Wouldn't this become pretty difficult to do at scale over time? Is there a way to compare the similarity in a database of responses without doing a search over every entry and comparing them? Because that would probably become pretty slow if literally every LLM output is saved and has to be scanned.