Large language models often encounter hallucination issues during the generation process and exhibit limitations in dynamic knowledge updates and complex reasoning. Retrieval-Augmented Generation (RAG) effectively addresses problems such as outdated knowledge and limited context in generative models. However, it relies on a retrieval mechanism based on vector similarity, which struggles to integrate dispersed knowledge, fails to support deep reasoning and cross-domain information integration, and performs poorly when handling global and complex problems. To address these limitations, we propose Multi-Semantic RAG (MSE-RAG), which incorporates a Knowledge Graph Structuring Module(KGSM) and a Hierarchical Semantic Reasoning Module (HSRM) to effectively overcome these shortcomings. The KGSM transforms textual content into a structured network of entities and relationships, providing clear semantic associations, thereby more effectively supporting the understanding and processing of complex information.The HSRM combines direct and indirect semantic path reasoning to capture the latent semantics of multi-hop associations within the graph, thereby improving the reasoning depth and comprehensiveness of answers for complex problems. The evaluation results on multi-domain long-text datasets show that MSE-RAG, in addressing global questions, detailed questions, and multi-hop reasoning problems, outperforms the RAG baseline significantly in terms of answer comprehensiveness, diversity, and reasoning depth.