Federated Graph Retrieval-Augmented Generation
GraphRAG usually assumes a centralized knowledge graph, which clashes with real-world settings where sensitive data is siloed across organizations that cannot share raw graphs due to privacy or ownership constraints. This paper proposes FedGRAG, a federated framework that routes queries to relevant clients using lightweight semantic prototypes and then reconstructs abstracted evidence subgraphs on a central server to enable cross-client reasoning. The problem is practically important because legal and competitive barriers prevent the global construction and indexing required by standard GraphRAG pipelines.
FedGRAG presents a modular and intuitively appealing architecture for distributed graph retrieval, and the ablation studies show that each proposed component contributes to end-to-end performance. However, the work overstates its privacy guarantees, offers only a vacuous theoretical bound for routing, and compares against centralized baselines that are adapted into straw-man federated variants, all of which substantially weaken its scientific contribution.
The ablation studies provide credible evidence that the proposed components are complementary. The authors show that the full FedGRAG consistently achieves the best performance, while removing any module degrades results across datasets. Additionally, the routing ablation demonstrates that the prototype-driven propagation strategy meaningfully improves client selection compared to random routing or direct matching.
The privacy justification is superficial. Although the paper describes controlled-disclosure evidence construction, the shared evidence package includes latent node embeddings $\mathbf{H}^q$, adjacency $\mathbf{A}^q$, and a local vocabulary $\mathcal{U}^q$, which collectively leak substantial semantic and structural information; there is no formal privacy definition, differential privacy budget, or empirical reconstruction attack to validate the claim that local graph memory cannot be shared directly. The theoretical guarantee for routing is also weak: Theorem 1 assumes the number of selected clients $B$ is at least the unknown number of evidence-bearing clients $|\mathcal{C}_{q}^{+}|$, and merely restates that zero hinge loss implies perfect ranking, which provides no practical coverage bound under realistic query conditions.
Baseline comparisons are further weakened by unfair adaptation. Centralized GraphRAG systems such as GraphRAG (MS) depend on global community detection and hierarchical summarization; forcing them to operate on isolated client shards with random routing strips away their core capability, so the reported gains partly reflect fixing a broken baseline rather than true architectural superiority.
The empirical gains are difficult to interpret because the centralized GraphRAG baselines are adapted in a way that cripples their core mechanisms. As originally designed, Microsoft GraphRAG constructs 'an entity knowledge graph from the source documents, then ... pregenerate community summaries for all groups of closely related entities' (Edge et al., 2024), which requires a unified corpus; the paper's own appendix confirms the federated adaptation merely performs local indexing combined with a random route strategy. This is not a competitive baseline but a straw man. Moreover, Table 1 reports only single-run point estimates without variances or significance tests, so the consistency of the reported improvements is unverified.
Reproducibility is hindered by missing implementation artifacts and incomplete model specifications. Although hyperparameters and training configurations are partially documented in Appendix B.4, the specific LLM used for final answer generation is only described as 'an external LLM' with no model name, version, or decoding parameters provided. The paper does not release code or data partitions, and Table 1 reports single-run metrics without standard deviations, leaving no way to assess experimental variance.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.