Although Retrieval-Augmented Generation (RAG) has shown promise in relation extraction, current methods mostly adopt single encoders, overlooking the potential gains from combining diverse encoding strategies. Multi-encoder frameworks can leverage the Mixture-of-Experts (MoE) architecture to adaptively select the most suitable encoder or encoder combination for a given input, thereby maximizing the effectiveness of feature representation. We propose RAG-MixSBERT-RE, a RAG framework that incorporates a gating network to dynamically select from multiple SBERT encoders based on input features. By fusing embeddings from diverse SBERT variants via weighted integration, our method enhances representation quality and adaptability in RE tasks. In our experiments, we evaluate our model on five benchmark datasets (TACRED, TACREV, Re-TACRED, SemEval, and Wiki80) and show that RAG-MixSBERT-RE consistently outperforms single-encoder baselines.