💭 also if you have n posts and two of them are "duplicates" or closely related in a way only a smart llm could detect, I think thats provably very expensive to detect? That's kind of what graph formation is about. I guess it'd be iterative and you'd not assume to find all connections but maybe fine obvious connections and gradually refine. Things that share 1 connection are more likely to share more so you could get more efficient that way.