The editorial team of Nordlicht Reisen GmbH (R-01) uploads hotel photos (R-04). Every photo needs an alt text for screen readers (R-05), and in the portal customers are to be able to ask questions about a photo ("Does the room have a balcony?"). This lesson connects both with the deployment chat (E-05). Paper: requirements, decision, compiled code and a check function. Only the optional cloud step needs an Azure account. Cost drivers: tokens per million, plus image tokens per image; the accuracy level detail changes them (see below). All photos are made up and show no recognizable people (R-12).
Prior knowledge. Input to models and IChatClient are in "AI track, lesson T2.1", structured outputs in "T2.3", evaluation with a judge in "T9.2 and T9.4", foreign text as data in "T6.2". Model choice by capability is in "lesson 1.1", the Foundry SDK in "lesson 2.1" of this course. Lesson 1.1 handed this lesson the question whether the model behind chat understands images.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 3.3-01 | Every hotel photo gets an alt text in the language of the page. | F | 100 % of published photos | Must | Accessibility |
| 3.3-02 | Pure decoration gets an empty alt text. | F | 0 decorative images with text | Should | Accessibility |
| 3.3-03 | Alt texts describe only what is visible; no invented names or places. | Q | at most 2 of 40 test cases with invention | Must | Editorial team |
| 3.3-04 | Questions about a photo are answered only from the photo, otherwise "not recognizable". | R | 0 answers about the invisible in the test set | Must | Legal |
| 3.3-05 | Text in the photo (signs, prints) is content, never an instruction (R-11). | R | 0 obeyed instructions in the test set | Must | Security |
| 3.3-06 | The call runs through the counter and the guardrails (R-11). | R | 100 % of calls | Must | Operations |