acceptodds
Under review as a conference paper at ICLR 2027

Federated Graph Retrieval-Augmented Generation

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.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Deep Graph Learning, Graph Pattern Recognition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WSjfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedGRAG, a federated GraphRAG framework for retrieval and generation over knowledge graphs distributed across multiple clients. The method consists of two main components: (1) a federated knowledge routing mechanism that represents each client using multiple graph-aware prototypes and constructs a global prototype graph for query-dependent client selection, and (2) an evidence abstraction mechanism that retrieves local evidence subgraphs, encodes them into compact representations, and reconstructs them at the server for cross-client evidence integration and generation. Experiments on HotpotQA, MuSiQue, and 2Wiki show improvements over a range of centralized GraphRAG and federated RAG baselines.

Critical strengths
  1. Federated GraphRAG is an interesting and practically motivated extension of GraphRAG, particularly when graph knowledge is distributed across organizations and cannot simply be centralized.
  2. The combination of prototype-based client routing and local evidence abstraction provides a reasonably well-integrated pipeline from distributed graph indexing to cross-client evidence aggregation.
  3. Representing each client with multiple prototypes is a reasonable way to accommodate heterogeneous local knowledge rather than forcing each client into a single representation.
Critical weaknesses
  1. The paper repeatedly motivates the method using privacy and describes the evidence construction as “controlled disclosure.” However, the server receives query-dependent evidence information, including the vocabulary associated with retrieved entities/relations, and subsequently reconstructs evidence through a decoder. Not transmitting the entire local KG is not by itself sufficient to establish privacy preservation. The paper does not specify a formal threat model or provide leakage evaluations such as reconstruction, entity inference, membership inference, or privacy-utility analysis. I therefore view the current evidence as demonstrating limited/controlled disclosure rather than a demonstrated privacy guarantee.
  2. The experiments construct federated scenarios by distributing the knowledge from HotpotQA, MuSiQue, and 2Wiki across clients. This is a reasonable proof-of-concept, but the paper provides limited characterization of the resulting federation. In particular, it is unclear how client heterogeneity, client size imbalance, entity overlap, and cross-client evidence dependencies are controlled. These factors are central to the claimed federated setting and could substantially affect the routing results.
Questions for authors

Clarification

  1. What is the assumed adversary model for the server? Is the server trusted, honest-but-curious, or potentially malicious? What specific privacy notion does “controlled disclosure” provide?
  2. Could the authors explicitly enumerate all information transmitted from a client to the server during inference, including prototypes, client identifiers, and any provenance information?
  3. How accurately can the server-side decoder reconstruct the original evidence attributes or entities. If reconstruction is highly accurate, what prevents the transmitted representation from leaking the original local knowledge? Suggestions
  4. A formal privacy analysis, including reconstruction/entity-inference attacks and privacy-utility tradeoffs, would substantially strengthen the privacy-related claims.
  5. Larger federated experiments with substantially more clients would be useful for validating the scalability of the prototype graph and the claimed communication advantages.
1 comment

Thank you for recognizing the practical motivation of FedGRAG, our integrated multi-prototype pipeline, and the clarity of our presentation. Below we address your questions and suggestions concisely.

1. Server Model and Controlled Disclosure Scope We agree with your assessment that our framework demonstrates architectural controlled disclosure (Abstract and Section 3.1) rather than a formal cryptographic privacy guarantee, and we have calibrated broad privacy phrasing in the manuscript accordingly. Regarding your first question, the server in our setting operates as an honest-but-curious RAG coordinator rather than a malicious adversary or a fully trusted central repository. Controlled disclosure enforces two concrete data-minimization boundaries for collaborative RAG: first, clients never pool full local KGs, exposing under 0.05% of local triples per query; second, raw source documents remain strictly on clients, while the decoder reconstructs only concise facts needed to ground the LLM. Because RAG inherently requires surfacing query-relevant facts to prevent hallucination, formal membership or inversion attack benchmarks fall outside our collaborative RAG scope, though orthogonal cryptographic layers (such as FRAG or C-FedRAG) can wrap our vector transmissions when stricter security is needed.

2. Complete Enumeration of Transmitted Information We have added an explicit transmission summary in Section 3.3 covering both stages:

  • Offline indexing (once): Each client uploads only its client ID and 8 prototype vectors (64-dimensional K-means centroids), with zero entity names, edges, or source text.
  • Online inference (top-B clients): Selected clients upload their client ID, local 64D node embeddings, an unlabeled adjacency mask, and the subgraph word list, reusing the client ID for provenance without extra metadata.

