KG-RAG: Graph-based Retrieval-Augmented Generation

• KG-RAG is a framework that uses knowledge graphs—structured as triples, subgraphs, or reasoning paths—to retrieve explicit and compositional evidence for answer generation.
• It employs diverse methods like path-based traversal, structural query alignment, and reinforcement of retrieval strategies to mitigate hallucination and improve multi-hop reasoning.
• Adaptations into multimodal, domain-specific, and interactive applications highlight KG-RAG's versatility in enhancing evidence granularity and answer accuracy.

Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) denotes a family of retrieval-augmented generation methods in which a LLM retrieves structured evidence from a knowledge graph rather than relying only on unstructured passages or latent parametric memory. In the formulations used across the literature, the external memory is typically a graph of triples such as $G=\{t=(e,r,e')\}$ or $G \subseteq E \times R \times E$, and the retrieved context is a set of triples, subgraphs, walks, or reasoning paths that conditions answer generation. The paradigm is motivated by the limits of chunk-based dense retrieval for knowledge-intensive tasks, especially hallucination, coarse evidence granularity, and difficulty with multi-hop or compositional questions (Sanmartin, 2024, Zhou et al., 7 Apr 2025).

1. Formal foundations and relation to standard RAG

A recurrent formalization treats the knowledge graph as a directed multirelational structure whose atomic unit is the fact triple $\langle h,r,t\rangle$. In the original KG-RAG pipeline, the graph is defined as $$G=\{t=(e,r,e') \text{ where } e,e'\in E,\, r\in R\}$$, and information retrieval is written as $\text{IR}=\arg\max_{w_z} P(w_z\mid q,G)$, where $w_z$ is a path in the graph. For multi-hop KGQA, answer selection is further expressed as $$a^*=\arg\max_{a\in A}\sum_z P(w_z\mid q,G)\cdot P(a\mid w_z)$$, making the retrieved path rather than the raw chunk the central evidential unit (Sanmartin, 2024).

The conceptual contrast with standard text RAG is explicit in multiple works. Standard semantic RAG retrieves passages or chunks that are similar to the query, but those chunks are often isolated, redundant, or too coarse for fact-level composition. KG-RAG instead retrieves structured evidence—triples, neighborhoods, or reasoning chains—and is presented as a way to search for “things, not strings.” The claim is not that graphs eliminate generation, but that they constrain it through editable, inspectable, and semantically explicit external memory (Sanmartin, 2024).

This family view is important because the literature does not define KG-RAG as a single architecture. Some systems retrieve graph paths, some retrieve induced subgraphs, some verbalize walks into text, and some fuse graph retrieval with dense vector retrieval. A plausible implication is that KG-RAG is best understood as a design space whose invariant is the use of knowledge-graph structure as the primary retrieval substrate, while the choice of retrieval unit, traversal policy, and generation interface varies by task.

2. Knowledge graph construction and representation

Graph construction is a decisive stage because KG-RAG systems are only as good as the graph they retrieve from. In the original KG-RAG pipeline, raw text is converted into a homogeneous knowledge graph by an LLM in a few-shot setting that extracts triples from each chunk $T$, modeled as a conditional generation problem $P(T'|T)=\prod_i P(t_i|T)$. That work also introduces triple hypernodes, allowing nested facts and contextualized relations to be encoded recursively rather than flattened into only atomic triples (Sanmartin, 2024).

Other systems keep an explicit link between graph facts and source text. KG$^2$RAG represents a document-derived graph as $G \subseteq E \times R \times E$0, where each extracted triplet is attached to the chunk $G \subseteq E \times R \times E$1 from which it came. This chunk-linked representation supports graph-guided expansion while preserving the provenance needed to reconstruct coherent evidence paragraphs at generation time (Zhu et al., 8 Feb 2025).

A more agentic construction model appears in RAGA, which treats KG building as a full lifecycle process rather than a one-pass extraction job. RAGA exposes 16 atomic tools spanning reading, search, entity and relation CRUD operations, review, and progress tracking, and it couples graph construction to evidence-anchored verification so that each knowledge entry carries provenance back to source text. In this formulation, KG quality, entity disambiguation, and auditability are construction-time concerns rather than downstream patches (Han et al., 16 May 2026).

