If you are generating documents you ignore all the stuff you don’t need. You are left with a sane dictionary based format with some optional compression. Ignore the compression until you feel the need to optimize. There is a bit of a complicated bit for random access where you need to remember and regurgitate the offset of various dictionaries. Beyond that it is just a bunch of drawing commands. Of course these are 20 year old memories and maybe the horror of part of it has burnt that from my memory.

I do agree that writing a reader would be brutal.