Abstract: Intelligent prescription formulation recommendation systems in traditional Chinese medicine (TCM) play a vital role in assisting clinical decision-making and promoting heritage innovation. However, most corpus data in the traditional TCM domain (e.g., ancient literature and clinical medical records) exists in the form of unstructured knowledge, with significant variation in data quality. Consequently, the field of TCM has long faced challenges in intelligent prescription formulation, including a lack of high-quality datasets, the sparsity of constructed structured knowledge graphs, and insufficient utilization of deep semantic information. To address these issues, this paper proposes an intelligent prescription formulation model that integrates generative large language models (LLMs) with dual-graph hierarchical decoupled propagation. The model comprises two core modules: a generative artificial intelligence (GenAI)-based semantic enhancement module and a hierarchical gated fusion-based dual-graph hierarchical decoupling propagation module.In the semantic enhancement module, the research leverage collected corpora of ancient TCM literature to guide GenAI in generating high-quality fine-tuning datasets via prompt engineering, and employ Low-Rank Adaptation (LoRA) technology to fine-tune the DeepSeek-R1 (8 B) model for domain-specific adaptation. Similarly, through prompt engineering, the domain-fine-tuned language model is invoked to generate structured textual descriptions for knowledge graph nodes. After encoding these descriptions into semantic vectors, a gating mechanism is adopted to adaptively fuse node semantic embedding vectors with node topological embedding vectors, thereby mitigating the issue of graph sparsity. In the dual-graph hierarchical decoupling propagation module, theresearch constructed a "symptom-herb interaction graph" and a "herb-herb knowledge graph", and innovatively proposed a dual-graph hierarchical decoupling propagation method. By decoupling information propagation across the dual graphs and adopting a hierarchical gated fusion mechanism, theresearch separately modeled the clinical associations between "symptoms and herbs" and the theoretical compatibility knowledge of "herbs and herbs". Finally, the obtained node embeddings are input into a graph convolutional neural network (GCNN) to generate prescriptions in a top-K manner. This study conducts comparative experiments on two datasets (including a self-constructed dataset) against a series of baseline models. The results demonstrate superior performance in metrics such as Recall, Precision, and F1-score. To analyze the impact of each module on overall performance, ablation experiments are conducted, and the results validate the effectiveness and necessity of each module. The intelligent prescription formulation method proposed in this study provides an interpretable, scalable, and end-to-end solution for knowledge computation and intelligent auxiliary diagnosis and treatment in the field of TCM.