The generation of Prior Samples is a way to define Dirichlet Priors. The expert prior knowledge is represented by a fully specified Bayesian network.
Basically, the definition of Prior Samples can be considered as sampling particles from the joint probability distribution encoded by the current network, and saving these particles in a Virtual Data Set. The actual data set and the virtual one are then used for machine learning the structure, and estimating the marginal and conditional probability distributions. As of version 9.0, this feature has then been moved to the Data menu.