I see your point here but please take a look at the “standard” unstructured pdf extraction algos they have a lot of problems as well. Llm based extraction is still (on avergad) a big improvement.