PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Retrieval-Augmented Generation: A Survey

Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, Siliang Tang

arXiv:2408.08921v2cs.AIcs.CLcs.IR

TL;DR

LLMs and traditional RAG struggle with missing external knowledge and complex entity relationships, motivating GraphRAG. This survey systematically organizes GraphRAG’s three-stage workflow, technologies, training methods, applications, evaluations, industrial systems, and future directions. It identifies the field’s lack of unified standard benchmarks as a key boundary for comparison.

  • Problem

    LLMs lack some domain-specific, real-time, and proprietary knowledge, while traditional RAG struggles to capture structured relational knowledge among entities.

  • Method

    The paper provides a systematic survey organized around Graph-Based Indexing, Graph-Guided Retrieval, and Graph-Enhanced Generation, plus training methods and use cases.

  • Results

    The survey categorizes GraphRAG techniques, training methodologies, downstream tasks, application domains, evaluation approaches, industrial systems, challenges, and future research directions.

  • Takeaways & Limitations

    GraphRAG leverages relational knowledge from graph datasets to improve the relevance, accuracy, and comprehensiveness of information retrieval compared with traditional RAG approaches.

  • Takeaways & Limitations

    GraphRAG lacks unified and standard benchmarks, limiting consistent comparison and objective assessment across methods.

Abstract

from arXiv · show

Recently, Retrieval-Augmented Generation (RAG) has achieved remarkable success in addressing the challenges of Large Language Models (LLMs) without necessitating retraining. By referencing an external knowledge base, RAG refines LLM outputs, effectively mitigating issues such as ``hallucination'', lack of domain-specific knowledge, and outdated information. However, the complex structure of relationships among different entities in databases presents challenges for RAG systems. In response, GraphRAG leverages structural information across entities to enable more precise and comprehensive retrieval, capturing relational knowledge and facilitating more accurate, context-aware responses. Given the novelty and potential of GraphRAG, a systematic review of current technologies is imperative. This paper provides the first comprehensive overview of GraphRAG methodologies. We formalize the GraphRAG workflow, encompassing Graph-Based Indexing, Graph-Guided Retrieval, and Graph-Enhanced Generation. We then outline the core technologies and training methods at each stage. Additionally, we examine downstream tasks, application domains, evaluation methodologies, and industrial use cases of GraphRAG. Finally, we explore future research directions to inspire further inquiries and advance progress in the field. In order to track recent progress in this field, we set up a repository at \url{https://github.com/pengboci/GraphRAG-Survey}.

1 Introduction

LLMs and traditional RAG can lack domain-specific, current, or relational knowledge, while GraphRAG uses graph structure to retrieve more precise relational information. The survey formalizes GraphRAG’s workflow and systematically reviews its methods, applications, evaluation, industrial systems, and future directions.

  • LLMs may hallucinate because their pre-training corpora lack domain-specific, real-time, and proprietary knowledge.
  • RAG supplements LLMs by dynamically retrieving factual information from large text corpora, improving contextual depth, factual accuracy, and specificity.
  • Traditional RAG can overlook structured relationships, such as citation links, because semantic similarity alone does not capture relational knowledge.
  • GraphRAG retrieves nodes, triples, paths, or subgraphs from graph databases, representing interconnections and shortening input through abstraction and summarization.
  • The survey organizes GraphRAG into G-Indexing, G-Retrieval, and G-Generation, covering technologies and training methods across these stages.
  • It also reviews downstream tasks, application domains, evaluation methodologies, industrial systems, challenges, and future research directions.

2 Comparison with Related Techniques and Surveys

GraphRAG extends RAG by retrieving relational knowledge from external graph databases rather than only text. The survey distinguishes it from LLM-on-graph methods, relates it to KBQA, and broadens the discussion to applications and potential improvements.

  • GraphRAG retrieves relevant relational knowledge from graph databases and incorporates structural information about relationships between texts.
  • Unlike previous LLM-on-graph surveys, this survey focuses on retrieving graph elements from external graph-structured databases and details related technologies and applications.
  • KBQA methods are commonly categorized as information retrieval, which retrieves relevant graph information, or semantic parsing, which executes generated logical forms against knowledge bases.
  • GraphRAG overlaps with KBQA because information-retrieval-based KBQA methods form a subset of GraphRAG approaches for downstream applications.
  • The survey extends beyond KBQA to cover GraphRAG applications across varied downstream tasks and discuss potential improvements.

