This is fascinating. I’m struggling to understand how that was causing such a large difference in the output. Is the “autoparser” vulnerable to injections somehow? How do you distinguish between user text, model text, and metadata, or is there ambiguity in the parsing?