Would you be able to parse it in a streaming fashion and just store the structure of the document in memory, with just offsets for all of the string locations, and then re-read those from disk as needed?

With modern SSDs and disk cache, that's likely enough to be plenty performant without having to store the whole document in memory at once.