Retrieval-Augmented Generation with Graphs (GraphRAG)

• GraphRAG captures relational knowledge through graph traversal, graph learning, and network analysis, overcoming semantic or lexical retrieval that can miss multi-hop connections and complex geometry.Graph structure can guide retrieval along relational paths, while specialized graph or geometric encoders preserve structural nuances that verbalizing neighborhoods cannot express efficiently.Source: §1 · pp. 2–3
• GraphRAG design must address three differences from conventional RAG: diverse-formatted information, interdependent information, and domain-specific information.These differences affect retrieval, organization, generation, and the transferability of graph encoders across domains and tasks.Source: §1 · pp. 2–3
• GraphRAG supports interconnected retrieval, structural organization, and richer generation by exploiting edges for multi-hop traversal, graph pruning, and structural signals such as positional encoding.Unlike independently indexed document chunks, graph-connected content can support tasks requiring multi-hop reasoning and long-form planning.Source: §1 · p. 3
• The survey addresses a fragmented and knowledge- or document-graph-centered research landscape by unifying GraphRAG globally while specializing its designs across domains.It reviews fundamental applications, graph construction methods and relational rationales, techniques for each domain, benchmark datasets, and tool resources.Source: §1 · pp. 3–4
• The proposed holistic GraphRAG framework comprises five components: query processor, retriever, organizer, generator, and graph data source.The survey reviews representative techniques within each component.Source: §1 · p. 3
• The survey categorizes GraphRAG designs into 10 application domains and reviews each domain’s applications, graph construction methods, component-specific techniques, benchmark datasets, and tools.It also highlights current research challenges and future opportunities for advancing GraphRAG.Source: §1 · pp. 3–4

• The section introduces a holistic framework of GraphRAG based on existing literature.Source: §2 · p. 5
• It establishes the basic problem setting for the framework.Source: §2 · p. 5
• It defines notation used throughout the framework.Source: §2 · p. 5

2.1 Problem Setting and Notations

• GraphRAG transforms query Q into answer A by preprocessing Q, retrieving content C from graph G, refining it into Ĉ, and generating A.The pipeline is defined as Q̂ = ΩProcessor(Q), C = ΩRetriever(Q̂, G), Ĉ = ΩOrganizer(Q̂, C), and A = ΩGenerator(Q̂, Ĉ).Source: §2.1 · p. 5
• The framework comprises a query processor, graph data source, retriever, organizer, and generator.The graph data source G stores information in graph-structured format, while the other components perform query preprocessing, retrieval, organization, and answer generation.Source: §2.1 · p. 5
• GraphRAG requires dedicated designs because graph-structured data encapsulates relational information unlike sequential text and grid-structured images.These designs must handle graph-structured inputs and outputs and support graph-based operations.Source: §2.1 · p. 5
• For retrieval, GraphRAG uses graph traversal methods such as entity linking and BFS/DFS and graph-based encoders such as GNNs, rather than conventional sparse/dense encoder search [196] [468].The graph-based encoders produce embeddings for retrieval.Source: §2.1 · p. 5

2.2 Task Applications and Example Query Q

• GraphRAG queries may use text, such as “What is the Capital of China?” for knowledge-graph question answering [272].Source: §2.2 · p. 5
• Queries can also use domain-specific formats, including SMILES strings for molecular graphs [132].Source: §2.2 · p. 5
• Some applications combine formats, such as a scene graph with a text instruction [147].Table 1 summarizes common GraphRAG task applications, exemplary queries, and representative references by domain.Source: §2.2 · p. 5, Table 1 · p. 6

2.3 Query Processor ΩProcessor

• GraphRAG query processing addresses the mismatch between text-formatted queries and graph-structured data through five complementary techniques.These techniques extract entities and relations, structure queries, decompose complex tasks, and expand underspecified queries.Source: §2.3 · p. 7

2.3.1 Name Entity Recognition

• Entity recognition identifies graph-grounded query entities as retrieval seeds and increasingly captures node types and structures to reduce cascading errors.Deep-learning methods such as EntityLinker and LLM-based extraction [186] support this process, which guides subsequent retrieval and generation [443] [347].Source: §2.3.1 · pp. 7–8

