Small-world phenomenon has been observed in existing peer-to-peer (P2P) networks, such as Gnutella and Freenet. Due to the similarity of P2P networks to social networks, the previous small-world model proposed by Duncan Watts can be adopted in the design of P2P networks: each node is connected to some neighbouring nodes, and a group of nodes keep a small number of long links to randomly chosen distant nodes. Unfortunately, current unstructured search algorithms have difficulty distinguishing these random long-range shortcuts. This paper presents small world architecture for P2P networks (SWAN) with a semi-structured P2P search algorithm that is used to create and find long-range shortcuts toward remote peer groups. In SWAN, not every peer node needs to be connected to remote groups, but every peer node can easily find which peer nodes have external connections to a specific peer group.