Retrosynthesis involves the strategic breakdown of complex molecules into simpler precursors, paving the way for the synthesis of novel molecules. Recently, there has been a development of AI-based methods for retrosynthesis, which allow learning reaction rules from the data of historically performed reactions. A central component of such systems is a model for single-step retrosynthesis that predicts what reactions could lead to a considered target molecule. Two dominant methodologies are used for single-step retrosynthesis. Template-based methods use a set of translation rules that represent the possible chemical transformations. Although these methods are characterized by speed and interpretability, they may require an extensive set of templates to cover a large space of chemical reactions, which limits their generalization capacity. Conversely, template-free approaches can produce arbitrary reactions without such constraints but are often computationally demanding, largely due to their dependency on autoregressive decoding [1, 2, 3, 4].