2.3.2 Relational Extraction

• Relational extraction constructs graph data through triplets and matches query relations to graph edges, guiding neighborhood selection and traversal direction.For example, identifying “capital of” enables corresponding-edge searches through vector similarity [119] [200] [272] [273].Source: §2.3.2 · p. 8

2.3.3 Query Structuration

• Query structuration converts natural-language questions into graph query languages such as Cypher, GraphQL, or SPARQL and can decompose complex queries into structured operations.These operations include node retrieval, feature fetching, neighbor checks, and degree assessment, improving querying precision and adaptability [186].Source: §2.3.3 · p. 8

2.3.4 Query Decomposition

• Query decomposition splits complex questions into logically connected subqueries whose results are aggregated, while question graphs represent subqueries as triplets for multistep retrieval and generation [316].This supports tasks requiring multistep reasoning and planning [248] [316] [355] [372].Source: §2.3.4 · p. 8

2.3.5 Query Expansion

• GraphRAG query expansion enriches ambiguous, brief, or uncertain queries with structured relations rather than relying only on textual similarity.Methods leverage neighboring nodes of mentioned entities [459] or convert queries into template-based subqueries [406].Source: §2.3.5 · p. 8

2.4 Retriever ΩRetriever

• GraphRAG retrievers retrieve relevant content from external graph sources after query processing, addressing workflows and structural signals that conventional text- and image-oriented retrievers overlook.GraphRAG workflows include Text-in, Text-out; Text-in, Graph-out; Graph-in, Text-out; and Graph-in, Graph-out processes.Source: §2.4 · p. 9

2.4.1 Heuristic-based Retriever

• Heuristic-based retrievers use entity linking, relational matching, graph traversal, graph kernels, and domain expertise to efficiently retrieve graph information, while fixed rules limit generalization.Entity and relation matching can resolve subtle ambiguities such as Byte versus Bit, while traversal methods mitigate neighborhood overload through adaptive filtering, pruning, or predefined rules.Source: §2.4.1 · pp. 9–10

2.4.2 Learning-based Retriever

• Learning-based retrievers embed nodes, edges, and graphs using shallow or deep graph encoders, enabling similarity search that captures proximity, structural roles, features, and task-relevant relations.Shallow methods preserve proximity or role information, whereas GNNs combine message passing with feature transformation and support inductive retrieval as graph content evolves.Source: §2.4.2 · pp. 11–12

2.4.3 Advanced Retrieval Strategies

• Advanced retrieval strategies address complex, multi-aspect, and multi-hop queries through integrated, iterative, and adaptive retrieval mechanisms.Integrated retrieval combines symbolic and neural signals; iterative retrieval expands evidence across dependencies; adaptive retrieval adjusts reasoning depth, although graph-based knowledge-conflict resolution remains open.Source: §2.4.3 · pp. 12–13

2.5 Organizer

• The organizer post-processes retrieved entities, relations, triplets, paths, or subgraphs together with the processed query to improve downstream generation.Its role is to refine retrieved content so it better fits the generator’s needs.Source: §2.5 · p. 13

2.5.1 Graph Pruning

• Graph pruning reduces noisy or oversized retrieved subgraphs while preserving task-relevant information, using semantic, syntactic, structural, or dynamic criteria.Methods score query relevance, analyze syntax, exploit PageRank, or remove nodes during training based on attention.Source: §2.5.1 · p. 14

2.5.2 Reranker

• Fine-grained reranking reorders retrieved triples, paths, or passages because information position and presentation order can affect LLM generation.Approaches use cross-encoders, reranker models, GNNs, or temporal ordering to prioritize context content.Source: §2.5.2 · p. 14

2.5.3 Graph Augmentation

• Graph augmentation enriches retrieved graphs by adding nodes and edges, enhancing node and edge features, or applying graph-level perturbations for robustness.Structure augmentation can connect queries or extracted phrases to graphs, while feature augmentation summarizes or expands sparse information.Source: §2.5.3 · p. 15

