A comprehensive survey of knowledge Graph-Augmented Generation (Graph-RAG) for trustworthy large language models

Large Language Models (LLMs), which are capable of generating diverse expressions or reasoning paths, are fundamentally limited by their reliance on Parametric Memory and hence suffer from the 'hallucination' problem. That is, LLMs are capable of generating false yet plausible statements that are not grounded in reality, which is problematic for scientific and clinical use. A popular solution for alleviating this problem is the retrieval-augmented generation (RAG) technique. However, the retrieved sources are still large passages of semi-unstructured documents, and rigorous, verifiable, and logically coherent reasoning processes are still lacking. In this review, we summarise the latest progress in the direction of Knowledge Graph Enhanced Generation (Graph-RAG) as an attempt to achieve rigorous, verifiable, and logical reasoning processes using structured retrieval sources. We design a computation framework to provide a comprehensive taxonomy, classifying Graph-RAG methods into three areas: Graph Indexing, Graph Guided Retrieval, and Graph Enhanced Generation. We also discuss in depth the potential future directions for Graph-RAG, with an emphasis on moving beyond simple retrieval towards causal graph reasoning and actionable graphs. Finally, we present the main challenges in scaling Graph-RAG and its evaluation protocols, and conclude that Graph-RAG is a critical step toward trustworthy AI.