RLG-RAG: Guiding the Knowledge Retrieval and Evaluation in Retrieval-Augmented Generation Framework by Reasoning Logic

The knowledge retrieval, integration, and evaluation processes in the RAG method lack the guidance of reasoning logic, leading to ongoing challenges in maintaining factual consistency. To address these issues, this paper proposes the RLG-RAG framework, which constructs a reasoning graph based on user queries to guide the knowledge retrieval, integration, and evaluation processes. By fully representing the reasoning logic of RAG, RLG-RAG dynamically models and integrates knowledge relationships during retrieval and defines a precise scope of relevant knowledge through sufficiency evaluation. This reduces inference-irrelevant knowledge that large language models may obtain. Experimental analyses on accuracy, factual consistency, and robustness demonstrate that RLG-RAG resists interference and provides accurate, factually consistent answers. The project URL is https://doi.org/10.5281/zenodo.14852250.