2.5.4 Verbalizing

• Verbalization converts retrieved triples, paths, or graphs into LLM-consumable language through rule-based tuples and templates or model-based graph-to-text and summarization.Graph-to-text methods preserve graph information, whereas summarization methods condense query-relevant knowledge and may alternate reasoning with summarization.Source: §2.5.4 · pp. 15–16

2.6 Generator

• GraphRAG uses three generator types: discriminative models for graph tasks, LLMs for text-based tasks, and generative models for producing new graphs.The generator is selected according to whether the task involves discrimination, natural-language generation, or graph generation.Source: §2.6 · p. 16

2.6.1 Discrimination-based Generator

• Discriminative generators apply GNNs and Graph Transformers to node, edge, and graph classification or regression, with architectures chosen by graph type and task.GCN [206], GraphSAGE [138], and GAT [402] target homogeneous graphs; RGCN and HAN target heterogeneous graphs; HGNN [111] and Hyper-Attention target hypergraphs, while graph transformers [296] [354] capture global dependencies.Source: §2.6.1 · p. 16

2.6.2 LLM-based Generator

• LLM-based generators address graph-to-text generation by verbalizing graph information or integrating graph representations into LLMs, because sequential LLMs may lose structure when graphs are converted to text.Embedding-fusion projects graph embeddings into the text space and injects them at different LLM layers, with either joint fine-tuning or a fixed LLM.Source: §2.6.2 · p. 16
• Positional embedding-fusion preserves retrieved graph structure by encoding node positions or pairwise information as inputs to LLMs.GIMLET [549] uses generalized position embeddings for graph structures and textual instructions, while LINKGPT [148] uses pairwise encoding from LPFormer.Source: §2.6.2 · p. 16

2.6.3 Graph-based Generator

• Graph-based generators support structure-sensitive scientific generation through transformer, Graph VAE, and diffusion architectures, while query-conditioned GNNs generate relevant KG answers.IRDIF uses SE(3)-equivariant diffusion with EGNNs to maintain molecular geometric consistency under rotation, translation, and reflection; RetMol supports varied encoder, decoder, generative-model, and molecule-representation choices, and KG methods use GNNs [110].Source: §2.6.3 · p. 17

Motivation

• Carefully curated graph data sources are essential because even strong query processing, retrieval, organization, and generation may not yield optimal GraphRAG performance.This motivates a data-centric perspective that emphasizes data quality and relevance alongside model design.Source: §2.7 · p. 17

Graph Construction

• Graphs can be constructed explicitly from predefined relationships or implicitly from derived connections such as word co-occurrence and tabular feature interactions.Examples include molecule, knowledge, citation, and recommendation graphs for explicit construction, and semantic or feature correlations for implicit construction.Source: §2.7 · p. 17

Graph Representation

• Graph data can be represented as adjacency matrices, edge lists, adjacency lists, node sequences, or natural-language descriptions.Node-sequence serialization may be irreversible, although Eulerian paths [552] and BFS/DFS for trees can preserve the original structure.Source: §2.7 · p. 17
• Basic graph representations require expansion for multi-relational attributed graphs, such as A ∈R|V|×|V|×|R| with relation-specific edge weights.Here, Ai,j,r denotes the weight of the edge between nodes i and j under relation r.Source: §2.7 · p. 18
• The ordering of graph descriptions significantly affects LLM comprehension of graph structure and performance across tasks [122].Therefore, graph representation should be selected according to task-specific requirements.Source: §2.7 · p. 18

3 Knowledge Graph

• Knowledge graphs encode entities, relationships, paths, and subgraphs in structured databases that may represent broad general knowledge or specialized domains.Their contents support downstream tasks across question answering, commonsense reasoning, fact-checking, recommendation, drug discovery, healthcare, and fraud detection.Source: §3 · p. 18

3.1 Application Tasks

• GraphRAG applies knowledge graphs to question answering, fact-checking, knowledge graph completion, and cybersecurity analysis by retrieving relational evidence for inference or validation.It retrieves supporting context for generated answers, facts and structures that support or refute claims, structural knowledge around incomplete triplets, and information about attack vectors and mitigation strategies.Source: §3.1 · p. 18

