We propose a personalized drug recommendation system leveraging ClinicalBERT-augmented graph neural networks. By integrating unstructured patient reviews with structured drug-condition relationships, our model captures both semantic and structural insights. ClinicalBERT generates contextual embeddings from reviews, while Node2Vec and a GraphSAGE-based encoder learn graph structure. A gated fusion module dynamically combines these signals, and a margin-based hinge loss guides training. This system outperforms traditional methods by understanding nuanced medical context and relationships between conditions and drugs, enabling more accurate and personalized recommendations. The use of a heterogeneous graph, combining condition and drug nodes, and a link prediction head for recommendation, ensures that our model is well-equipped to support high-precision clinical decision-making and promote better health outcomes.