The small-world phenomenon includes both small average distance and the clustering effect. Randomly generated graphs with a power law degree distribution are widely used to model large real-world networks, but while these graphs have small average distance, they generally do not exhibit the clustering effect. We introduce an improved hybrid model that combines a global graph (a random power law graph) with a local graph (a graph with high local connectivity defined by network flow). We present an efficient algorithm that extracts a local graph from a given realistic network. We show that the underlying local graph is robust in the sense that when our extraction algorithm is applied to a hybrid graph, it recovers the original local graph with a small error. The proof involves a probabilistic analysis of the growth of neighborhoods in the hybrid graph model.