3.2 Knowledge Graph Construction

• Knowledge graphs are constructed manually, with rules and extractors, or with LLMs that identify, link, summarize, and organize entities and relations from documents.Examples include WikiData and UMLS for manual construction, ConceptNet, Freebase, and REANO for rule-based construction, and CuriousLLM, Graph-RAG, and AutoKG for LLM-based construction.Source: §3.2 · pp. 18–19

3.3 Retriever

• KG retrieval generally identifies query-relevant seed entities and then retrieves facts or entities using traversal, subgraph, rule, GNN, similarity, relation, fusion, or agent-based methods.Retrieval designs trade off the kinds, efficiency, and amount of retrieved facts, with methods ranging from path and neighborhood extraction to learned scoring, query planning, and tool-based search.Source: §3.3 · pp. 19–20

3.4 Organizer

• Organizers format retrieved graph information as tuples, text, code-like structures, or embeddings, and may re-rank it to improve the context supplied to the generator.Common inputs include triples, paths, subgraphs, entity text, probabilities, and historical reasoning programs; ordering can use relevance, recency, logical sequence, time, or task-specific scores.Source: §3.4 · p. 22

3.5 Generator

• Most GraphRAG systems use LLMs to generate responses from the query and organized context, while other systems use GNNs, masked language models, entity scoring, or triple classification.LLM generators commonly use models such as ChatGPT, Gemini, Mistral, and Gemma, with LoRA-based fine-tuning sometimes applied to open-source models.Source: §3.5 · p. 23

3.6 Resources and Tools

• The section catalogs widely used KG data resources, including Freebase, ConceptNet, and WikiData, alongside Graph RAG and LangChain as supporting tools and implementations.Freebase provides general facts, ConceptNet links words and ideas semantically, WikiData is crowdsourced, Graph RAG is an open-source implementation, and LangChain supports LLM applications including KG-based RAG.Source: §3.6.1 · p. 23, §3.6.2 · pp. 23–24

4 Document Graph

• Document graphs capture heterogeneous relational structure across documents and their components, enabling GraphRAG to recover indirectly connected evidence for tasks such as multi-hop question answering [90].They may represent hyperlinks, citations, sentence relations, and entity connections, providing semantic and syntactic context.Source: §4 · p. 24

4.1 Application Tasks

• Document graphs support multi-document summarization, text generation, document retrieval, classification, and question answering by compressing corpora, retrieving related documents or chunks, leveraging label homophily, and enabling long-range and multi-hop reasoning.Source: §4.1 · pp. 24–25

4.2 Document Graph Construction

• Document graphs are constructed explicitly from existing links or implicitly from semantic and syntactic structure, with nodes ranging from words to topics and heterogeneous, hierarchical, or dynamic edges.Implicit edges include word-word, word-sentence, sentence-sentence, sentence-document, and document-document relations based on parsing, similarity, references, clustering, and embeddings.Source: §4.2 · pp. 25–26

4.3 Retriever

• Document-graph retrieval uses pre-retrieval, graph similarity, iterative, and topology-based strategies to reduce construction cost and follow indirect or structurally similar evidence paths [392] [278] [10] [429].Pre-retrieval builds graphs only from query-relevant information, while iterative retrieval supports multi-step question answering.Source: §4.3 · p. 26

4.4 Organizer

• Document-graph organizers mainly prune irrelevant subgraphs and rerank retrieved passages, using clustering, path structure, dynamic node dropping, similarity thresholds, communities, and GNNs [539] [229] [98] [498] [90] [239].These operations reduce computation and arrange graph-expanded evidence for generation.Source: §4.4 · pp. 26–27

4.5 Generator

• Generators vary with input format: GNNs and Graph Transformers process graphs, RNNs and LLMs process text, and integrated models combine graph encoders with language generators [320] [392] [491] [300] [313] [90].LLM-based approaches often serialize retrieved subgraphs into text, while graph-informed self-attention embeds graph structure directly into generation [237].Source: §4.5 · p. 27

