This is an ugly half-solution. I want a vision model that only rips out the significant text content and related images from the page and displays them in a lovingly typeset emacs org-mode buffer.