LLMs still beat a clarifier, because they're able to extract more signals than a text embedding.

It's very difficult to beat an LLM + prompt in terms of semantic extraction.