4.6 Resources and Tools

• Document GraphRAG relies on NLP, entity-linking, extraction, graph-processing, orchestration, and database tools, including CoreNLP, spaCy, BLINK, OpenIE, GraphRAG [98], LangChain, Neo4j, LlamaIndex [253], and Haystack.These tools support graph extraction, parsing, entity linking, community summarization, graph storage, querying, and LLM application pipelines.Source: §4.6 · p. 27

5 Scientific Graph

• GraphRAG retrieves scientific graphs to improve molecule generation and medical question answering by reducing invalid structures, hallucinations, and search spaces while preserving scientific validity.Scientific graphs primarily encompass molecular and medical graphs, common in drug discovery and biomedicine [265] [484] [255] [78] [185] [443].Source: §5 · pp. 28–29

5.1 Application Tasks

• Scientific GraphRAG supports molecule generation, molecular property prediction, and domain-specific question answering through retrieved molecular structures, similar-molecule demonstrations, or knowledge graphs.Molecule generation and property prediction target drug discovery, while scientific QA uses literature and databases converted into knowledge graphs [265] [484] [185] [443] [225] [175] [449].Source: §5.1 · p. 29

5.2 Scientific Graph Construction

• Scientific graphs are built from public and private datasets and, for chemistry, represent molecules as 1D SMILES, 2D molecular graphs, 3D molecular graphs with coordinates, or text captions.Text-based construction extracts entities and relationships from scientific text to build knowledge graphs.Source: §5.2 · pp. 29–30

5.6.1 Data Resources

• Scientific GraphRAG uses public and private data sources, including authoritative molecular resources PubChem, ChEMBL, and ZINC, plus biomedical sources PubMed and ClinicalTrials.Source: §5.2 · p. 29

5.3 Retriever

• Retrievers use heuristic or deep-learning methods, including similarity, matching, knowledge-graph, fusion, protein-interaction, biomedical-document, and LLM-based retrieval.Examples retrieve similar molecules [433], linked biomedical entities [318, 225], reasoning chains [185], context triples, target-aware ligands, and biomedical evidence [175] [225].Source: §5.3 · pp. 30–31

5.4 Organizer

• Organizers either concatenate retrieved knowledge with queries or fuse retrieved examples into input embeddings using cross-attention.Query-based methods include reasoning-chain insertion, context pruning, and data rebalancing, while embedding-based methods fuse molecular or ligand representations [185] [449] [225] [484] [78].Source: §5.4 · p. 31

5.5 Generator

• Generators comprise Transformer, diffusion, and LLM-based models, supporting chemistry generation, target-pocket ligand design, molecular property prediction, and medical QA.Chemformer powers RetMol, IRDIFF uses protein-molecule interactions for strongly binding molecules, and LLM systems include Med-GraphRAG [449], MolecularGPT [265], DALK [225], and HyKGE [185].Source: §5.5 · p. 32

5.6.2 Tools

• Supporting tools include RDKit for cheminformatics processing and CADRO [23] for extracting Alzheimer’s-disease-related samples and terminology from medical QA datasets.RDKit supports molecular representations, fingerprints, similarity, reactions, and predictive modeling, while CADRO organizes Alzheimer’s-related terms into a three-tier classification.Source: §5.6.2 · p. 33

6.1 Application Tasks

• Social-graph GraphRAG supports entity-property prediction, text generation, recommendation, question answering, and fake-news detection by combining social structure with textual or interaction information.Applications include personalized answers, sparse-interaction recommendation, structurally informed generation, and credibility assessment from related posts.Source: §6.1 · p. 34

6.2 Social Graph Construction

• Social graphs are constructed around proximity-, role-, and personalization-based rationales using user-user, user-item, item-item, metadata, and agent-agent interactions.User-user, user-item, and metadata relations are naturally formulated, whereas item-item relations are manually extracted and agent-agent relations are simulated.Source: §6.2 · pp. 34–35

6.3 Retriever

• Social-graph retrievers range from ID-based and collaborative-filtering retrieval to relational neighborhood retrieval and integrated neural-symbolic retrieval.Neural-symbolic designs combine rule-based precision over identifiers, relationships, or interaction patterns with embedding-based adaptability and contextual similarity.Source: §6.3 · p. 35