These representation choices matter because they determine what retrieval can recover. Hypernodes favor richer event structure, chunk-linked graphs favor source-grounded reconstruction, and tool-based lifecycle management favors auditable and incrementally maintainable graph memory. This suggests that “the graph” in KG-RAG is not merely a storage backend; it is the operative representation of evidence granularity.

3. Retrieval algorithms and reasoning regimes

The retrieval layer is the main locus of methodological diversity. In the original KG-RAG system, Chain of Explorations (CoE) performs reasoning-guided sequential traversal. It begins from a few-shot plan generated from the query, alternates between node exploration and relation exploration, searches top-$G \subseteq E \times R \times E$2 candidate nodes in a vector database, retrieves adjacent relations via Cypher, reranks them with dense similarity, and repeatedly asks the LLM to select the most relevant continuation. An evaluation function checks alignment with the original query and may refine the plan and restart if traversal drifts (Sanmartin, 2024).

SimGRAG replaces free-form traversal with structural query alignment. It first converts the natural-language question into a pattern graph, then retrieves isomorphic candidate subgraphs and ranks them by a graph semantic distance (GSD) that sums embedding-space distances over matched nodes and relations. The reported optimized retrieval algorithm identifies the top-$G \subseteq E \times R \times E$3 subgraphs within 1 second on a 10-million-scale KG, positioning structural matching rather than only local exploration as the retrieval primitive (Cai et al., 2024).

Walk&Retrieve adopts a different simplification strategy. It precomputes BFS or random walks around entities, verbalizes each walk into natural language, indexes both entity-level and walk-level embeddings, and at inference retrieves the most similar entities and verbalized walks before making a single LLM call. The method is explicitly zero-shot and is designed to adapt to KG updates without retraining retrievers or generators (Böckling et al., 22 May 2025).

AnchorRAG addresses an assumption common in earlier KG-RAG pipelines: the existence of a correct anchor entity. Its predictor agent identifies candidate anchors from the query, retriever agents explore those anchors in parallel, and a supervisor agent determines whether the accumulated paths are sufficient or whether retrieval should continue. The reported implementation uses $G \subseteq E \times R \times E$4 retrieval agents, maximum depth $G \subseteq E \times R \times E$5, and width $G \subseteq E \times R \times E$6, turning anchor uncertainty into a multi-agent search problem rather than a single-linking bottleneck (Xu et al., 1 Sep 2025).

KERAG shifts emphasis from exact path finding to recall-oriented neighborhood acquisition. Its planning quadruplet $G \subseteq E \times R \times E$7 determines domain, topic entity, filtered relations, and hop count; retrieval then returns a broader neighborhood scope that may be needed for answering, after which filtering and summarization reduce noise and concentrate reasoning in the summarization stage. On Head2Tail, the reported retrieval recall is 0.952 for KERAG versus 0.844 for ToG, illustrating a deliberate preference for coverage before compression (Sun et al., 5 Sep 2025).

Closed-loop variants further alter the retrieval regime. Know$G \subseteq E \times R \times E$8-RAG uses KG embeddings to estimate whether retrieval is needed, augments queries with linked, local-neighbor, and predicted entities, and filters generated references by semantic relevance and KG consistency. MetaKGRAG adds a Perceive–Evaluate–Adjust cycle that diagnoses completeness deficiency and relevance deficiency, then performs trajectory-connected correction from a pivot point instead of restarting blindly. KG-IRAG, in turn, couples retrieval to iterative temporal reasoning by letting one LLM propose a starting plan and another judge whether the currently retrieved time-location evidence is sufficient, expanding the search window until a boundary decision can be made (Liu et al., 19 May 2025, Yuan et al., 13 Aug 2025, Yang et al., 18 Mar 2025).

