Many real-world prediction tasks are hampered by the problem of having limited knowledge available for use in making predictions. Additional knowledge can often be acquired with an investment of resources, however there is great interest in minimizing the resources invested in the process of data gathering, without compromising the quality of prediction. In this paper, we present a framework for adaptive feature acquisition, comprised of three replaceable components, which tackles this problem. At the core of our method is the search for the next best question (TNBQ) to ask, given the data currently available, in order to optimally acquire features. We evaluate the framework using various datasets and provide an analysis of its performance with different configurations. We also demonstrate the benefits of the proposed framework and compare it to the state-of-the-art method.