<p>Recommender systems have made significant strides in integrating auxiliary knowledge from knowledge graphs. However, existing knowledge-based recommendation models often encounter challenges, including sparse supervision signals and insufficient capacity to capture user implicit intents. Although graph neural networks (GNNs) effectively model high-order relational structures, their training is heavily based on sparse user-item interaction data. Moreover, user-item interactions frequently stem from distinct semantic intents that current models do not adequately represent. To overcome these issues, we propose a novel recommendation framework that synergistically combines multilevel cross-view contrastive learning with intent-aware item embeddings. Our model leverages a knowledge attention mechanism to selectively aggregate salient knowledge graph information, thus constructing intent-oriented item representations. Additionally, we perform multilevel contrastive learning across collaborative, semantic, and global structural views, enabling the extraction of rich self-supervised signals and significantly enhancing representation learning. Extensive experimental evaluations confirm that our approach achieves substantial performance improvements over state-of-the-art baselines in the precision of the recommendations. Further ablation studies verify the effectiveness of our proposed contrastive learning paradigm, knowledge attention mechanism, and semantic similarity graph to improve the expressiveness and generalization of the model.</p>