Structural libraries of fragments are commonly used to model or design the 3D structure of biomolecules (drugs, peptides, nucleic acids). They typically approximate all possible local conformations of these molecules within a given precision, by a set of wellchosen representative fragments. Such a set can be obtained by clustering a larger set of fragments whose structures have been solved experimentally, using suitable clustering algorithm and measure of dissimilarity between fragments. A commonly used measure of dissimilarity in structural biology is the root mean square deviation (RMSD), whose exact computation requires a pairwise structural alignment. But this alignment is highly time-consuming and not applicable for a very large initial set of fragments. We propose here an approach based on feature extraction to perform an effective clustering, while avoiding a computationally expensive full pairwise alignment. Using as example poly-A RNA fragments of 3 nucleotides (3-nt), we searched for internal coordinates whose differences can best approximate the RMSD between two fragments without any superposition. We found that the simple differences of internal distances and angles can provide a lower bound on the RMSD, allowing us to filter out pairs of which the RMSD does not need to be computed. We can then compute the exact values for only the small RMSDs, and use it to apply more effective clustering methods. We present this strategy and its application on 39431 RNA 3-nt, which could be approximated by only 3258 representative prototypes with 1 Å accuracy.