We present a new provenance model for generating fine-grained data and service dependencies within XML data processing workflows. Our approach follows the widely used black box transformation semantics [15] in which service components produce new outputs from their inputs (without transformation). The heart of the model are data dependency rules which are evaluated on XML documents assembling all data produced by some workflow execution (similar to nested collections [5]). Dependency rules are defined in XPath extended with variables and can directly be compiled into XQuery expressions for generating provenance information in RDF-PROV [8]. We also present an implementation of our model, using the WebLab platform [19], showing step-by-step how our model works in a typical media mining use-case.