A large language model knowledge-guided composite multi-attention (LKCM) method was proposed to address the shortcomings in current document-level relation extraction (DRE) methods, namely, the insufficient differentiation of internal feature importance in various semantic information and the limited, hard-to-expand scale of external domain knowledge. By integrating a composite multi-attention framework, the attention mechanism was utilized to meticulously extract features at the word, sentence, and document levels to effectively distinguish the varying importance of internal features across different semantic information. A large language model was fine-tuned as a dynamic domain knowledge base component and its extensive commonsense knowledge and reasoning capabilities were leveraged to continuously provide guidance for the model. This design effectively mitigates the issues of limited knowledge scale and difficult real-time expansion. Experimental results on a real-world medical relation dataset showed that the average F1 score of the LKCM was 1.54 percentage points higher than that of the best baseline. The comprehensive analysis demonstrated that this method not only enhanced the capture of long-distance, cross-sentence relations but also improved the identification of key features. The LKCM method exhibits strong performance and broad applicability.