6.4 Organizer

• Organizers curate retrieved social-graph content through keyword extraction, profile summarization, and hierarchical graph aggregation and summarization.Hierarchical summarization compresses each neighborhood layer before propagation, controlling context growth while expanding the receptive field.Source: §6.4 · p. 36

6.5 Generator

• Generators are either LLM-based text generators for recommendations, explanations, and reviews or prediction-based models for recommendation and social-property tasks.Ground-truth text can ground LLM outputs against platform items, while prediction-based systems use graph neural networks, transformers, or multilayer perceptrons for direct outputs.Source: §6.5 · p. 36

6.6 Resources and Tools

• Resources for social-graph GraphRAG include relational and textual datasets such as STARK-Amazon, Amazon-Review, MovieLens, Netflix, Yelp, Weibo, Brexit, Diginetica, and Yoochoose, plus X and Reddit APIs and RecBole.These resources provide product, review, preference, social-network, session, and user-item interaction data for retrieval, recommendation, and social-network research.Source: §6.6.1 · p. 37

6.6.2 Tools

• The X Developer Platform and Reddit API expose real-time or historical social content, while RecBole provides a Python/PyTorch framework with over 100 recommendation algorithms and 43 preprocessed benchmark datasets.Their user interactions and textual metadata can serve as GraphRAG inputs.Source: §6.6.2 · p. 38

7.1 Application Tasks

• Planning and reasoning GraphRAG supports sequential plan retrieval, asynchronous planning, commonsense and defeasible reasoning, tool use, and embodied planning.Plan retrieval returns tool or action subgraphs; asynchronous planning incorporates time constraints, while graph-based search enables adaptive tool selection using A* or Monte Carlo Tree Search [570] [142].Source: §7.1 · p. 39

7.2 Reasoning and Planning Graph Construction

• Graph construction adds edges for resource, temporal, inclusive, causal, and analogy dependencies, while distinguishing resource sharing from cause-and-effect relations.Resource edges can connect tools when one output matches another input; temporal edges encode order, inclusive edges encode shared categories or hierarchies, and causal and analogy edges represent additional reasoning structures [277] [499] [503].Source: §7.2 · pp. 39–40

7.3 Retriever

• Retrievers traverse graphs from query-seeded nodes and select neighbors using either embedding similarity, heuristic rewards, or thought propagation.Embedding-based retrieval matches decomposed queries to neighboring APIs, heuristic retrieval can use A* with accumulated and anticipated costs, and thought propagation solves analogous problems before aggregating their solutions [454] [499].Source: §7.3 · p. 40

7.4 Organizer

• Planning and reasoning GraphRAG generally omits organizer modules because multi-round reasoning-aware retrieval and reward-based search provide sufficient precision without fine-grained reranking.Unlike one-shot top-K retrieval used for many document or knowledge graphs, these methods integrate embedding similarity with reasoning steps to improve plan fidelity [452].Source: §7.4 · pp. 40–41

7.5 Generator

• Generators either return retrieved plans directly or provide graph-structured plans and execution results to an LLM for downstream response generation.Existing systems commonly textualize graph formats such as adjacency matrices, adjacency lists, edge lists, or CSR representations when fusing graph information with language models [454] [355] [248].Source: §7.5 · p. 41

7.6 Resources and Tools / 7.7 Data Resources

• The section summarizes resources for GraphRAG on planning and reasoning graphs, including tool repositories, asynchronous planning, and API-based task graphs.Hugging Face and Multimedia encode resource-dependent tool connections; Daily Life APIs encode temporal dependencies; RestBench models categorical associations and resource dependencies; AsyncHow represents ordering constraints and causal links in DAGs.Source: §7.6 · p. 41, §7.7 · pp. 41–42

7.7 Data Resources

• The surveyed data resources construct graph instances through sampled tool subgraphs, expert annotations, LLM-assisted preprocessing, or iterative graph verification.RestBench annotates and verifies gold API paths, AsyncHow contains 1.6K DAG-based planning examples, and EXPLAGRAPHS verifies and refines commonsense explanation graphs; GSM8K and PrOntoQA provide structured multi-step reasoning data [372] [454] [248] [342].Source: §7.7 · pp. 41–42

