Retrieval augmented generation (RAG) system is very important in improving the performance of large language models (LLMs), but the existing methods are often limited to single granularity retrieval units (such as paragraphs or sentences), which is difficult to balance the retrieval efficiency and semantic integrity. Paragraph level indexes may introduce redundancy, while sentence level indexes may lack context. Although the proposition level index provides finer granularity, it faces the problems of high construction cost and ambiguity. To address these challenges, this paper proposes a two-level index generation method based on multi-granularity retrieval unit fusion, which innovatively combines the two-level index structure of Title level short index and proposition level long index. The optimal index granularity is dynamically selected through the query routing layer, and the retrieval fusion layer is used to integrate the retrieval results of different granularity. The experiment is carried out on the self-built ELD elevator field dataset, which verifies the effectiveness of the method in improving the performance of information retrieval, and improves the accuracy and recall rate of retrieval.