that is fine, I can use http raw request and then use transformResponse to parse the response, but how exactly promptfoo can do evaluation once I have a response? just based on the text response from the model or it can do something with the already scored strings "bad" or "good" in the response (the API returns a score for prompt and response string)?