Federated Retrieval-Augmented Generation: A Systematic Mapping Study

Abhijit Chakraborty, Chahana Dahal, Vivek Gupta

Findings of the Association for Computational Linguistics: EMNLP 2025, pages 7362–7374Peer reviewed

• PDF
• ACL Anthology
• arXiv
• Code

Summary

Federated RAG combines two ideas. Federated learning trains models across parties without moving their raw data, and retrieval-augmented generation (RAG) grounds a language model's output in retrieved documents. This paper is the first systematic mapping study of the area, covering work published from 2020 to 2025.

Following evidence-based software engineering guidelines, we classified each paper by research focus, contribution type and application domain, then traced architectural patterns and how the field changed over time. Three open problems kept coming up: retrieval that preserves privacy, differences between clients, and the limits of current evaluation. They matter most in privacy-sensitive domains such as healthcare, finance and personal assistants.

The full abstract is on ACL Anthology.

Related work from this project

• SYNAPSE: Federated Tool Routing via Typed Compendium Artifacts
• Typed Federated Artifacts for the Agentic Web

Cite

@inproceedings{chakraborty-etal-2025-federated,
  title     = {Federated Retrieval-Augmented Generation: A Systematic Mapping Study},
  author    = {Chakraborty, Abhijit and Dahal, Chahana and Gupta, Vivek},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2025},
  year      = {2025},
  pages     = {7362--7374},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.findings-emnlp.388/}
}