A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models (Jan 2025, arXiv)
- arXiv link: https://arxiv.org/abs/2501.13958
- Authors: Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, et al. (12 authors in total)
- Type: Survey
- Category: RAG
- Relational and multi-hop knowledge that spans multiple documents;
- Structural context (entity relations, hierarchies) lost in chunk-based indexing;
- Domain adaptation in verticals such as enterprise search, recommendation, and conversational search.
- A unified perspective that consolidates scattered GraphRAG work into a comparable framework;
- A clear decomposition of method components: graph data preparation, graph indexing, retrievers, rerankers/planners, generators, and feedback mechanisms;
- Coverage of benchmarks, datasets, and evaluation protocols, lowering the entry cost for new researchers;
- Discussion of interfaces with LLM tool use, agents, and multi-hop reasoning;
- Explicit open problems: evaluation reliability, latency and cost, hallucination and safety, cross-lingual and multimodal extension.
- 2019–2021: BERT rerankers and DPR lay foundations for neural retrieval;
- 2022–2023: RAG-style retrieval-generation fusion becomes mainstream;
- 2024: Agentic search and generative recommendation expand the design space;
- 2025+: GraphRAG, RL-trained search agents, and deep research systems become growth frontiers.
- Datasets: MS MARCO, BEIR, Natural Questions, knowledge-graph QA sets, and domain corpora;
- Metrics: nDCG@10, MRR, Recall@k, Hit@k, task success rate, citation accuracy, latency, and token cost;
- Baselines: BM25, dense retrieval, cross-encoder reranking, and retrieval-free LLMs.
- Lack of unified GraphRAG benchmarks and reproducible private-data experiments;
- Evaluation bias in LLM-based judging;
- Safety and cost constraints of agentic systems on the open web;
- Deeper fusion with knowledge graphs and structured databases;
- Multimodal, cross-lingual, and causal/fairness-aware extensions.
- Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG (arXiv:2501.09136)
- RAFT: Adapting Language Model to Domain Specific RAG (2024)
- RAG vs. GraphRAG: A Systematic Evaluation and Key Insights (arXiv:2502.11371)
- A Survey on Retrieval-Augmented Text Generation for Large Language Models (arXiv:2404.10981)
- Original paper: A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models. https://arxiv.org/abs/2501.13958
One-line summary
This survey systematically reviews Graph Retrieval-Augmented Generation (GraphRAG), an emerging approach that uses graph-structured data and graph-based retrieval to customize and strengthen large language models beyond plain-text RAG.
Background and Motivation
Large language models (LLMs) suffer from hallucinations, outdated knowledge, and weak domain customization. Conventional RAG pipelines retrieve flat text chunks, which struggles with:
GraphRAG addresses these gaps by organizing knowledge as graphs—knowledge graphs, document/semantic graphs, or graph indexes over corpora—so retrieval can traverse relations and return more precise, context-aware evidence for generation.
Core Contributions
Taxonomy
The survey organizes GraphRAG approaches along two main dimensions:
| Dimension | Subcategories | Characteristics | |-----------|---------------|-----------------| | Retrieval source | Graph-data-based retrieval | Graphs (e.g., knowledge graphs) exist as the data itself; queries are matched against graph nodes/relations | | | Graph-index-based retrieval | Graphs are constructed as indexes over plain corpora to structure retrieval | | Adaptation | Training-free | Prompting / in-context use of graph evidence; low cost, model-agnostic | | | Training-based | Fine-tuning or instruction tuning to align LLMs with graph inputs; better grounding, higher cost |
Evolution Timeline
Four main retrieval paradigms are typically contrasted: dense retrieval (high recall, low latency), late interaction such as ColBERT (higher precision, larger index), generative IR (directly generating document identifiers), and agentic search (sequential decision-making with multi-hop and self-reflection).
Evaluation Paradigm
Typical datasets and metrics discussed in this line of research:
Note: the forum summary is based on the abstract and public metadata; exact quantitative results should be verified against the original PDF.
Key Takeaways for Search / Rec / Personalization
1. Architecture: cascade retrieve-rerank-generate remains dominant, but agentic paradigms make the number and strategy of retrieval steps itself learnable; 2. Data: high-quality instruction data and session logs matter as much as architecture; synthetic data needs leak/shift checks; 3. Evaluation: the gap between offline metrics and real user satisfaction is widening; LLM-as-judge should be cross-validated with human evaluation; 4. Product: latency, cost, explainability, and safety are hard industrial constraints—do not optimize academic benchmarks alone.
Open Problems and Future Directions
Engineering Checklist
| Item | Question | Suggestion | |------|----------|------------| | Data | PII in index? Versioning? | Partitioned indexes, anonymization, rollback-able embedding versions | | Latency | p99 budget? Retrieval steps? | Cascade + early stop, query caching, async reranking | | Quality | Does offline gain transfer online? | Interleaving experiments, human audits, citation verification | | Safety | Open retrieval poisoning/bias? | Source allowlists, adversarial detection, output filtering | | Cost | Token and GPU cost per query? | Small-model routing, distillation, hybrid sparse-dense retrieval |
Glossary
| Term | Meaning | |------|---------| | RAG | Retrieval-Augmented Generation | | GraphRAG | Graph-based Retrieval-Augmented Generation | | Agentic Search | Search modeled as sequential decisions and tool calls | | nDCG | Normalized Discounted Cumulative Gain, a ranking-quality metric | | Gen-IR | Generative Information Retrieval |