The segmentation of triangular mesh model in CAD model reconstruction cannot embody significance. To solve this problem, the knowledge base composed by basic modeling features and machining features was constructed to provide prior knowledge for segmentation. Firstly, depending on the fitting error of quadric surface and curvature, every patch of the model corresponding to the surface was extracted successively and the significance of segmentation was reflected by the primitive design element of surface. Then, both features in the knowledge base and the surface set of preliminary segmentation were represented by the attributed adjacency graph (AAG). Finally, the AAG of the features were applied to match the parts of the model which had the isomorphic AAG. This embodied more meaning from the view of engineering semantics. The experimental results demonstrate that the proposed efficient algorithm can get meaningful segmentation.