I have thousands of structure charts created in PowerPoint that represent the relationships between different nodes defined by PowerPoint shapes using edges defined by PowerPoint connector lines. I need a way to programmatically extract all nodes and their associated relationships either by accessing PowerPoint coordinates or doing image processing. However, I do not know how to proceed.

1. I tried using some popular llm vision models but they proved incapable of the task as the charts are often quite complicated, and llm capabilities are just not there yet. I also tried a programmatic algorithm based on PowerPoint coordinates but found there to be far too many edge cases to account for. I am now thinking about using some combination of open-cv and a graph neural network to parse them, but I am not finding very much information online at all about such a use case. Any help at all or any resources anyone could provide would be greatly appreciated.