Federated Graph Retrieval-Augmented Generation

Xiaobao Wang; Renda Han; Zhiliang Xia; Meng Liu; Renxiang Guan; Di Jin

cs.LGOct 22, 2026

Federated Graph Retrieval-Augmented Generation

Authors: Xiaobao Wang, Renda Han, Zhiliang Xia, Meng Liu, Renxiang Guan, Di Jin

Organizations: College of Intelligence and Computing, Tianjin University, Tianjin, China · Henan University, Zhengzhou, China · National University of Defense Technology, Changsha, China

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) leverages centralized knowledge graphs to enhance large language models with structured knowledge retrieval. However, real-world knowledge is inherently distributed across multiple organizations rather than being centrally accessible. Privacy constraints further prohibit sharing entire local knowledge graphs, preventing GraphRAG from effectively retrieving and utilizing distributed knowledge. To this end, we propose Federated GraphRAG (FedGRAG), a unified framework for distributed graph retrieval and knowledge-enhanced generation over multiple clients with knowledge graphs. Specifically, we address two fundamental challenges in federated GraphRAG. First, the absence of a global knowledge graph makes unified retrieval impossible. To address this, we develop a Federated Knowledge Routing mechanism that efficiently identifies relevant clients through lightweight knowledge indices without exposing local graph structures. Second, local graph memory cannot be shared directly because it contains sensitive entities and knowledge information. We therefore propose a controlled-disclosure evidence construction method. It extracts retrieved evidence subgraphs and converts them into an evidence abstract, enabling evidence conversion and reconstruction on the server. Extensive experiments demonstrate that Federated GraphRAG consistently outperforms existing baselines in both retrieval and downstream generation tasks, establishing a new paradigm for deploying GraphRAG in federated environments.

Figures & tables

Figure 1: Motivation of Federated GraphRAG. Existing GraphRAG relies on centralized graph memories, while directly extending Federated RAG fails to capture semantic relationships among distributed graph memories.

Figure 2: The architecture of FedGRAG. This process consists of a pre-process and a query phase. During the pre-process phase, the client and server perform collaborative training to extract low-dimensional graph embeddings; the client then uploads prototypes to the server to construct a prototype graph. In the query phase, the server routes the query to the appropriate semantic cluster on the client. Subsequently, an evidence subgraph is constructed within the cluster, and the server aggregates these reconstructed evidence facts to generate retrieval results.

Models	HotpotQA				MuSiQue				2Wiki
Models	5 Clients
	R@2	R@5	EM	F1	R@2	R@5	EM	F1	R@2	R@5	EM	F1
GraphRAG	10.9	13.4	39.1	51.2	8.9	14.7	13.1	21.9	17.1	20.6	31.0	34.7
LightRAG	12.3	17.7	37.8	40.5	11.9	12.2	18.5	24.1	13.2	19.0	22.3	23.9
HippoRAG	9.6	21.0	28.7	31.1	12.3	13.1	20.2	22.6	12.9	22.4	33.9	37.3
SubgraphRAG	15.0	19.1	26.7	30.3	7.0	16.2	23.1	25.7	14.4	18.6	25.0	38.2

Table 1: Performance comparison with different state-of-the-art GraphRAG methods on three datasets.

Figure 3: Model performance of FedGRAG and its three variants.

Dataset	Method	R@2	Hit@2	R@5	Hit@5
HotpotQA	Random Routing	7.2	31.6	9.4	34.0
	Direct Matching	8.5	42.4	12.3	46.2
	Ours	19.4	66.7	31.7	71.4
MuSiQue	Random Routing	6.9	13.2	12.0	34.3
	Direct Matching	8.8	25.1	14.6	42.2
	Ours	14.7	68.6	23.9	73.2

Table 2: Client routing performance on three multi-hop QA datasets (%) under 10 clients.

Figure 4: Model performance under varying $K\in[1,20]$ .