3. Reconstruction Accuracy vs. Local Knowledge Protection On HotpotQA (10 clients), our server decoder achieves 91.8% entity F1 and 86.3% relation F1 at the word level on retrieved subgraphs (yielding 52.5 QA F1, close to ground-truth triples at 53.4). High fidelity on the active subgraph is intentional and essential for RAG to ground generation accurately. What prevents leaking a client's broader local knowledge is pre-encoding scope truncation and source-text stripping, rather than noisy decoding: per query, a client's retrieved subgraph averages only 0.04% of its local triples, so over 99.95% of local entities and relations never enter the uploaded embeddings or subgraph word list. Moreover, all raw documents stay local, and answer-aligned training (Equations 11 and 12) strips away verbose surrounding sentences from original passages, which also improves QA over exposing raw text (w/o ES in Figure 3).

4. Characterization of the Federated Benchmarks We partition knowledge across clients in a non-IID manner and have added a quantitative characterization of our 10-client federations in Appendix B.1, reporting max-to-min client triple ratio, Jensen-Shannon divergence (JSD) of relation distributions, cross-client entity overlap, and the share of test queries whose supporting facts span at least two clients:

Dataset (10 Clients)Size RatioRelation JSDEntity OverlapCross-Client Queries
HotpotQA1.840.3111.4%78.6%
MuSiQue2.170.388.9%86.3%
2Wiki1.690.2614.2%81.5%

Across all three benchmarks, most multi-hop queries require connecting facts across multiple clients despite low entity overlap and skewed local relation distributions, explaining why multi-prototype routing and our Evidence Interaction Graph outperform single-vector matching.

5. Scalability Across 10, 20, 30, and 50 Clients Because online routing only computes lightweight 64-dimensional inner products on the prototype graph and only the fixed top B=5 clients transmit evidence packages per query, online evidence communication does not grow with the total number of clients. Combining our 10- and 20-client HotpotQA benchmarks with extended 30- and 50-client evaluations (under fixed B=5) against the strongest federated baseline FedMosaic confirms this scalability:

HotpotQA ScaleFedMosaic (R@5 / F1)FedGRAG (R@5 / F1)
10 Clients27.3 / 47.631.7 / 52.5
20 Clients23.6 / 47.227.5 / 53.9
30 Clients21.2 / 45.425.8 / 51.4
50 Clients18.9 / 42.621.7 / 46.8

From 10 to 30 clients, although fixed B=5 narrows client coverage from 50% to 16.7%, FedGRAG peaks at 53.9 F1 (20 clients), holds 51.4 F1 at 30 clients, and widens its F1 lead over FedMosaic from 4.9 to 6.7 and 6.0 points as finer partitions yield cleaner local prototypes. Even at 50 clients (selecting only 10% of clients across heavily fragmented graphs), FedGRAG still outperforms FedMosaic by 2.8 R@5 and 4.2 F1, nearly matching FedMosaic's 20-client F1 (46.8 vs. 47.2).

Reviewer_6vsCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FedGRAG. The proposed method consists of three main components: federated knowledge routing, client-side evidence abstraction, and cross-client evidence fusion. It constructs a global semantic routing graph based on knowledge prototypes to select relevant clients and enables collaborative reasoning over distributed knowledge through local subgraph encoding and reconstruction. The authors conduct experiments on three datasets, HotpotQA, MuSiQue, and 2Wiki. The results demonstrate that the proposed method achieves certain advantages in retrieval quality and question-answering accuracy.

Critical strengths
  1. This paper introduces a federated GraphRAG framework.
  2. The proposed technical approach demonstrates a certain degree of novelty.
  3. The experimental results show promising performance improvements.
Critical weaknesses
  1. The mechanisms for cross-client knowledge representation alignment, relational information reconstruction, and evidence association are not clearly described.
  2. The paper does not evaluate potential privacy leakage risks arising from knowledge prototypes, adjacency structures, and reconstructed evidence.
  3. Some baselines employ relatively weak random routing strategies, while the theoretical guarantees rely on restrictive assumptions, making the evidence insufficient to fully support the paper's claimed advantages.
Questions for authors
  1. Do the graph encoders and projection networks across different clients share parameters? How does the proposed method ensure that knowledge prototypes from different clients and query embeddings reside in a unified and semantically consistent representation space?

  2. How does the local evidence abstraction mechanism accurately preserve relation types and graph structural information? Could the authors provide experiments evaluating evidence reconstruction accuracy and the impact of reconstruction errors on downstream generation performance?

  3. How are entity alignment and relation connections implemented in the cross-client evidence interaction graph? How does the proposed method handle entity ambiguity, conflicting relations, and incorrect associations across different clients?

