Efficient search for unstructured peer-to-peer information retrieval systems is a challenging problem. This paper proposes a semantic overlay based on the small world phenomenon that facilitates efficient search for information retrieval in unstructured P2P systems. In the semantic overlay, each node maintains a number of short-range links which are semantically similar to the node, together with a small collection of long-range links that help increase recall rate of information retrieval as well as reduce network traffic. In this paper we present how to construct and maintain a semantic small world overly, and the search strategy utilizing the above semantic overlay. Experimental results show that our model can improve performance by 150% compared to Gnutella and by up to 50% compared to the Interest-based model, a similar shortcut-based search technique.