Initial models for complex networks consisted of various random graphs satisfying powerlaw degree distributions and small world properties. However, such models fail to satisfy important finer properties observed in real networks, including sparse cuts and negative assortativity. We consider models which preserve the mathematical clarity of of random graphs while introduce network semantics matching real data. In particular, we consider networks whose nodes are generated from probability distributions in high dimensional spaces (one dimension for each relevant attribute of the data), and links are determined according to the inner product, or more general kernel functions, depending on the application. We present and review analytical and experimental results for suitably chosen generating distributions and kernel functions.