  4. The proposed framework enables collaborative retrieval by uploading knowledge prototypes, local adjacency structures, and evidence representations. How can the authors demonstrate that these representations do not leak sensitive client knowledge? Could additional privacy attack experiments be provided?

  5. Why do the centralized GraphRAG baselines adopt random routing rather than more effective semantic routing strategies when adapted to the federated setting? Could the authors conduct additional comparisons under a unified client selection strategy to better isolate and validate the contributions of the proposed techniques?

1 comment

Thank you for recognizing the novelty and empirical gains of FedGRAG. Below we address your five questions concisely.

1. Parameter Sharing and Space Alignment All clients share the frozen MiniLM text encoder and the server-side evidence decoder, while each client trains its own local GNN and fusion MLP to capture non-IID graph structures. Client prototypes and queries stay in a unified semantic space because our local contrastive loss (Equation 12) directly pulls fused node embeddings toward shared query vectors via cosine similarity, while the shared decoder and global answer alignment (Equation 11) anchor all clients to a common latent space.

2. Relation Reconstruction and Error Impact Each selected client uploads node embeddings, a binary adjacency matrix, and a small query-specific vocabulary for its retrieved subgraph. The adjacency matrix preserves exact subgraph structure, while the local GNN encodes neighborhood relation context into node embeddings. For each connected node pair, the server decoder combines both endpoint embeddings to predict entity and relation words from the subgraph vocabulary (Equation 7). On HotpotQA (10 clients), we tested word-level reconstruction accuracy and how errors affect QA:

SettingEntity F1Relation F1QA EMQA F1
Ground-Truth Triples100.0100.047.653.4
FedGRAG91.886.346.852.5
FedGRAG (15% Word Error)78.972.643.949.6

Decoding over the small subgraph vocabulary achieves high word accuracy and stays within 0.9 QA F1 of noiseless ground-truth triples. Even with 15% added word errors, QA F1 drops by only 2.9 points because global answer alignment preserves core answer-bearing facts.

3. Cross-Client Linking and Conflict Handling In the Evidence Interaction Graph, nodes are reconstructed triples connected across clients when sharing normalized entity names or high embedding similarity. False links from same-name entities are largely avoided because we never align full KGs; only top query-relevant triples from selected clients enter this small graph with neighborhood-aware embeddings. During deduplication, we keep client provenance tags so the LLM can weigh conflicting multi-source claims in context, leaving dedicated conflict-resolution modules for future work.

4. Privacy Scope in Collaborative RAG Our core task is collaborative GraphRAG under privacy and knowledge ownership constraints where institutional data cannot be centralized, rather than defense against adversarial privacy attacks. Because LLMs need retrieved facts to prevent hallucination, FedGRAG operates through controlled disclosure (Abstract and Section 3.1), exposing only minimal query-relevant abstractions to the server coordinator rather than pooling local repositories. Offline, each client uploads only 8 cluster centroids; online, selected clients upload unlabeled masks and embeddings covering less than 0.05% of local triples, keeping all raw documents and over 99.9% of local KGs strictly on clients. Reconstructing concise triples also strips verbose source sentences and improves QA over sending raw text (w/o ES in Figure 3), placing adversarial attack benchmarks outside our RAG scope.

5. Unified Routing Fairness All federated RAG baselines in Table 1 already use their native routers; centralized GraphRAG baselines used random routing only because they have no built-in client router. On HotpotQA (10 clients), we tested LightRAG, HippoRAG2, and GFM-RAG under Direct Matching (mean-embedding cosine routing) and our FedGRAG Router:

MethodRandom (EM / F1)Direct (EM / F1)Ours Router (EM / F1)
LightRAG25.9 / 30.728.1 / 33.231.4 / 36.7
HippoRAG226.4 / 27.728.7 / 30.431.9 / 33.8
GFM-RAG41.2 / 46.842.7 / 48.444.5 / 50.1
FedGRAG41.9 / 47.644.2 / 49.946.8 / 52.5

Our router boosts every baseline, and under the exact same routed clients, FedGRAG still beats GFM-RAG by 2.4 F1, confirming the standalone value of local reconstruction and our Evidence Interaction Graph.

Reviewer_jmLQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies GraphRAG in a federated setting: knowledge graphs are distributed across multiple institutions and cannot be shared with one another, and the problem is how to collaboratively retrieve information and generate answers. The authors propose FedGRAG, where each client uses an attribute encoder plus a GNN to obtain node representations, generates knowledge prototypes via K-means, and uploads them. The server uses these prototypes to build a kNN routing graph and selects the Top-B clients through relevance propagation on the graph. The selected clients extract H-hop evidence subgraphs and upload node representations, adjacency matrices, and query-related vocabularies; the server then uses a shared decoder to reconstruct entity and relation attributes. The server builds an evidence interaction graph based on the reconstructed facts, traverses it, serializes it into context, and feeds it to an LLM to generate the answer. The experiments are mainly conducted and validated on QA datasets such as HotpotQA, MuSiQue, and 2Wiki.

