CAD standard parts have been widely used in designing process for accelerating designers’ workflows. Generally, the access to retrieve CAD standard parts relies on catalogs provided by some manufacturers or software. Considering that catalogs might not update timely and might be confusing for users without corresponding specific knowledge, this paper proposes a 2-stage method to retrieve parts based on sketches. An unsupervised reranker is added to the traditional retrieval process, which alleviates the problem of unbalanced data and provides optimal ranking of the retrieval results. Further, considering that the training results of a triplet-loss network might be unstable since input samples are fed randomly, a method to control the selection probability of negative samples and a new training workflow are designed in this paper. Meanwhile, because of the scarcity of a dataset which contains CAD standard parts classified according to their mechanical definition and their free-hand sketches, a new training dataset is constructed firstly in this paper. Finally, in experiments, we compare the retrieval methods with and without reranker, and the workflows with and without the negative sample selection strategy, and verify the method proposed in this paper on a public dataset. (The CAD standard part dataset and the code are at https://github.com/dishy313/CADStdPartRetrieval .)