This paper presents a 3D shape retrieval algorithmbased on the Bag of Words (BoW) paradigm.For a given 3D shape, the proposed approach considersa set of feature points uniformly sampled on the surfaceand associated with local Fourier descriptors; thisdescriptor is computed in the neighborhood of each featurepoint by projecting the geometry onto the eigenvectorsof the Laplace-Beltrami operator, it is very informative,robust to connectivity and geometry changesand also fast to compute. In a preliminary step, a visualdictionary is built by clustering a large set of featuredescriptors, then each 3D shape is described by anhistogram of occurrences of these visual words, hencediscarding any spatial information. A spatially-sensitivealgorithm is also presented where the 3D shape is describedby an histogram of pairs of visual words. Weshow that these two approaches are complementary andcan be combined to improve the performance and therobustness of the retrieval. The performances have beencompared against very recent state-of-the-art methodson several different datasets. For global shape retrievalour combined approach is comparable to these recentworks, however it clearly outperforms them in the caseof partial shape retrieval.