Collaborative Filtering (CF) recommendation systems plays a key role in recommendation algorithms. However, collaborative filtering often relies heavily on a large number of tags, which in real-world scenarios usually do not exist. The current mainstream approach to addressing this issue is twofold: We enhance collaborative filtering through two key innovations: first by incorporating self-supervised contrastive learning to reduce label dependency, and second by leveraging Large Language Models (LLMs) to extract semantic information from item descriptions, addressing data sparsity issues.. However, the current contrastive learning approach introduces additional views, which increases the complexity of collaborative filtering recommendation algorithms. Moreover, in hot scenarios, the performance of current LLMs is not as good as that of collaborative filtering algorithms. We consider employing dual modalities of collaborative interaction and textual modalities for recommendation. We introduce HRLLM (Hierarchical Relationship LLM), a novel hierarchical contrastive learning framework enhanced by large language models, which achieves single-view contrastive learning to extract interaction information and utilizes BERT to extract textual information, demonstrating superior performance in both cold and hot scenarios.