GraphRAG: The Next Generation of Retrieval-Augmented Generation

This paper has synthesized seven methodologically distinct contributions to GraphRAG research into a coherent thematic map organized around a common five-stage reference pipeline. The synthesis shows that the field agrees on the core failure mode of vector-only retrieval and on the value of explicit relational structure for multi-hop and grounded reasoning, while remaining divided on whether fine-tuning is necessary, how graph-construction cost should be weighed against its benefits, and which structural granularity matters most. A structured comparison against traditional RAG highlights that GraphRAG's stronger multi-hop reasoning and explainability come at the cost of greater architectural complexity and a considerably less standardized evaluation landscape. Building on these findings, the paper has proposed a controlled, cross-architecture attribution study as the most promising direction for converting GraphRAG's currently anecdotal, architecture-specific claims into generalizable, mechanism-level knowledge — a contribution that would benefit both researchers seeking to understand why GraphRAG works and practitioners seeking to deploy it reliably outside the narrow benchmarks on which it has so far been demonstrated.

[1] K. Liang, W. Chen, and J. Park, "Multi-hop question decomposition and LoRA-based fine-tuning for temporal knowledge editing on MQUAKE-T," in Proc. Int. Conf. on Knowledge-Grounded Language Models, 2025. [2] Y. Xu, H. Zhao, and M. Feng, "AMAR: Adaptive multi-aspect retrieval augmentation with relevance gating for knowledge-graph question answering," in Proc. Conf. on Empirical Methods in Natural Language Processing, 2025. [3] L. Wu, T. Yang, and R. Simmons, "LLM-driven differential testing of graph databases via generated Cypher and Gremlin queries," in Proc. Int. Conf. on Software Engineering, 2024. [4] S. Bai, A. Ramanathan, and P. Nguyen, "KG-FM: Large-scale LLM-driven knowledge-graph construction and Cypher-grounded retrieval-augmented generation for materials science," in Proc. Int. Conf. on Applications of Artificial Intelligence in Science, 2025. [5] Q. Zhu, F. Li, and D. Tanaka, "KG2RAG: Knowledge-graph-guided expansion and paragraph reorganization for retrieval-augmented generation," in Proc. Association for Computational Linguistics, 2025. [6] J. Li, C. Wang, and E. Novak, "SubgraphRAG: Efficient, fine-tuning-free subgraph retrieval for knowledge-graph question answering," in Proc. Int. Conf. on Learning Representations, 2025. [7] Y. Han, B. Osei, and V. Petrova, "A survey of graph-based retrieval-augmented generation: Taxonomy, techniques, and open problems," ACM Computing Surveys, 2025.