Few-Shot Relation Extraction (FSRE) focuses on extracting relational facts from limited labeled data, a crucial task in low-resource natural language processing. Recent methods leverage pre-trained language models (PLMs) and contrastive learning to encode relations as label embeddings and predict instance-label similarity. However, existing approaches often underutilize instance-label pairs, neglect auxiliary relational information, and suffer from encoding gaps, leading to suboptimal performance. To overcome these issues, we introduce a method called SadpFSRE, a novel collaborative anchoring contrastive pre-training framework that aligns representations using symmetric instance-anchored and prototype-anchored objectives. Our framework improves prototype quality through auxiliary information and a semantic alignment-based prototype aggregation method, enhancing alignment and unifying the representation space while maximizing mutual information across relational perspectives. Experiments on FewRel 1.0 and 2.0 demonstrate the effectiveness of SadpFSRE, achieving significant performance gains over baseline models. The framework also exhibits faster convergence and achieves higher adaptability to domain adaptation and Zero-shot Relation Extraction (ZSRE) tasks. Codes are accessible at https://github.com/Cishenn/SadpFSRE .