Retrieval-Augmented Generation (RAG) is a powerful method currently widely applied to leverage external knowledge sources to answer user queries or generate text. However, when queries require multiple retrieval steps from different types of documents, the process becomes more complex. Standard methods may have difficulty in selecting and combining relevant evidence. In this paper, we introduce a solution with a strategy to split the query into smaller sub-queries to improve the efficiency of retrieved information and to synthesize them. This allows focusing on each detailed aspect of the original query, helping to simplify the retrieval process and avoid missing relevant evidence, thereby improving its quality. By integrating the results of the sub-queries, we can produce a final text of higher quality. Our code is available at: https://github.com/Quyet160903/Multi-step-queries .