Abstract: The rapid digital and intelligent transformation of railway information systems has created an urgent demand for fine-grained, explainable safety knowledge recommendations. To address the fragmentation of cross-modal associations and insufficient alignment with operational rules exhibited by traditional approaches, a framework integrating multimodal feature fusion with generative reasoning is investigated. A hierarchical railway safety knowledge graph is constructed, and topological features under business constraints are extracted via the Node2Vec algorithm. Simultaneously, a lightweight Transformer encoder (GTE) captured deep semantic embeddings of individual safety clauses. To balance contributions from graph and text features, a tunable weighting strategy is introduced, dynamically controlling the fusion ratio of text vectors and graph embeddings and applying a dual-constraint mechanism based on cosine similarity and predefined rules to generate candidate recommendations. A three-stage progressive retrieval architecture is designed to achieve precise multimodal alignment and suppress noise. Finally, the DeepSeek-R1 large language model served as the reasoning engine, with domain-specific prompting converting retrieved candidates into executable decision plans, thereby enhancing coherence and interpretability. Experiments on 27 safety documents from a railway operator, using a similarity threshold of 0.85 and a maximum of 10 recommendations per query, demonstrated a recommendation accuracy of 95% (an 8-percentage-point improvement over traditional methods) along with significant gains in contextual relevance and explainability. This investigation confirms the synergistic benefits of multimodal retrieval and generative reasoning, providing a robust technical foundation for evolving railway safety knowledge services from precise recommendation to intelligent decision support.