Retrieval Augmented Generation (RAG) has become a common practice to alleviate the hallucination of Large Language Models (LLMs). The retrieval phase of RAG, however, usually solely depends on the original query, which, to some extent, suffers from the problem of semantic gap and thus degrades the quality of the retrieved external knowledge. To address this problem and enhance the performance of the traditional RAG, we propose a rEwrite-sElect-votE-rEad paradigm ( ) that first paraphrases the original query into N rewritten ones to bridge the semantic gap from different perspectives and then determines the most valuable retrieved external knowledge via a voting manner. Besides, in the midst of the above procedures, a certain query-selecting strategy is also required to filter out the extra noise introduced by the query-rewriting process. Following this proposed paradigm, we provide our implementation of . Experimental results of our implementation on long context reading comprehension datasets from LongBench demonstrate the effectiveness of our proposed paradigm and provide a profound insight into the whole enhanced RAG process.