The growing reliance on online platforms for mental health support underscores the need for accurate, real-time suicide risk detection. Most existing datasets utilized in training AI models for suicide risk detection consist of conversation level training examples, missing critical turning points within interactions. Research is especially limited in low-resource languages like Hebrew, restricting broader applicability. We propose a hybrid inductive-deductive framework for crisis hotline chats: an inductive component uses large language models to generate message-level pseudo-labels, distilled into a compact classifier for resource-constrained settings; a deductive component links predictions to a clinically curated taxonomy of risk factors. Experiments on real-world data from a national support service show that pseudo-labeling with curriculum learning improves performance over strong baselines. Our results highlight the promise of combining inductive and deductive strategies for resource-efficient, clinically meaningful AI in suicide prevention.