Guiding Graph-Based Retrieval Augment Generation:Guided Strategies for Enhancing Graph Structure Retrievaland Generation

This research presents a retrieval-augmented generation model called G3-RAG (Guiding Graph-Based Retrieval Augment Generation), designed to address the issue of information loss commonly encountered by large language models (LLMs) when processing long texts. Existing RAG models often miss critical information in lengthy documents, leading to inaccurate results. G3-RAG enhances the processing of comprehensive information and the integration of multi-source knowledge by constructing the Evi-KG knowledge graph using relation extraction techniques and integrating graph-based retrieval methods. The research also introduces the concept of a chain of thought, breaking down complex questions into sub-questions and guiding multiple rounds of reasoning to gradually find the answers, particularly excelling in multi-hop reasoning tasks. Additionally, G3-RAG incorporates the use of stage answers to enhance memory during the reasoning process. Experimental results show that G3-RAG can retrieve more precise information and make more accurate inferences when dealing with complex problems, significantly improving the overall performance of LLMs.