FD-RAG: Federated Dual-System Retrieval-Augmented Generation

A novel framework known as FD-RAG (Federated Dual-System Retrieval-Augmented Generation) has been introduced to overcome the challenges faced by traditional RAG systems in edge computing environments. Unlike standard RAG, which relies on centralized knowledge and significant computational resources, FD-RAG is tailored for decentralized settings where knowledge is dispersed across devices, sharing raw data is not feasible, and frequent LLM calls are costly. This framework separates lightweight memory access from LLM reasoning on demand. It develops semantic-aware adaptive hypergraphs from local data and compresses them into efficient QA memories. During inference, FD-RAG utilizes direct memory matching for well-covered queries and resorts to LLM reasoning only when necessary, ensuring traces of retrieved memories align with hypergraph-based evidence. Additionally, it consolidates anonymized memory statistics across devices to address knowledge fragmentation, promoting efficient and privacy-conscious RAG in edge scenarios.