3 Preliminaries

The preliminaries represent GraphRAG data as Text-Attributed Graphs and introduce GNNs and LMs as models supporting retrieval and generation. GNNs propagate information through graph neighborhoods, while LMs provide discriminative or generative language modeling capabilities.

  • 3.1 Text-Attributed Graphs: Text-Attributed Graphs uniformly represent GraphRAG data with textual attributes on nodes and edges.Knowledge graphs are a typical instance, with entities as nodes, relations as edges, and names as textual attributes.
  • 3.2 Graph Neural Networks: GNNs model graph data by updating node representations through message passing over neighboring nodes and edges.The update aggregates messages computed from a node, its neighbor, and their connecting edge.
  • 3.2 Graph Neural Networks: A readout function such as mean, sum, or max pooling converts node representations into a global graph-level representation.
  • 3.2 Graph Neural Networks: In GraphRAG, GNNs can represent graph data for retrieval and model retrieved graph structures.
  • 3.3 Language Models: Language models comprise discriminative models for conditional probability estimation and generative models for tasks such as translation and text generation.Recent GraphRAG research increasingly enhances information retrieval for LMs to address complex tasks and mitigate hallucinations.

4 Overview of GraphRAG

GraphRAG decomposes response generation into graph indexing, graph-guided retrieval, and graph-enhanced generation. The workflow constructs a task-aligned graph database, retrieves relevant graph elements, and generates outputs from the retrieved information.

  • Overview: GraphRAG jointly models retrieval and answer generation, approximating answer probabilities with an optimal retrieved subgraph.The retriever extracts the optimal subgraph, after which the generator produces the answer from it.
  • Overview: The workflow consists of Graph-Based Indexing, Graph-Guided Retrieval, and Graph-Enhanced Generation.
  • Graph-Based Indexing: Graph-Based Indexing constructs or identifies a task-aligned graph database and establishes indices on it.Sources may include public knowledge graphs, graph data, or proprietary textual and other data.
  • Graph-Guided Retrieval: Graph-Guided Retrieval extracts relevant entities, triplets, paths, or subgraphs from the graph database in response to a natural-language query.Retrieval can use semantic similarity while narrowing the subgraph search range for efficiency.
  • Graph-Enhanced Generation: Graph-Enhanced Generation synthesizes responses from the query, retrieved graph elements, and an optional prompt.A conversion function transforms graph data into a form the generator can process.

5 Graph-Based Indexing

Graph-based indexing prepares graph data and indices for GraphRAG retrieval, using open or self-constructed sources and graph, text, vector, or hybrid indexing.

  • Graph Data: Graph data comes from open knowledge graphs or self-constructed graphs tailored to proprietary and domain-specific sources.Open graphs include general and domain knowledge graphs; self-constructed graphs can integrate documents, tables, and databases.
  • Open Knowledge Graphs: General knowledge graphs store broad structured knowledge, including encyclopedic and commonsense information.Examples include Wikidata, Freebase, DBpedia, YAGO, ConceptNet, and ATOMIC.
  • Open Knowledge Graphs: Domain knowledge graphs provide specialized information and relationships for fields such as biomedicine.They support deeper understanding of complex professional relationships and domain-specific questions.
  • Graph Data: Self-constructed graphs model relationships among documents, passages, and entities using signals such as co-citation, co-topic, co-venue, and shared keywords.These graphs are closely tied to each method’s design.
  • Indexing Methods: Graph indexing preserves structure, text indexing converts graph data into retrievable descriptions, and vector indexing enables efficient similarity search.Graph search can use BFS or shortest-path algorithms, while vector search can use query embeddings and LSH.
  • Indexing Methods: Hybrid indexing is often preferred because it combines structural access, textual retrieval, and efficient vector search.HybridRAG retrieves vector and graph data simultaneously.

6 Graph-Guided Retrieval

Graph-guided retrieval extracts relevant graph elements for a query, but must manage expanding candidate subgraphs and similarity measures that capture both text and structure.

