An LLM-vector-based search engine has the potential to be fantastic. No autocomplete - just make it search for similar vectors.