7.7.1 Tools

• ToolBench benchmarks open-source LLM tool manipulation with diverse real-world software tools and an accessible execution-success evaluation infrastructure.It targets the performance gap between proprietary model APIs and open-source LLMs in software tool use [355].Source: §7.7.1 · p. 42

8 Tabular Graph

• Tabular graphs address serialization’s disruption of feature permutation invariance, preserve primary/foreign-key relationships across tables, and enable efficient retrieval when tables exceed LLM context windows.Tabular data may occupy single or multiple relational tables, motivating graph representations for structural reasoning and retrieval.Source: §8 · p. 43

8.1 Application Tasks

• Applications include node-level prediction, link-level prediction, graph-level prediction, table question answering, and semantic table retrieval.Examples include fraud detection, data imputation, table similarity prediction, graph-enhanced Table QA by Zhang and Zhang et al. [531], and natural-language table retrieval [61].Source: §8.1 · p. 43

8.2 Tabular Graph Construction

• Tabular graph construction uses instance, feature, instance-feature, cell, hypergraph, knowledge-graph, tree, and cross-table representations.Instance graphs use rules, similarity or learnable edge weights; feature graphs use learned relationships, similarity, or expert heuristics; cross-table graphs connect rows through primary-foreign keys without manual table merging [115, 71, 517].Source: §8.2 · p. 43

8.3 Retriever

• Tabular-graph retrieval remains early-stage and largely follows established retrieval methods, including timestamp-based time-consistent subgraphs and deterministic heuristic selection.Fey et al. [115] retrieve subgraphs using timestamps, whereas Cvitkovic [71] uses a deterministic heuristic.Source: §8.3 · p. 44

8.4 Generator

• Generators predominantly use GNNs or Graph Transformers, with some systems combining them with tabular predictors or gradient-boosted trees, while tabular graphs can help LLMs retain structural information.Wang et al. [417] fuse GNNs with DeepFM [129], FT-Transformer [125], XGBoost [50] and AutoGluon [101], and Ivanov and Prokhorenkova [171] jointly train GBDT and GNN.Source: §8.4 · p. 44

8.5.1 Data Resources

• Resources include relational and tabular benchmarks such as RelBench [339] [115], TabGraphs, Tabular-benchmark, DBInfer Benchmark [417], HiTab [61], OpenML, Kaggle, and the UC Irvine repository.These resources cover node, link, temporal, static, classification, regression, question-answering, and multi-table learning scenarios.Source: §8.5.1 · p. 45

8.5.2 Tools

• The listed tools support tabular preprocessing, model training, evaluation, heterogeneous data, and AutoML, including PyTorch Tabular, DeepTables, and PyTorch Frame [156].PyTorch Frame [156] supports numerical, categorical, time, text, and image columns and integrates with diverse architectures, including Large Language Models.Source: §8.5.2 · pp. 45–46

9 Other Domains

• The review consolidates GraphRAG research in infrastructure, biological single-cell, scene, and random graphs, where domain-specific structures support retrieval, reasoning, and generation.Infrastructure GraphRAG is especially sparse, while the other domains illustrate graph construction and application patterns relevant to GraphRAG.Source: §9 · p. 46, §9.1 · p. 46, §9.4 · p. 49

9.1 Infrastructure Graph

• Infrastructure graphs model physical and logical networks for utility prediction, simulation, vulnerability analysis, maintenance, and optimization, but no existing work directly explores GraphRAG there.Prior RAG and generative-AI systems address pedestrian behavior, traffic tasks, networking data retrieval, and operator assistance, providing foundations for infrastructure GraphRAG.Source: §9.1 · pp. 46–47

9.2 Single-cell Graph

• Single-cell graphs represent gene, cell, peak, spatial, and multimodal relationships through weighted, KNN, or heterogeneous constructions for cell annotation, deconvolution, and regulatory analysis.Multi-omics graphs connect cell nodes with gene and peak nodes, while spatial graphs integrate expression, location, and morphology to capture intra- and cross-modality relationships.Source: §9.2 · pp. 47–48

