Bridging the semantic gap between signal level and semantic level descriptions is the main problem to solve for multimedia indexing. The most advanced approaches take into account several descriptor types, several modalities and/or the context to enhance concept detection. For mastering the complexity linked to the integration of the heterogeneous data and processes that this requires, we propose an approach based on operators organized into dataflow networks. A unique type of data, the numcepts, and a unique type of processing unit, the operators, are used at all levels in these networks. This approach offers a great flexibility for the design of indexing systems and it eases the experiments with the corresponding architectures. We describe an instance of this model and several numcepts and operators. We have tested the approach on TRECVID 2004 and 2005 corpora. We have implemented 7 of operators and studied the influence of variations on the accuracy of concept detection