another approach is HyDE, where you ask the LLM to come up with a plausible (but likely wrong) answer, and use the embedding of the wrong answer to find the app

another approach is HyDE, where you ask the LLM to come up with a plausible (but likely wrong) answer, and use the embedding of the wrong answer to find the appropriate chunk, pretty clever