Figure 5: Model performance on varying $L$ $\in$ [ $1$ , $10$ ] under three datasets.

Clients	Ours		FRAG		C-FedRAG
Clients	T (s)	C (KB)	T (s)	C (KB)	T (s)	C (KB)
1	24.8	50.3	130.3	213.0	37.1	66.4
2	51.3	100.6	270.2	256.4	68.5	134.9
3	78.6	150.9	300.4	375.5	105.3	201.5
4	110.1	201.1	418.3	420.3	137.0	274.6
5	140.9	256.4	566.5	458.5	174.5	355.3

Table 3: Communication overhead comparison under the HotpotQA datasets. “T” and “C” denote Time (s) and Cost (KB), respectively.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Dataset	# Query-Doc Pairs	# Documents	# KGs	# Entities	# Relations	# Triples
HotpotQA	20,000	204,822	20	1,930,362	967,218	6,393,342
MuSiQue	20,000	410,380	20	1,544,966	900,338	4,848,715
2Wiki	20,000	122,108	20	916,907	372,554	2,883,006
Total	60,000	737,310	60	4,392,235	2,240,110	14,125,063

Appendix

Table 4: Statistics of the datasets and knowledge graphs (KGs) used for training.

Graph Retrieval-Augmented Generation

Graph Retrieval-Augmented Generation connects language models to knowledge graphs to retrieve relational evidence for grounded, multi-hop answers.

See all 232 papers

Guiding Federated Graph Recommendation with LLM-encoded knowledge
Jun 13, 2026Thi Minh Chau Nguyen, Hien Trang Nguyen, Duc Anh Nguyen +3Federated Learning Aggregation Graph Representation Learning
Beyond Raw Context Transfer: Representation-based Federated Retrieval-Augmented Generation
Aug 26, 2026Can Peng, Yu Liu, Yingyu Yang +4Retrieval-Augmented Generation Federated Learning
University of Oxford, Oxford, United Kingdom · National University of Singapore, Singapore
FD-RAG: Federated Dual-System Retrieval-Augmented Generation
May 22, 2026Tianhao Gao, Kai Yang, Yiyang LiRetrieval-Augmented Generation Memory-Augmented Language Models
School of Computer Science and Technology, Tongji University

Jun 13, 2026J/K move · Enter open · S save

Guiding Federated Graph Recommendation with LLM-encoded knowledge

Thi Minh Chau Nguyen, Hien Trang Nguyen, Duc Anh Nguyen, Van Ho-Long+2

Federated Learning Aggregation Graph Representation Learning Non-IID Federated Learning Federated Graph Learning

Open paper AskPDF arXiv

Graph-based recommender systems are highly effective at extracting collaborative signals from user--item interactions, and federated learning (FL) allows these models to be trained while preserving user privacy. However, aggregating graph representations across distributed, non-IID clients remains a challenge; structural embeddings learned locally often misalign, and naive averaging fails to capture meaningful cross-client relationships. Most existing federated graph methods rely exclusively on structural aggregation, neglecting the rich, global semantic context available in large language models (LLMs). In this paper, we propose a novel framework that uses LLM-encoded knowledge to guide federated graph recommendation. Specifically, clients learn structural representations from local graphs while simultaneously summarizing their typical interaction patterns into compact semantic vectors via a frozen LLM. The central server then uses these LLM-encoded semantic signals to discover related preference patterns across clients, guiding the selective aggregation of their structural representations. This enables semantically informed cross-client collaboration without exposing raw data. Extensive experiments on standard benchmarks show that guiding structural alignment with LLM-encoded knowledge consistently improves recommendation accuracy over existing federated graph baselines.

Abstract

Graph Retrieval-Augmented Generation

Explore similar work

Guiding Federated Graph Recommendation with LLM-encoded knowledge

Beyond Raw Context Transfer: Representation-based Federated Retrieval-Augmented Generation

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