Parametric many-query problems demand models that balance accuracy and computational cost. In this presentation, we discuss hierarchies for parametric problems consisting of multiple models with different accuracies and computational complexities. These models interact in an adaptive manner based on a posteriori error estimation by enriching each other with additional training data. For instance, a full order model creates snapshots to extend the reduced space of a reduced basis model while every solve of the reduced basis model generates new training data for a machine learning based model. We analyze the resulting hierarchy in terms of reduced basis construction from a theoretical perspective and establish connections to weak greedy algorithms. Further, we showcase the performance of the adaptive hierarchy on an example of a parametric partial differential equation.