You say that like the LLM must be built and retrained on ads rather than a separate scan of the prompt