Efficient ambulance dispatch is essential for timely emergency medical care under limited fleet capacity and uncertain patient demand. This paper presents an event driven semi Markov decision model in which dispatch decisions are triggered by patient call arrivals or newly idle ambulances. The model allows strategic ambulance withholding and ambulance redirection. To solve the problem, we develop a transformer reinforcement learning framework that learns contextual ambulance and patient location assignments through attention based communication. We further distill the learned policy into a surrogate decision tree to support explainable dispatch recommendations. Experiments on synthetic cases and Toronto paramedic data show that the proposed approach reduces patient waiting and improves ambulance operating efficiency.