To address the challenges of dispersed information sources, rapid event evolution, complex emergency-plan matching, and high multi-actor coordination demands in urban rail transit operational risk identification and emergency response, a safety-constrained framework integrating large language model (LLM), retrieval-augmented generation (RAG), and human-in-the-loop (HITL) for urban rail transit operational risk governance was studied through literature analysis, scenario analysis, framework construction, normative analysis, and prototype validation, encompassing risk information input, knowledge-enhanced retrieval, model-assisted generation, human review and authorization, and closed-loop governance feedback. The results show that: a safety-constrained framework consisting of five links, i.e., multi-source information input, knowledge-enhanced retrieval, model-assisted generation, human review and authorization, and closed-loop governance feedback, is constructed; five applicable scenarios, i.e., risk information parsing, emergency-plan matching, situation-summary generation, cross-departmental collaborative communication, and event review and learning, are identified; a mapping relationship among risk levels, model permissions, and human review mechanisms is established; RAG failure modes including outdated knowledge bases, incorrect retrieval recall, retrieval omission, context truncation, misinterpretation of rules, unsupported generation, and untraceable sources are summarized along with corresponding control strategies. The results show that hierarchical knowledge organization with multi-granularity indexing achieves a hit@5 of 0.886 with a retrieval latency of about 24ms, constrained-template generation achieves an average usability score of 4.39 out of 5 with a citation grounding rate of 0.89, and a generation latency of about 14.2 s remains the main response-time bottleneck. Accordingly, engineering implementation suggestions regarding data security governance, responsibility boundary control, and response latency control are proposed. It can be concluded that large language models should not replace dispatchers or automatically execute emergency instructions in urban rail transit safety governance; rather, they are more appropriately positioned as safety-constrained human-machine collaborative decision-support tools, providing a framework reference for LLM-assisted risk information processing, emergency knowledge retrieval, and intelligent emergency platform development in urban rail transit enterprises.