The raw multimedia data is separated from its domain knowledge in this paper. Thus, different application domains can share the raw multimedia resources at great extent. By this way, the multi-view property of multimedia can be easily supported. In CDB/M, Frame is used to model the domain concepts, which are abstracted from multimedia data, and Sub Frame to represent the subconcept of another concept. Therefore, the PART_OF relationships among the domain concepts can be well modeled. Besides the PART_OF relationships, each concept may have independent attributes Mi and derived attributes Fj, which respectively represent user input annotations and visual features of multimedia data. However, different users may have large disparities in their specifying the similarity thresholds and fuzzy values of multimedia data. To solve this problem, CDB/M maintains a Context for each of the users of the system. As a matter of fact, the system provides a Context template, whose parameters can be input and modified by the individual users according to their special requirements. In many situations, a user needs visual similarities in sub-concepts other than the whole multimedia object. The hierarchical structures of domain concepts in CDB/M can easily support this problem. This paper also investigates multimedia query processing procedures. Two kinds of index, traditional B+ tree for simple comparisons and VP tree for visual similarities, are employed by the system in order to increase its searching speeds.