Similarity search is important in many applications including data cleaning, near duplicate detection, plagiarism detection and collaborative filtering. Existing methods make use of a filter-and-verification framework. They first generate candidate strings and then verify the candidates. It is very important to generate candidates as small as possible. The prefix-filtering technique is widely used to reduce the number of candidates. In this paper, we show that we can generate candidates using partial elements in the prefix of a query. Based on the observation, we propose a technique that reduces the number of candidates significantly. Experimental results show that our technique achieves high efficiency.