Abstract: With the deep application of large language models in areas such as intelligent customer service, knowledge engineering, and text generation, they have become the core driving force behind the deployment of artificial intelligence technologies. However, in multi-stage question answering scenarios, existing methods generally adopt a direct retrieval approach based on the original question, which struggles to capture the implicit logical associations within the question, leading to semantic discontinuities and path deviations in complex reasoning. To address these limitations, this paper proposes a multi-stage question answering framework based on a question-decomposition chain of thought. First, through a query-aware entity tree paragraph filtering mechanism, the framework systematically extracts core entities and their semantic correlations, constructs a structured question representation, and achieves efficient retrieval of relevant paragraphs while suppressing noise. Second, using the GPT-3 model, the framework takes the complex question together with the entity-tree-related paragraphs as prompts to generate semantically independent, high-quality sub-questions. Finally, during the iterative retrieval-and-reasoning process, the sequence of sub-questions serves as the query basis for multi-stage retrieval. This decomposition strategy captures the core semantic units of complex questions, transforms compound questions into analyzable modular subtasks, and provides structured knowledge representations and diversified retrieval perspectives for multi-stage reasoning. Experimental results on public datasets show that the proposed method improves retrieval and question answering F1 scores by 1.2% to 4.5% compared to existing methods. Moreover, it achieves favorable performance in railway bidding system processing, promoting the application of large language models in the railway vertical domain.