Document-level relation extraction (DocRE) is an important challenge in the field of artificial intelligence and natural language processing. The task focuses on detecting relational links between entity pairs presented across entire documents. Our analysis of popular document-level datasets, such as Re-DocRED, reveals a pronounced imbalance in the distribution of relation types, particularly highlighting the long-tail issue. The quantity of head-relation labels might be several thousand times greater than that of tail-relation labels. Models trained in this manner may exhibit suboptimal performance when predicting on tail data. Prior studies have primarily concentrated on obtaining better embedding representations for entity pairs in documents, thus neglecting the problem of long-tail data distribution. This paper presents a Retrieval-Augmented strategy and a dual-branch relation extraction model with reweighting. Through the Retrieval-Augmented module, this model enables the interaction between tail data with a smaller sample size and similar sample data, thereby enhancing the model’s representational and generalization capabilities for tail data. The experimental findings on the Re-DocRED dataset confirm that the proposed approach surpasses existing state-of-the-art models with comparable parameter sizes, achieving superior performance in both complete datasets and long-tail scenarios. The relevant code can be found at https://github.com/QiyiJiang/RAB-DOCRE.git .