Graph Retrieval-Augmented Generation (GRAG) addresses the limitations of traditional Retrieval-Augmented Generation (RAG) methods in graph-based contexts by enhancing the generation capabilities of Large Language Models (LLMs) through the retrieval of query-relevant textual subgraphs. GRAG preserves both textual and topological information essential for accurate reasoning, using a divide-and-conquer strategy with k-hop ego-graphs and a soft pruning mechanism to mitigate irrelevant entities. It introduces dual prompting with hard and soft prompts to maintain semantic nuances and graph topology. GRAG significantly outperforms state-of-the-art RAG methods and LLM baselines in multi-hop graph reasoning tasks, particularly in scenarios requiring detailed, multi-hop reasoning on textual graphs. Empirical results show that GRAG not only addresses the NP-hard problem of exhaustive subgraph searches but also demonstrates that a frozen LLM with GRAG can outperform fine-tuned LLMs with lower training costs. The approach effectively mitigates hallucinations and improves the generation quality by integrating both textual and topological information, making it a significant advancement in integrating graph-based information retrieval and generation.Graph Retrieval-Augmented Generation (GRAG) addresses the limitations of traditional Retrieval-Augmented Generation (RAG) methods in graph-based contexts by enhancing the generation capabilities of Large Language Models (LLMs) through the retrieval of query-relevant textual subgraphs. GRAG preserves both textual and topological information essential for accurate reasoning, using a divide-and-conquer strategy with k-hop ego-graphs and a soft pruning mechanism to mitigate irrelevant entities. It introduces dual prompting with hard and soft prompts to maintain semantic nuances and graph topology. GRAG significantly outperforms state-of-the-art RAG methods and LLM baselines in multi-hop graph reasoning tasks, particularly in scenarios requiring detailed, multi-hop reasoning on textual graphs. Empirical results show that GRAG not only addresses the NP-hard problem of exhaustive subgraph searches but also demonstrates that a frozen LLM with GRAG can outperform fine-tuned LLMs with lower training costs. The approach effectively mitigates hallucinations and improves the generation quality by integrating both textual and topological information, making it a significant advancement in integrating graph-based information retrieval and generation.