The social behavior of peers in peer-to-peer network can be inferred from the observable factors of the system and its components as it is created, lives and evolves. Following a social metaphor, it should be possible to use the observation of these behaviors to organize the network of peers for purposes as various as improving the retrieval performance, efficiently managing storage, improving robustness and increasing security, for instance. In order to concretely illustrate this idea and to precisely quantify its benefits in a concrete scenario, we consider the important example of the improvement of retrieval performance. We propose an unstructured peer-to-peer architecture in which the system, adaptively and in a decentralized manner, learns the expertise and interest of peers, and dynamically re-organizes itself by creating efficient communities (groups) of peers.