A new contrastive recommendation model was proposed to alleviate the performance degradation caused by the excessive reliance on structural perturbations for data augmentation. The embedding representations of knowledge graphs were leveraged to guide the contrastive learning process for highly effective item recommendation. A relation-aware knowledge aggregation module was designed to capture heterogeneous relational information from the knowledge graphs, thereby obtaining knowledge embeddings. A graph neural network encoder was utilized to learn entity representations from the user-item interaction graphs. The knowledge embeddings were incorporated into the representation learning through a knowledge-enhanced contrastive recommendation module, to enhance the user and item embeddings and improve the recommendation accuracy. Extensive experiments were conducted on three datasets of enterprise services, books, and news. The results demonstrated that the proposed model had significant advantages in handling sparse datasets. Compared with the baseline models KGAT and CKAN, the proposed model achieved average improvements of over 20% on the Recall and NDCG metrics. Compared with the state-of-the-art contrastive learning models such as KGIN, KGCL and MGDCF, an average performance gain of 10% was realized. These results demonstrate that the proposed method has comprehensive performance advantages.