MARL-RAGDoc

Reference implementation of the MARL-RAGDoc framework described in the design notes. The project demonstrates how to build a hierarchical multi-agent retrieval-augmented generation (RAG) pipeline that can reason over multimodal industrial documents.

Project Layout

• src/marl_ragdoc/data_types.py – shared dataclasses (e.g., Element, Candidate).
• src/marl_ragdoc/parsing/multimodal_parser.py – document ingestion, OCR stubs, embedding hooks.
• src/marl_ragdoc/retrieval/candidate_pool.py – projection layers, similarity scoring, top-k pooling.
• src/marl_ragdoc/agents/ – coordinator and modality agent policies plus neural networks.
• src/marl_ragdoc/environment/rag_env.py – hierarchical MARL environment that glues parsing, retrieval, and reasoning.
• src/marl_ragdoc/reasoning/collaborative_reasoner.py – weighted aggregation, reflection loop, LLM interface.
• src/marl_ragdoc/training/trainer.py – RL training harness, losses, rollout helpers.
• src/marl_ragdoc/pipeline.py – high-level orchestration entry point.

Getting Started

poetry install
poetry run python -m marl_ragdoc.pipeline --config configs/demo.yaml

You must provision external OCR, table extraction, and LLM endpoints. The included components expose abstract hooks so you can plug in the tools used in your environment.

Status

The code focuses on architecture clarity: state/action definitions, reward wiring, and module interfaces match the MARL-RAGDoc description. Production deployments should replace the mocked subsystems (OCR, embeddings, LLM scoring) with real services, add persistence for experience replay, and extend evaluation.