Been building various LLM+PDF pipelines at work. As soon as you need to e.g. parse tables etc. it becomes a lot of hard work!