27 Sep

No prompt: facts in, typed questions out

A chat model returns prose that has to be parsed. It also reads the star rating, so it can't notice when the words disagree with it, which is exactly the case a business cares about.

So each review goes out as one call with 15 typed questions, and every answer comes back typed, with a probability, so code can act on it. Asking a chat model for JSON gives free text with no confidence per answer. One call per question costs 15× the calls for the same reading. And sending the stars too would make "stars vs words" impossible to measure.

What Jev receives

Only the product and the numbered sentences. No system prompt, no stars:

{
  "review": {
    "product": "Aero X2 Earbuds",
    "sentences": {
      "s0": "Oh great, the battery died mid-flight after three hours.",
      "s1": "Five stars for the lovely box, I guess."
    }
  }
}

Why

If the AI only judges and plain code does the counting, every number in the report can be traced and tested.