not really an LLM thing, more of a computer vision thing
but I think this would run into some pretty big problems because a computer can't infer context the way a human can. like, suppose your schematic is spread across multiple pages with some indication of how the pages relate to each other. a human can handle that without really thinking about it, but a computer would probably choke completely
if the data was in a computer-friendly format, like a graph (in the nodes and edges sense), and if you associated the nodes with local geographic data, then I think you could have a pretty handy little computer assistant. but converting the schematics into graphs would require the human touch even if it was largely automated, and then you'd want to make sure the schematics were actually correct, and that the data didn't go stale with undocumented changes to the system...