Running summarizing tools on a new article

Abstract: Retrieval-Augmented Generation (RAG) stands out as a promising paradigm that can be applied to improve large language models by serving text generation with external sources of knowledge. Nevertheless, traditional RAG models are mostly based on flat text corpora and similarity-based retrieval and thus are not able to realize intricate relations, multi-hop dependencies, and structured semantics. The survey provides a high-level introduction to graph-based Retrieval-Augmented Generation (Graph-RAG), which is an emerging methodology adding systematic knowledge representations to the retrieval and generation process. Graph-RAG supports relational knowledge reasoning, contextual cohesion and explicit dependency modeling, complementary to text-only retrieval by organizing knowledge into graphs e.g., knowledge graphs, citation networks, and semantic graphs. The survey discusses the principles underlying RAG, gives a description of the weaknesses of the vector-based and token-based retrieval frameworks, and explains how the graph-aware architecture overcomes these weaknesses. Graph-RAG methods are systematically classified based on the architectural design options, retrieval schemes, learning methods, and reasoning methods, such as graph neural networks, hybrid graph-vector retrieval, and multi-hop inference. The main areas of application including healthcare, finance, and scientific question answering are discussed, and their areas are improved in factual grounding, interpretability, and robustness. Lastly, there are open issues that are concerned with scalability, graph dynamic updates, interpretability, and evaluation.