We introduce a new nearest neighbor search al-gorithm. The algorithm builds a nearest neighborgraph in an offline phase and when queried witha new point, performs hill-climbing starting froma randomly sampled node of the graph. We pro-vide theoretical guarantees for the accuracy and thecomputational complexity and empirically showthe effectiveness of this algorithm.