample.to_jsonl(out_file, result_pdf)

Usage

To launch the mmore pipeline, follow the specialised instructions in the docs.

1. :page_facing_up: Input Documents Upload your multimodal documents (PDFs, videos, spreadsheets, and m(m)ore) into the pipeline.

2. :mag: Process Extracts and standardizes text, metadata, and multimedia content from diverse file formats. Easily extensible! You can add your own processors to handle new file types. Supports fast processing for specific types.

3. :file_folder: Index Organizes extracted data into a hybrid retrieval-ready Vector Store DB, combining dense and sparse indexing through Milvus. Your vector DB can also be remotely hosted and then you only have to provide a standard API. There is also an HTTP Index API for adding new files on the fly with HTTP requests.

4. :robot: RAG Use the indexed documents inside a Retrieval-Augmented Generation (RAG) system that provides a LangChain interface. Plug in any LLM with a compatible interface or add new ones through an easy-to-use interface. Supports API hosting or local inference.

5. :globe_with_meridians: Web Search Augments RAG answers with live web search results using an iterative sub-query loop. DuckDuckGo is the default provider (free, no API key needed). Tavily is available as an optional higher-quality provider.

     # Install web search dependencies
     uv pip install "mmore[rag,websearch]"
   
     # Optional: use Tavily instead of DuckDuckGo
     export TAVILY_API_KEY=your_key_here
   

6. :tada: Evaluation Coming soon An easy way to evaluate the performance of your RAG system using Ragas.

:construction: Supported File Types

License

This project is licensed under the Apache 2.0 License, see the LICENSE :mortar_board: file for details.

Cite mmore

If you use mmore in your research, please cite the paper:

@misc{sallinen2025mmoremassivemultimodalopen,
      title={mmore: Massive Multimodal Open RAG & Extraction}, 
      author={Alexandre Sallinen and Stefan Krsteski and Paul Teiletche and Marc-Antoine Allard and Baptiste Lecoeur and Michael Zhang and Fabrice Nemo and David Kalajdzic and Matthias Meyer and Mary-Anne Hartley},
      year={2025},
      eprint={2509.11937},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2509.11937}, 
}

Star History