G-Retriever is a retrieval-augmented generation model designed for textual graph understanding and question answering. It enables users to interact with graphs through a conversational interface, allowing them to ask questions about the graph and receive textual answers with highlighted relevant parts. Unlike previous methods that focus on conventional graph tasks or simple queries on small graphs, G-Retriever is tailored for real-world textual graphs and handles complex, real-world applications such as scene graph understanding, common sense reasoning, and knowledge graph reasoning. The model introduces a new Graph Question Answering (GraphQA) benchmark, which includes data from multiple tasks and evaluates models on various graph-related questions. G-Retriever uses a retrieval-augmented generation approach, formulating the task as a Prize-Collecting Steiner Tree optimization problem to resist hallucination and scale effectively to larger graphs. The model outperforms baselines in multiple domains, scales well with larger graph sizes, and mitigates hallucination. It also enhances scalability and efficiency by allowing selective retrieval of relevant graph parts. The model's architecture integrates the strengths of graph neural networks, large language models, and retrieval-augmented generation, enabling efficient fine-tuning while preserving the LLM's pretrained language capabilities. The model's performance is validated through experiments on three datasets, demonstrating its effectiveness in answering graph-related questions and reducing hallucination.G-Retriever is a retrieval-augmented generation model designed for textual graph understanding and question answering. It enables users to interact with graphs through a conversational interface, allowing them to ask questions about the graph and receive textual answers with highlighted relevant parts. Unlike previous methods that focus on conventional graph tasks or simple queries on small graphs, G-Retriever is tailored for real-world textual graphs and handles complex, real-world applications such as scene graph understanding, common sense reasoning, and knowledge graph reasoning. The model introduces a new Graph Question Answering (GraphQA) benchmark, which includes data from multiple tasks and evaluates models on various graph-related questions. G-Retriever uses a retrieval-augmented generation approach, formulating the task as a Prize-Collecting Steiner Tree optimization problem to resist hallucination and scale effectively to larger graphs. The model outperforms baselines in multiple domains, scales well with larger graph sizes, and mitigates hallucination. It also enhances scalability and efficiency by allowing selective retrieval of relevant graph parts. The model's architecture integrates the strengths of graph neural networks, large language models, and retrieval-augmented generation, enabling efficient fine-tuning while preserving the LLM's pretrained language capabilities. The model's performance is validated through experiments on three datasets, demonstrating its effectiveness in answering graph-related questions and reducing hallucination.