Clustering and sampling are key methods for the study of relational data. Learning efficient representations of such data relies on the identification of major geometric and topological features and therefore a characterization of its coarse geometry. Here, we introduce an efficient sampling method for identifying crucial structural features using a discrete notion of Ricci curvature. The introduced approach gives rise to a complexity reduction tools that allows for reducing large relational structures (eg, networks) to a concise core structure on which to focus further, computationally expensive analysis and hypothesis testing.