Addressing key challenges of general large language models in geography test item development—such as insufficient scientific rigor and misalignment with assessment objectives—the paper introduces Retrieval-Augmented Generation (RAG) technology to establish an intelligent, collaborative item-writing approach characterized by“teacherled, AI-assisted”synergy. A practical case focusing on “storm surges in the Pearl River Estuary” demonstrates that this approach, by constructing a controllable knowledge space, eff ectively enhances both the scientifi c validity and contextual appropriateness of test items. The paper further proposes a dual-cycle human-AI collaborative item development model based on RAG, off ering a practical paradigm for innovation in geographical assessment.