The usage of machine learning (ML) based solutions for finding defects on manufactured products has enjoyed plentiful attention from the academy and is being steadily deployed into industry. However, a common problem with ML models in general is the lack of their interpretability and adaptability to changing situations without an abundance of data. A popular task in defect detection is anomaly detection, with PatchCore being one of the most popular methods. It re-uses a pre-trained feature extractor, moving the optimization away from a statistical and towards a geometrical basis. It relies on modeling of a distance field of nominal samples using an automatically extracted point cloud of image patch features. Recently introduced SequentialPatchCore further enables efficient application to large image sizes and datasets. In addition, it introduced a method of melding coresets to rapidly join extracted knowledge from multiple domains without increasing computational and memory costs of the final model. Although the aforementioned models were presented for industrial inspection, they have no domain specific assumptions and can be applied to other domains such as medical, geographical and security. Explicit geometrical representation of knowledge enables simpler control over the model allowing us to directly influence the computational costs and model performance. Overall, this opens a new direction to explore geometrical analysis and modeling of the model knowledge and adapt the model given geometrical constraints.