  • Retrieval Challenges: Graph retrieval faces exponentially growing candidate subgraphs as graph size increases and requires efficient search strategies.Heuristic search algorithms are used to explore and retrieve relevant subgraphs.
  • Retriever Types: GraphRAG retrievers are categorized as non-parametric, LM-based, or GNN-based according to their underlying models.Preprocessing models such as query encoders and entity linkers are excluded from this categorization.
  • Retriever Types: Non-parametric retrievers use heuristic rules or graph-search algorithms to achieve high retrieval efficiency without deep-learning models.Examples retrieve k-hop paths or optimize relevant subgraphs with an extended PCST algorithm.
  • Retriever Types: LM-based retrievers leverage language understanding, whereas GNN-based retrievers encode graph structure and score retrieval candidates against the query.GNN-based methods may score entities or iteratively retrieve relevant paths.
  • Retriever Trade-offs: Non-parametric retrieval is efficient but may be less accurate, while LM-based and GNN-based retrieval can improve accuracy at higher computational cost.Hybrid and multi-stage approaches combine retriever types to balance these properties.

6.2 Retrieval Paradigm

GraphRAG uses once, iterative, and multi-stage retrieval paradigms, trading response speed against retrieval refinement and accuracy according to application needs.

  • Once Retrieval: Once retrieval gathers relevant information in a single operation using embeddings, predefined rules, or structured graph extraction.G-Retriever uses an extended PCST algorithm to retrieve a relevant subgraph.
  • Iterative Retrieval: Iterative retrieval performs dependent searches that progressively deepen or complete the retrieved information.It is classified into non-adaptive and adaptive retrieval.
  • Iterative Retrieval: Non-adaptive retrieval follows a fixed sequence and terminates after a preset number of iterations, time limit, or threshold.PullNet, for example, retrieves problem-relevant subgraphs through T iterations.
  • Iterative Retrieval: Adaptive retrieval lets models determine when retrieval should stop using hop predictions, special signals, or generated answers.LLM agents can select information, invoke retrieval tools, and cease retrieval based on retrieved content.
  • Multi-Stage Retrieval: Multi-stage retrieval divides processing into sequential stages that can use different retrievers and interleaved enhancement or generation steps.This design tailors retrieval techniques to different aspects of a query.
  • Trade-offs: Once retrieval generally offers lower complexity and shorter response times, whereas iterative retrieval can improve accuracy through refinement but increase processing time.Paradigm selection should balance accuracy and time complexity for the use case.

6.3 Retrieval Granularity

GraphRAG retrieves nodes, triplets, paths, subgraphs, or combinations of these, with granularity choices balancing relational detail, contextual coverage, noise, and efficiency.

  • Granularity Overview: Retrieval granularity comprises nodes, triplets, paths, and subgraphs, each suited to different task scenarios and indexing types.The appropriate choice depends on the desired retrieval content and efficiency.
  • Nodes: Node retrieval targets individual entities or attribute-bearing textual nodes for precise information extraction.It can provide details about node attributes, relationships, and context.
  • Triplets: Triplets represent subject-predicate-object relations, supporting structured retrieval when entity relationships and contextual relevance matter.Their structured format organizes relational data clearly.
  • Paths: Path retrieval captures sequences of entity relationships and contextual dependencies, but possible paths grow exponentially with graph size.Methods constrain traversal with predefined rules or hop limits.
  • Subgraphs: Subgraph retrieval captures broader relational contexts, complex patterns, dependencies, and semantic connections within larger graph structures.Candidate subgraphs may be rule-based, embedding-based, or induced from retrieved paths.
  • Hybrid Granularity: Hybrid granularity retrieves multiple forms of graph information to combine detailed relationships with broader context while reducing noise.GraphRAG systems may use LLM agents to coordinate these retrievals.

6.4 Retrieval Enhancement

Retrieval enhancement improves GraphRAG by enriching queries and refining retrieved knowledge. The survey categorizes these strategies into query expansion, query decomposition, knowledge merging, and knowledge pruning.

