Focusing on the low quality of expansion source of traditional Pseudo Relevance Feedback (PRF) algorithms, which lead to low retrieval performance, a retrieval result based sorting model, namely REM, was proposed. Firstly, the first-pass retrieval result was considered as a pseudo relevant set. Secondly, documents in the pseudo relevant set were re-ranked based on rules of maximizing the relevance between the user query intention and the documents of pseudo relevant set and minimizing the similarity between documents. Finally, the top ranked documents of the re-ranking were regarded as the expansion source to the second-retrieval. The experimental results show that, compared with two classical PRF methods, the proposed model can improve the performance of retrieval and obtain more relevant feedback document to the user query intention.