The Primula system is a versatile software tool for modeling and reasoning with probabilistic relational structures based on the symbolic Relational Bayesian Networks (RBN) language. The new version 3 of Primula extends previous versions by adding support for categorical variables, and by integrating Graph Neural Networks (GNN) as model components into a full generative RBN model, thus combining the predictive power and scalable learning tools of GNNs with the high expressivity and flexible inference capabilities of RBNs.