9.3 Scene Graph

• Scene-graph methods retrieve spatial and semantic subgraphs for visual question answering and embodied planning, enabling LLMs to reason over relevant scene context.G-Retriever indexes JSON objects and attributes before subgraph construction, whereas P-RAG converts observations into scene graphs and retrieves historical trajectory context.Source: §9.3 · p. 48

9.4 Random Graph

• Random graph models provide controlled structures for analyzing GraphRAG retrieval, organization, and generation, and for evaluating GNN and LLM graph reasoning.Studies test tasks including edge existence, counting, shortest paths, cycle checking, substructure counting, and bipartite matching across generated graph families.Source: §9.4 · p. 49

10.1 Graph Construction

• Graph construction must adapt node and edge granularity, entity and relationship extraction, disambiguation, alignment, coreference, representation, multimodality, and dynamic updates to each application.Choosing whether to use one or multiple graphs and maintaining efficient, effective updates remain difficult.Source: §10.1 · p. 49

10.2 Retriever

• GraphRAG retrievers must distinguish symbolic from neural knowledge, reconcile internal and external information, balance accuracy with diversity and novelty, and support adaptive reasoning and planning.External knowledge may overlap with or conflict with internal knowledge, requiring calibration, validation, reconciliation, and selective updating.Source: §10.2 · p. 50

10.3 Organizer

• Organizers must compress large retrieved graphs without losing essential context, structure content effectively, align heterogeneous modalities, and augment data without adding irrelevant or redundant information.Different tasks may require different structuring methods, including preserving graph structure for structure-sensitive tasks.Source: §10.3 · p. 50

10.4 Generator

• Generators need suitable prompting formats beyond text and structural encodings that enable LLMs to interpret retrieved subgraphs and their geometric relationships.No systematic study has established whether LLMs distinguish existing graph encodings and recognize their corresponding structures.Source: §10.4 · pp. 50–51

10.5 GraphRAG as a System

• A cohesive GraphRAG system must globally coordinate construction, retrieval, organization, and generation while scaling storage, querying, training, serving, fine-tuning, evaluation, and maintenance.Scalability requires techniques such as subgraph sampling, pathfinding, efficient fine-tuning, hardware acceleration, compression, and online maintenance.Source: §10.5 · p. 51
• Trustworthy GraphRAG requires reliability, robustness, safety, privacy, and explainability that address multi-hop uncertainty, structural attacks and leakage, and faithful reasoning paths.Graph-specific risks include accumulated multi-step errors, attacks targeting structurally influential nodes, indirect privacy inference through neighbors, and the challenge of retrieving contextually relevant explanatory subgraphs.Source: §10.5 · pp. 51–52

10.6 Evaluation of GraphRAG.

• GraphRAG evaluation needs component-specific datasets and metrics, end-to-end benchmarks covering quality, latency, efficiency, and resources, diverse task-domain tests, and trustworthiness assessments.Trustworthiness benchmarks should measure adversarial robustness, privacy protection, transparency, explainability, and reliability.Source: §10.6 · pp. 52–53

10.7 New Applications

• New applications such as code generation [262] and robust cyber defense require tailored GraphRAG strategies based on each field’s requirements, data structures, and graph configurations.Adapting construction, retrieval, organization, and generation is necessary to maximize effectiveness across domains.Source: §10.7 · p. 53

11 Conclusion

• The framework unifies GraphRAG architectures into five components: graph construction, retriever, organizer, generator, and data source.The survey reviews representative techniques for each component.Source: §11 · p. 53
• Domain-specific reviews show GraphRAG’s flexibility across knowledge, document, scientific, and social graphs for diverse task demands.The survey examines designs tailored to different graph structures and applications.Source: §11 · p. 53
• The survey concludes by outlining challenges and opportunities that could advance GraphRAG beyond its current capabilities.These directions aim to push the boundaries of GraphRAG research.Source: §11 · p. 53