Abstract: Content search is an essential function, but it presents a very difficult and challenging problem for large-scale peer-to-peer systems. In this paper, a new learning-based algorithm——SmartSearch is introduced. SmartSearch learns passively interest similarity between nodes from history search results, divides nodes into interest groups, and constructs friend relations between nodes with similar interest which can be used to locate content effectively. Simulation results show that, compared to the Gnutella algorithm, SmartSearch improves query efficiency by up to ten times without a significant increases in load.