In the context of the study and enhancement of historical architectural heritage, the construction of a cognitive model represents a crucial step in delineating the identity of a work, interpreting its compositional dynamics, and defining methods of intervention. According to Gestalt theory, an observer does not merely perceive individual elements but tends to reconfigure the whole into meaningful and coherent forms. This cognitive approach, when applied to computer vision, can be leveraged for the study of architectural heritage through processes of information acquisition, processing, and abstraction. Consequently, a laser survey not only represents a point cloud accurate to the original geometry but can also become an interpreted and semantically organized information model. In this study, a computational analysis methodology is proposed in which the point cloud is submitted to a hierarchical decomposition process, also grounded in perceptual criteria, through the application of foundation models. These models are generalized learning frameworks based on pre-trained deep learning neural networks, specifically designed for the computer vision domain. The research aims to verify, inside the Palazzo Fizzarotti building in Bari, an eclectic structure inspired by Venetian Gothic style, the ways in which these digital segmentation models can be integrated to support the understanding and study of a complex architecture, as a preparatory phase for subsequent operations of reading, representation, and information modeling.