Might be useful to have an LLM generate a few versions of the query in order to account for imperfect recall — it seems like something gpt-3.5-turbo and friends

Might be useful to have an LLM generate a few versions of the query in order to account for imperfect recall — it seems like something gpt-3.5-turbo and friends would be pretty good at doing.