Herbal Prescription Recommendation (HPR) is a critical task that bridges Traditional Chinese Medicine (TCM) theory with clinical practice, holding significant value for aiding diagnostic and therapeutic decision-making as well as facilitating the transmission of TCM knowledge. However, existing approaches exhibit notable limitations in deeply integrating structural information with semantic representations, achieving robust learning from sparse and noisy data, and ensuring cross-scale interpretability of recommendation outcomes. To address these gaps, our study proposes a Graph Transformer and Contrastive Learning based HPR model (HPR-GTCL). First, we fine-tune the Bidirectional Encoder Representations from Transformers (BERT) encoder on a self-built corpus to generate high-quality semantic embeddings for symptoms and herbs. In the encoding phase, we design a graph Transformer encoder that integrates symptom−symptom and herb−herb co-occurrence matrices as learnable structural biases into the multi-head attention, and employs bidirectional cross-attention to explicitly model symptom−herb feature interactions. During optimization, we introduce a contrastive learning strategy with an improved Information Noise-Contrastive Estimation (InfoNCE) loss for self-supervised training, thereby enhancing the discriminability and robustness of learned representations. HPR-GTCL achieves significant improvements across both datasets. On Dataset 1, it improves F1-score (F1) @5, F1@10, and F1@20 by 13.39%, 16.82%, and 16.25%, respectively, over the best-performing baseline. On Dataset 2, the corresponding improvements surpass the strongest baseline by 13.40%, 12.39%, and 12.51% for F1@5, F1@10, and F1@20, respectively. HPR-GTCL not only delivers a substantial advance in predictive accuracy, but also enhances the credibility and interpretability of its recommendations through cross-scale validation. HPR-GTCL provides a precise, reliable, and interpretable method for intelligent decision support in TCM.