In this paper, a new assembly retrieval approach is presented, based on which, users can input flexible queries, either rough or precise, to retrieve efficiently the whole or partial assemblies they want from the product library. First, a multilevel assembly descriptor supporting various searching requirements is elaborated, which collects different levels of information in assembly models. Then, the corresponding matching and similarity assessment methods with well-balanced efficiency and discriminability are given to evaluate the differences between assembly models. Moreover, an indexing mechanism for accelerating assembly retrieval, especially the partial retrieval, is presented to filter the unmatchable models quickly. Finally, an assembly retrieval prototype system is implemented, and the experimental results are analyzed to verify the advantages of the flexible assembly retrieval approach.