In unstructured P2P systems, peers organize themselves into a random overlay. A challenging problem in these systems is to efficiently locate appropriate peers to answer a specific query. Recently, research works have focused on methods based on query history, which use the historical information of past queries and query hits to build a local knowledge base per peer. When a peer forwards a given query, it runs a learning algorithm that evaluates the query against the local knowledge base in order to select a set of relevant peers to whom the query will be routed. If the current peer fails to select a sufficient number of relevant peers it floods the query through the random overlay network, which badly affects the routing efficiency and effectiveness. To address the unsuccessful relevant peers search problem, we propose to organize the P2P network into semantic clusters of peers sharing similar knowledge bases. We implemented the proposed approach, and tested (i) its retrieval effectiveness in term of recall and precision, (ii) its routing efficiency in term of messages traffic. Experimental results show that our approach improves the recall and precision metrics while it dramatically reduce network traffic.