The pilot customers of Nordlicht Reisen GmbH sometimes find the answers of the travel assistant too long and, for the same question, not always the same (assumption). The technology has many knobs for this, and each one acts differently. This lesson puts them in order by effect and cost and shows a loop in which a reviewer critiques the draft before a customer sees it.
Prior knowledge. "AI track, lesson T1.1" explains what a language model does when it generates, "T3.1 to T3.3" system prompts, examples and the versioning of prompts, "T7.6" the reviewer agent, "T9.2" measuring with a test set. That is not repeated. New: the parameters of the Foundry page, the particularities of reasoning models according to Learn and the loop of draft, critique and revision.
New in this lesson. The parameters of a deployment live in the configuration like its model (section Abstimmung, R-08), never in code. A revision after the critique of the pruefer is allowed, at most one per answer. That changes the assumption 2.5-05 (two calls): Only an objected draft costs a third call.
The requirement
All numbers are assumptions of the exercise.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 2.7-01 | The same question leads to the same statement in ten repetitions. | Q | at least 9 of 10 | Should | Customers |
| 2.7-02 | Answers stay short enough for web and phone. | Q | at most 600 output tokens | Should | Customers |
| 2.7-03 | Every change to the prompt or parameters is measured with the test set before it goes live (R-10). | R | Run with 40 cases | Must | Quality |
| 2.7-04 | The loop has a limit for checks and a token cap (R-09). | R | 1 revision; cap per answer | Must | Cost |
| 2.7-05 | Hidden reasoning is neither stored nor displayed. | R | 0 stored reasoning contents | Must | Data protection |
Cost frame. Cost drivers: tokens per million, and every round of the loop is another model call with its waiting time; reasoning models count their hidden reasoning too. Prices: Foundry pricing page.