4. Applications and modality extensions

Although early work focused on Freebase-style KGQA, later systems extend KG-RAG into domain-specific, multimodal, and interactive settings. DO-RAG targets technical closed-domain QA in database and electrical engineering corpora by combining multi-level KG construction with semantic vector retrieval and a grounded refinement stage. In the reported experiments, its SunDB.AI implementation improves composite scores by up to 33.38% over external baselines, and with KG integration Contextual Recall reaches 1.000 for both DeepSeek-R1 and DeepSeek-V3 (Opoku et al., 17 May 2025).

A distinct line of work adapts KG-RAG to GUI agents. The GUI-oriented KG-RAG framework transforms UI Transition Graphs into a searchable knowledge graph/vector memory of intent–trajectory pairs, then retrieves trajectories as actionable navigation evidence during online decision-making. On DroidTask with GPT-4 as backend, the reported results are 75.8% Success Rate, 84.55% Decision Accuracy, and 4.10 Average Steps; the same system also reports transfer gains of +40% SR on Weibo-web and +20% on QQ Music-desktop (Guan et al., 30 Aug 2025).

Multimodal extensions treat the KG itself as a multimodal object. mKG-RAG constructs document-level multimodal KGs by aligning textual subgraphs with scene-graph-style visual relations through MLLM-based vision-text matching, then performs dual-stage retrieval over candidate documents and local graph evidence. Its strongest reported configuration, mKG-RAG* with QM-Retriever, achieves 36.3% overall on E-VQA and 40.5% overall on InfoSeek (Yuan et al., 7 Aug 2025).

MegaRAG generalizes this logic to long, visually rich corpora such as textbooks, slide decks, and reports. It defines each page as $G \subseteq E \times R \times E$9, builds an initial multimodal knowledge graph, refines it with page-specific subgraphs, retrieves both graph and page-image evidence in a unified embedding space, and performs two-stage answer generation to reduce modality bias. Reported results include 64.85% on SlideVQA (2k) and an average overall win rate of 89.5% on the paper’s multimodal global QA benchmark (Hsiao et al., 26 Nov 2025).

These extensions show that KG-RAG is not confined to symbolic text QA. The same core idea—retrieve structured external memory, then condition generation on that structure—has been adapted to expert technical corpora, multimodal VQA, long-document reasoning, and interactive software navigation. This suggests that the central abstraction is portable across modalities as long as the retrieval graph can be constructed and queried effectively.

5. Evaluation practices and empirical behavior

Empirical results in KG-RAG are heterogeneous rather than uniformly favorable. The prototype KG-RAG evaluation on ComplexWebQuestions v1.1 reported lower answer correctness than an embedding-based baseline—19 EM, 25 F1, and 32 accuracy for KG-RAG versus 28 EM, 37 F1, and 46 accuracy for Embedding-RAG—but a lower hallucination count, dropping from 30 to 15. The authors further note that CoE typically takes about 4–5 steps to reach answer nodes and that average response time is about 30 seconds because multiple Azure API calls are required (Sanmartin, 2024).

A central methodological concern is graph incompleteness. Under controlled deletion experiments on WebQSP and CWQ, the evaluation study on incomplete KGs shows that all tested KG-RAG methods—RoG, ToG, and G-Retriever—degrade as triples are removed, with the more severe reasoning path disruption causing larger drops than random deletion. For RoG under 20% random deletion, the reported relative accuracy drop is about 5.99% on WebQSP and 7.41% on CWQ; yet even incomplete KGs still outperform the corresponding “w/o retrieval” settings, indicating that structured retrieval remains useful despite missing facts (Zhou et al., 7 Apr 2025).

Evaluation methodology itself is an active research topic. A KG-based evaluation framework inspired by RAGAS converts input and context into graphs, adds similarity edges between matched entities, and then scores outputs with two graph-native procedures: multi-hop semantic matching via Dijkstra reachability under a path-cost threshold, and community-based semantic overlap via Louvain clustering. The reported findings are moderate to high correlation with both RAGAS and human judgments for factual correctness, faithfulness, and answer relevancy, but notably weaker correlation for context relevancy (Dong et al., 2 Oct 2025).

