Leveraging deep learning models to capture geometric features has received a great deal of research interest. Most of these models can easily handle relatively simple vector graphics, but are often helpless when faced with complex geometric structures due to the thousands of geometric elements. We propose a novel Graph Structural Similarity Network (GSSN) to solve the problem of high-complexity geometric graph retrieval. Unlike existing graph similarity neural networks, GSSN does not rely on any explicit labels during training. It uses contrastive learning to evaluate the similarity of a random pair of graphs. When calculating the similarity, we adopt a block-by-block approach at data input, mapping high-complexity geometries into multiple low-dimensional latent space vectors, and constructing the final representation vector by feature alignment and splicing to calculate the similarity of complex geometries. Specifically, we evaluate GSSN on a self-processed Computer Aided Design (CAD) drawing dataset and propose a flexible matching approach for labelling differences between these retrieved similar CAD drawings. In the retrieval task, GSSN achieved an impressive F1-score of 87.04%. For the matching task, the approach attained an accuracy of 96.11%, validating the accuracy and availability of GSSN.