  • Retrieval Enhancement: Query enhancement comprises query expansion and query decomposition, while knowledge enhancement comprises merging and pruning.Together, these strategies optimize retrieval by enriching queries and refining retrieved results.
  • Query Enhancement: Query expansion supplements short or information-limited queries with relevant terms, concepts, relation paths, aliases, or reweighted terms.Examples use relation paths, entity aliases from Wikidata, and consensus-based term reweighting.
  • Query Enhancement: Query decomposition breaks complex or ambiguous questions into smaller sub-queries, often corresponding to distinct relations.Systems can sequentially retrieve pertinent triplets for each decomposed sub-question.
  • Knowledge Enhancement: Knowledge merging compresses and aggregates retrieved information to provide a more comprehensive view while reducing input-length pressure.KnowledgeNavigator condenses retrieved subgraphs through triple aggregation.
  • Knowledge Enhancement: Knowledge pruning removes less relevant or redundant information using reranking-based and LLM-based approaches.Reranking can use stronger models, query-information similarity, or metrics incorporating impact, recency, and confidence.

7 Graph-Enhanced Generation

Graph-enhanced generation combines queries with retrieved graph data to produce task-specific responses. The survey covers generator selection, graph-to-generator transformations, and cascaded or parallel integration of GNNs and language models.

  • Graph-Enhanced Generation: Graph-enhanced generation transforms retrieved graph data into generator-compatible formats before producing responses conditioned on the query.Generator selection depends on the downstream task, and generative enhancement can intensify query–graph interaction.
  • Generator Selection: GNNs and discriminative language models support discriminative tasks, while encoder-decoder and decoder-only models support discriminative and generative tasks.GNN representations can be mapped through an MLP to answer logits, whereas models such as T5, GPT-4, and LLaMA process textual inputs.
  • Hybrid Models: Hybrid generation follows cascaded or parallel paradigms that combine GNN structural representations with LM textual representations.Cascaded systems pass GNN-processed graph data to an LM, while parallel systems process both inputs concurrently and merge outputs or representations.
  • Cascaded Paradigm: Cascaded approaches commonly prepend GNN-encoded graph data as a prefix to LM input embeddings and optimize the GNN through downstream tasks.This prompt-tuning pattern lets the LM generate text after receiving graph-derived encodings.
  • Parallel Paradigm: Parallel approaches merge GNN and LM outputs or representations using weighted summation, attention mechanisms, concatenation, or dedicated modules.GreaseLM Layers integrate graph and textual representations through a two-layer MLP at each layer.
  • Graph Languages: Graph languages represent retrieved subgraphs as adjacency or edge tables, natural language, node sequences, code-like forms, or syntax trees.These formats adapt graph inputs to the requirements of different generators.
  • Graph Languages: Graph languages should be complete, concise, and comprehensible because their choice can affect downstream-task performance.Lengthy representations may cause context loss or exceed LM input limits, while incompleteness can omit essential graph information.
  • Graph Embeddings: Graph embeddings avoid long textual graph inputs but face difficulties preserving precise entity information and generalizing well.Integrating graph embeddings with textual representations into a unified semantic space remains a central challenge, especially for closed-source LMs.

8 Training

GraphRAG training strategies are organized into training-free, training-based, and jointly trained retriever–generator approaches. The survey describes their supervision sources, optimization methods, and deployment settings.

  • Training-Free Approaches: Training-free methods rely on prompts or predefined retrieval procedures and are commonly used with closed-source LLMs such as GPT-4.Training-free retrievers may be non-parametric or use pretrained language models and embeddings.
  • Training-Based Retrievers: Training-based retrievers optimize query–ground-truth similarity for nodes or triplets and often use autoregressive path generation for path retrieval.Metric learning aligns relevant facts with queries while distinguishing unrelated facts.
  • Retriever Supervision: Most datasets lack ground truth for retrieval content, so researchers use distant supervision, implicit intermediate signals, or reinforcement learning.Examples derive paths between query and answer entities, enforce convergence of bidirectional searches, or reward inclusion of answers in retrieved information.
  • Retriever Supervision: Noisy distant or implicit supervision motivates self-supervised retriever pretraining, including masking randomly sampled subgraph passages.SKP pretrains DPR by converting sampled subgraphs into passages and masking them.
  • Training-Free Approaches: Training-free generators feed retrieved graph data and queries to LLMs, relying on prompts and the models’ inherent graph-and-query understanding.This setting targets closed-source LLMs or scenarios where avoiding high training costs is important.
  • Training-Based Generators: Training-based generators use supervised fine-tuning for generative LLMs or task-specific losses for GNN and discriminative generators.Inputs may include task descriptions, queries, and graph data, with outputs compared against downstream-task ground truth.
  • Joint Training: Joint training optimizes retrievers and generators together, either within one model or through alternating updates after separate initialization.Alternating training fixes one component while training the other, then reverses the roles using feedback from generation.

