As structured repositories capturing intricate entity-relation semantics, Knowledge Graphs (KGs) fundamentally rely on Knowledge Graph Completion (KGC) to realize their comprehensive utilization potential. Recently, text-based KGC approaches harness the semantic understanding capabilities of pre-trained Large Language Models (LLMs) during the encoding of knowledge triples, subsequently optimizing model parameters through targeted fine-tuning to enhance the accuracy of knowledge inference. Despite their excellent performance, they only focus on entity neighborhoods and ignoring semantic patterns between relationships. To address this kind of challenges, this paper proposes a general framework, which harnesses the power of LLMs by introducing relation-aware context as well as typed sampling strategy into their native input construction procedure. Especially: (i) The relational context supplements the global relational pattern (e.g., the multi tailed entity distribution of “1-N” relation as well as “N-N” relation), enabling the model to more accurately capture logical constraints; and (ii) Typed sampling effectively reduces redundant information and preserves highly correlated contexts (e.g., “N-1” relationships to avoid duplicate tail entities). In practice, the proposed work can function as a readily integrable plug-in unit, enhancing the performance of various existing LLM-based KGC models.