Social networking and collaborative tagging have taken off at an unexpected scale and speed. Huge opportunities to significantly boost the search experience are out there in the Web but the amount of information to be dissected is seemingly Herculean. Moreover, users might be reluctant to publicize their profiles in order to facilitate the navigation of other users. We present Gossple, the first decentralized system to personalize the user search experience by expanding queries with information derived from anonymous social acquaintances. Underlying Gossple lies the intuition that, while social networks can provide you with news from your old buddies, you can learn a lot more from people you do not know, but with whom you share many interests. Considering a collaborative tagging system with active participants annotating content, Gossple manages each user profile and dynamically creates her personalized "social" network by gossiping and computing a distance between users, without revealing which profile is associated with which user. Using the information in this personalized social network, each user extracts knowledge about the relations between tags which she locally leverages to improve her own search experience through a personalized query expansion mechanism. We evaluate Gossple on traces crawled from CiteUlike and Delicious, with 33,834 and 20,000 users. We do so in a real distributed system of 170 PlanetLab nodes as well as by simulating a large-scale system involving thousands of peers. In short, we show that by sharing their tagging behaviors with small numbers of neighbors, users benefit from personalized and efficient query expansion, increasing the number of query results (recall) while significantly improving on quality (precision).