9 Applications and Evaluation

GraphRAG spans diverse downstream tasks and application domains, with evaluation organized around task datasets and dedicated multi-domain benchmarks. The survey’s overview connects tasks, benchmarks, methods, metrics, and industrial scenarios.

  • Evaluation: Table 1 summarizes GraphRAG tasks, benchmarks, methods, and evaluation metrics.It provides a consolidated view of techniques and their evaluation settings across the surveyed applications.
  • Downstream Tasks: GraphRAG downstream tasks include question answering, information retrieval, information extraction, fact verification, link prediction, dialogue systems, and recommendation.Question answering includes KBQA and CSQA, while information retrieval includes entity linking and relation extraction.
  • Question Answering: KBQA retrieves and reasons over structured knowledge bases to answer questions involving entities, relationships, or operations between entity sets.CSQA instead uses multiple-choice questions and external commonsense graphs such as ConceptNet.
  • Information Retrieval: Entity linking connects text mentions to knowledge-graph entities, while relation extraction identifies and classifies semantic relationships between entities.Graph structures support contextual use of interdependencies for these information-retrieval tasks.
  • Other Tasks: GraphRAG also supports fact verification, link prediction, dialogue systems, and recommendation by retrieving or structuring relational evidence.Applications include validating factual assertions, predicting missing connections, organizing conversation context, and forecasting user–product links.
  • Application Domains: GraphRAG applications span e-commerce, biomedicine, academic research, literature, legal analysis, intelligence reporting, patent similarity, and software understanding.These domains use graphs representing interactions, biomedical relations, scholarly affiliations, citations, events, patent phrases, or software structures.
  • Application Domains: Additional real-world applications include intelligence report generation, patent-phrase similarity detection, and software understanding.Examples retrieve event-graph aspects, patent-phrase ego networks, or software-related graph information.

10 Future Prospects

Future GraphRAG research must address dynamic data, multimodal and large-scale graphs, graph foundation model integration, long-context compression, benchmarking, and broader applications.

  • 10.1 Dynamic and Adaptive Graphs: Most GraphRAG methods use static databases, creating a need for efficient dynamic updates and real-time integration of emerging entities and relationships.
  • 10.2 Multi-modal GraphRAG: Knowledge graphs commonly lack images, audio, and video, while integrating these modalities presents substantial technical challenges.
  • 10.3 GraphRAG on Large-scale Graphs: Industrial knowledge graphs may contain millions or billions of entities, but many current retrieval methods target graphs with only thousands.
  • 10.4 Combination with Graph Foundation Model: Graph foundation models offer a promising direction because their graph-structured inputs may process graph data more efficiently than LLMs.
  • 10.5 Lossless Compression of Retrieved Context: GraphRAG’s graph-organized retrieval produces long contexts, making lossless compression important but difficult because current methods trade compression against information preservation.
  • 10.6 Standard Benchmarks: GraphRAG lacks unified standard benchmarks, limiting consistent comparison and objective assessment across methods.
  • 10.7 Broader Applications: Applications extend beyond customer service, recommendation, and KBQA toward healthcare, finance, legal compliance, smart cities, and IoT.

11 Conclusion

The survey organizes GraphRAG’s techniques, training methodologies, and application scenarios while highlighting its retrieval benefits and the field’s evaluation and research needs.

  • The survey provides a comprehensive retrospective that systematically categorizes GraphRAG’s fundamental techniques, training methodologies, and application scenarios.
  • GraphRAG improves retrieval relevance, accuracy, and comprehensiveness by leveraging relational knowledge from graph datasets.
  • The survey delineates benchmarks, analyzes prevailing challenges, and identifies future research directions for the relatively nascent GraphRAG field.
Loading 2408.08921v2…