Server selection is typically defined as maximizing network performance under the assumption that each server holds an exact replica of all data. We propose and evaluate methods of server selection when servers are not exact replicas such that we maximize both network performance and information retrieval (IR) precision (i.e., the relevance of retrieved data). We show that naive composition of previously proposed techniques from networking and IR perform poorly. We propose improving the performance of current IR selection techniques by using language model/based selection to construct local replicas of databases that network selection predicts are likely to be poor network performers and that IR selection predicts will have relevant results. In our experiments our technique is capable of selecting servers and retrieving information that is as accurate as techniques that focus on IR performance; moreover, our techniques reduce network latency significantly (30-67% over those IR selection techniques), as the cost of local storage (16-33% of all data).