Adaptive Federated Retrieval-Augmented Generation over Distributed Big Data Systems

Saideh Naderi¹ Seyyed Hamid Ghafouri² Azam Sadat Mirtajodini³ Raziyeh Khojasteh nejad⁴

1) 1. Ph.D Student Department of Computer Engineering, Islamic Azad University Kerman Branch, Iran,
2) 2. Department of Computer Engineering, Ke.C. Islamic Azad University, Kerman,Iran,
3) 3. Ph.D Student Department of Computer Engineering, Islamic Azad University Kerman Branch, Iran,
4) 4. Ph.D Student Department of Computer Engineering, Islamic Azad University Kerman Branch, Iran,

IOI Code : AAHD-ABCHCD
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    The rapid adoption of Retrieval-Augmented Generation (RAG) has enabled large language models (LLMs) to incorporate external knowledge; however, existing RAG architectures predominantly rely on centralized indexing and data aggregation, which limits scalability and raises privacy concerns in distributed big data environments. This paper proposes an Adaptive Federated Retrieval-Augmented Generation (AFRAG) framework designed for decentralized big data systems where data remain locally stored across multiple nodes. The primary objective is to enable scalable, privacy-preserving, and high-quality knowledge retrieval and generation without transferring raw data. The proposed architecture integrates federated embedding learning, adaptive query routing, and privacy-aware aggregation mechanisms to dynamically select relevant nodes and optimize retrieval performance under communication and latency constraints. A distributed experimental setup is implemented to simulate heterogeneous data silos, and the framework is evaluated against centralized and baseline federated RAG approaches using standard retrieval and generation metrics, communication overhead, and latency measurements. Experimental results demonstrate that AFRAG improves retrieval relevance and answer accuracy while significantly reducing communication cost and preserving data locality. The findings indicate that adaptive routing combined with federated embedding optimization provides an effective solution for scalable and privacy-conscious LLM-driven analytics in large-scale distributed data ecosystems.

Keywords : federated learning retrieval augmented generation (rag) distributed big data systems adaptive query routing privacy preserving analytics large language models (llms)