In this paper we study how to use sparse representations for symbol description in retrieval tasks. To the best of our knowledge, this is the first attempt of using this kind of representation in symbol retrieval tasks. Unlike classifications tasks, we need to find how to link a sparsest symbol representation with a retrieval purpose. To take advantage of these representations we need to solve two issues. The first one concerns the visual vocabulary construction. We have solved it by means of the K-SVD algorithm using the SCIP descriptor. The second issue concerns the retrieval phase. We propose the sparse vector model by extending the tf-idf model [33] to sparse representations.