The big challenge of constructing P2P applications is how to implement efficient distributed file searching in complex environment which implies huge-amount users and uncontrollable nodes. FriendSearch is introduced to improve the efficiency and scalability of distributed file searching. FriendSearch introduces a new hybrid architecture in which the storage and search of raw file is based on DHT network, but the storage and search of meta-data is based on unstructured P2P network. FriendSearch learns interest similarity between participating nodes and uses it to construct friend relations. The forwarding of queries is limited to friend nodes with similar interests. Simulation tests show that FriendSearch algorithm is both efficient and scalable.