Retrieval-Augmented Generation (RAG) has emerged as a groundbreaking methodology for mitigating hallucination phenomena in large language models (LLMs). However, original RAG often suffers from significant reductions in retrieval effectiveness due to the semantic ambiguity and information deficiency in user queries. Additionally, previous query rewriting methods tend to focus on being either human-friendly or generator-friendly, but may not fully suit retrieval systems. In this paper, we propose a general query rewriting with multi-stage retrieval feedback framework, QR-MSRF, which leverages supervised learning and reinforcement learning to optimize the query rewriter by integrating static and dynamic feedback for retrieval tasks. The goal is to rewrite the original query into one that better meets the needs of the retrieval system. Additionally, to enhance the coverage and accuracy of retrieval, we introduce a query rewriting strategy called the Information Alignment Strategy that achieves adaptive alignment between query expressions and retrieval targets in both quantity and semantics. Experimental results on three open-domain question answering datasets demonstrate that QR-MSRF significantly improves the retrieval and generation quality of various baseline frameworks, fully validating its effectiveness and versatility.