Benchmark leaders are also task-specific. SimGRAG reports MetaQA Hits@1 of 98.0, 98.4, and 97.8 on 1-hop, 2-hop, and 3-hop settings, together with 86.8 accuracy on FactKG. AnchorRAG reports, with Qwen-Plus, 73.3 Hit@1 / 56.4 accuracy on WebQSP, 62.7 / 56.0 on GrailQA, 47.0 / 47.0 on CWQ, and 60.9 / 50.5 on WebQuestions, alongside average improvements of 10% Hit@1 and 11.2% Accuracy over the strongest baseline. KERAG reports 0.732 accuracy, 0.202 hallucination, 0.066 miss, and 0.529 truthfulness on CRAG, and 0.908 accuracy, 0.049 hallucination, 0.043 miss, and 0.860 truthfulness on Head2Tail (Cai et al., 2024, Xu et al., 1 Sep 2025, Sun et al., 5 Sep 2025).

A common misconception is that knowledge graphs automatically improve all QA metrics. The published results are more selective: graph grounding often helps hallucination control, evidence coherence, multi-hop reasoning, or domain transfer, but some graph-centric systems underperform strong embedding baselines on exact-match accuracy, and many gains are benchmark-, modality-, or task-dependent.

6. Limitations, robustness, safety, and ongoing directions

Several limitations recur across the literature. Graph construction by LLM extraction can miss triples or extract incorrect ones, and such errors propagate into retrieval. Coverage is often constrained by cost, especially when graph construction is performed over only a subset of available documents or snippets. Retrieval latency can also be substantial; the original KG-RAG prototype reports an average response time of about 30 seconds. At the evaluation level, incompleteness is only one failure mode: real KGs also contain noise and incorrect triples, and there is no universally standardized evaluation protocol for KG-RAG (Sanmartin, 2024, Zhou et al., 7 Apr 2025).

Reliability problems extend beyond raw accuracy to confidence estimation. Ca2KG treats prompt variants as causal interventions on retrieval-dependent reasoning and rescales predictions through a Causal Calibration Index. On GPT-3.5, the reported calibration results include MetaQA 1-hop Acc 0.876, ECE 0.067, BS 0.078, and AUC 0.947, with similarly improved calibration on MetaQA 3-hop and WebQSP settings. The point is not merely higher confidence, but confidence that better tracks correctness under retrieval uncertainty (Ren et al., 14 Jan 2026).

Explainability has become a parallel concern. XGRAG augments a GraphRAG pipeline with graph-native perturbation, removing or modifying nodes and edges and measuring the semantic change in the generated answer. Across NarrativeQA, FairyTaleQA, and TriviaQA, the reported improvement in explanation quality over RAG-Ex is 14.81% by F1 alignment, and the resulting importance rankings show positive correlation with degree centrality and PageRank (Li et al., 27 Apr 2026).

Security work shows that structure does not imply robustness. Under a practical insertion-only poisoning attack, carefully chosen perturbation triples can complete misleading inference chains and materially damage KG-RAG performance. For RoG on WebQSP, the reported clean-to-attack degradation is from F1 70.34 to 38.06 and from EM 45.76 to 8.85, while the analysis identifies retrieval rather than generation as the main vulnerability surface (Zhao et al., 9 Jul 2025).

At the same time, newer systems treat the KG as an adaptive object rather than a fixed database. EvoRAG uses feedback-driven backpropagation to assign response-level feedback to path utility and then to triplet-level contribution scores, refining retrieval priorities and evolving the graph through reinforcement, suppression, and shortcut relation creation. Across RGB, MultiHop, and HotpotQA, the reported gains over state-of-the-art KG-RAG frameworks are +7.34% ACC, +9.84% EM, and +7.29% F1 (Fu et al., 17 Apr 2026).

This suggests that mature KG-RAG systems will require more than improved retrieval heuristics. They will also require graph construction with provenance, robustness to incompleteness and poisoning, path-aware refinement, calibration under uncertain evidence, and explanation mechanisms that are native to graph structure rather than borrowed from text-only RAG.