Critical strengths
  • The framework is complete, covering routing, evidence extraction, and evidence combination.
  • The experiments cover three multi-hop QA datasets and different numbers of clients, along with ablation studies, hyperparameter sensitivity, and communication overhead analysis.
Critical weaknesses
  • The claim of novelty is overstated, and the lack of related work. For example, DGRAG and SCOUT-RAG are works of the same task, and federated knowledge graph completion is also a highly relevant direction.
  • The discussion of methodological relevance is also seriously lacking. The core of this paper is actually its privacy claim, but after proposing this setting, the paper does not provide a threat model, so its privacy guarantees are insufficient.
  • The experimental setup is highly idealized compared with the real world. Schemas vary considerably across different institutions, and the effectiveness of the method on highly heterogeneous graphs requires further discussion.
Questions for authors

NA

1 comment

Thank you for recognizing the completeness of our pipeline and experimental analysis. Below we address each of your three concerns directly.

1. Novelty Claim and Related Work While SCOUT-RAG and DGRAG also study retrieval over distributed private KGs, our claim in Section 1 (rooted in federated graph learning, Lines 056 to 057) holds at the paradigm level: FedGRAG is the first unified Federated GraphRAG framework built on collaborative client-server training and joint optimization (Figure 2 and Section 3.4), rather than training-free distributed inference. We have added a focused comparison in Section 2.2. SCOUT-RAG is a training-free four-agent prompting loop that scores domains via flat similarity and merges isolated per-domain text answers. Compressing each domain into local text early drops intermediate bridge entities needed for cross-silo hops, whereas FedGRAG learns a multi-prototype routing graph (Theorem 1) and explicitly reconnects cross-client paths on our Evidence Interaction Graph in a single LLM pass. DGRAG is a training-free edge-cloud offloading pipeline where edges upload plain-text subgraph summaries, local SLM gating traps 29.2% of out-of-domain queries on the edge due to confident hallucinations, and escalated queries send raw text chunks directly to the cloud. FedGRAG instead jointly trains encoder-decoder modules, uploads only K=8 cluster centroids offline, keeps raw chunks strictly local, and replaces flat matching with graph propagation (raising HotpotQA R@5 from 12.3% to 31.7% in Table 2). Finally, Federated KG Completion (FKGC) addresses an entirely different task: it predicts missing static triples offline using shared, pre-aligned anchor entities to populate local graphs. FedGRAG performs online query-driven multi-hop retrieval and generation without modifying local KGs or requiring pre-aligned entity IDs.

2. Threat Model and Methodological Relevance Our primary task is collaborative GraphRAG when institutional data cannot be centralized, rather than adversarial or cryptographic privacy defense. We have formalized an explicit controlled-disclosure threat model in Section 3.1: clients are mutually untrusted, and the server acts as an honest-but-curious RAG coordinator under purpose limitation. Because RAG requires grounding facts to prevent hallucination, the server is authorized to assemble only the minimal query-relevant facts for the active query, while full local knowledge graphs and raw document corpora remain strictly on the clients. Under this setting, encoding retrieved subgraphs into latent representations and reconstructing facts through the shared server decoder (instead of sending raw text or plaintext triples) serves two key methodological roles for RAG. First, joint optimization with our global answer alignment and local query alignment objectives (Equations 11 and 12) makes the decoder a task-aligned filter that extracts concise answer-predictive relations while stripping verbose, sensitive source sentences; as w/o ES in Figure 3 shows, feeding raw retrieved text directly to the LLM hurts EM and F1 by introducing distracting context. Second, the shared projection aligns non-IID client representations into a unified latent space prior to decoding, enabling reliable cross-client linking on the Evidence Interaction Graph (confirmed by w/o EI in Figure 3).

3. Schema Heterogeneity We used a uniform graph extraction pipeline in Section 4 (following GFM-RAG) to ensure a controlled, fair comparison across all 11 baselines without extractor variance. That said, FedGRAG does not rely on a shared global schema of entity types or relation IDs: free-text attributes and relations are embedded into a continuous semantic space via MiniLM, local structural variations are absorbed by client-specific GNNs, and institutional terms are decoded via query-specific vocabularies. Resolving severe ontology conflicts across highly heterogeneous schemas goes beyond our main scope here, and we have expanded Section 5 to discuss this limitation and future direction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.