(One reason I think nobody seems to have tried implementing this prior to LLMs is that parsing arbitrary PDFs well enough to trust money to the results was almost intractably hard, and now it's something which is probably single digit engineers times single digit weeks to build.)

→ View original post on X — @patio11