Knowledge graphs are essential in various artificial intelligence tasks, but they frequently encounter the issue of incompleteness. Given the long-tail distribution of relationships within real-world knowledge graphs, researchers have incorporated few-shot learning methods for knowledge graph completion (FKGC). Large language models (LLMs) encompass extensive knowledge bases that can effectively mitigate the long-tail problem. This paper introduces the LLM-AR framework, which integrates LLMs with FKGC models. In this framework, FKGC functions as a retriever, and the retrieval results are encoded into instructions using a knowledge prompting strategy to guide the LLM in enhanced retrieval. LLM-AR is compatible with the majority of existing FKGC models without incurring additional training overhead. LLM-AR was utilized in experiments on the NELL and Wiki datasets with two FKGC models, FAAN and CIAN. The experimental results demonstrate varying degrees of improvement across all metrics. On the Wiki dataset, Hits@1 for FAAN and CIAN increased by 16.3% and 15%, respectively.