We introduce a self-organizing similarity search system for a large-scale unstructured peer-to-peer network, called the Metric Social Network. This system does not rely on any centralized control and does not define any data clustering or partitioning principle. It combines multiple strategies into a single system which results in abilities to scale to a large number of peers, to adapt to different data distributions, and to be robust to abrupt peer disconnections. We prove these abilities by running various experimental trials on real-life, as well as, synthetic data sets stored on up to 2,000 peers. Additionally, different data distributions among the peers, ranging from clustered to totally non-clustered and real-life data distributions, are also considered.