Shape matching is one of the crucial problems in 3D model retrieval system. A usual approach is to exhaustively search through the database comparing each database model with the query model. This approach is inefficient especially for a large database. In this paper, an efficient mechanism for 3D model retrieval is proposed. In the preprocessing stage, a set of reference models are selected from the database using cluster analysis, and distances between database models and reference models are computed and stored. Then in the query stage, for a certain query model, searching is accelerated by reducing the large amount of model comparisons using triangle inequality based on the reference distances computed before. The proposed retrieval mechanism is implemented and experiment result shows that retrieval efficiency is greatly improved than the usual approach without any precision loss.