fed-rag 0.0.4

A framework for federated fine-tuning of retrieval-augmented generation (RAG) systems.

FedRAG

---

FedRAG is a framework for federated fine-tuning of Retrieval-Augmented Generation (RAG) systems, wherein a server and potentially several client nodes share an overall model architecture. Since the model is common to all participants of the system, fine-tuning can be done collaboratively without any of the raw data leaving any of the client nodes. Instead only the model weight updates are shared between the client and the server.

./fed_rag
├── __init__.py
├── base
│   ├── __init__.py
│   ├── loss.py
│   └── models
│       ├── __init__.py
│       ├── generator.py # BaseGeneratorModel
│       ├── rag.py # BaseRAGModel
│       └── retriever.py # BaseRetrieverModel
├── loss
│   ├── __init__.py
│   ├── generator  # Losses for generation task
│   │   └── __init__.py
│   └── retriever # Losses for retrieval task
│       └── __init__.py
├── models
│   ├── __init__.py
│   ├── generators # Generator models
│   │   └── __init__.py
│   └── retrievers # Retrieval models
│       └── __init__.py
├── ops # module for running fed system
│   └── __init__.py
└── types
    ├── __init__.py
    ├── client.py # Wrapper for flwr.Client
    └── server.py # Wrapper for flwr.Server

Getting Started

Contributing

Environment setup:

# while in root directory of project `fed-rag/`
uv sync --all-extras --dev