Nordlicht Reisen GmbH (R-01) has ordered the travel assistant but has not chosen anything yet: no model, no search service, no memory store. This lesson does not ask how to deploy something (that follows in 1.2 and 2.1), but which service fits which requirement and which limit flips the choice. Everything here is paper: a requirements sheet, a decision note and a configuration sketch. You don't need an Azure account, and no costs arise.
Prior knowledge. How to tell large, small and reasoning models apart, read a model card and weigh Global Standard against Data Zone Standard is in "AI track, lesson T1.3". Retrieval and hybrid search are built in "AI track, lessons T4.4 and T5.4", memory in the Agent Framework in "AI track, lesson T7.5", tools through MCP in "AI track, lesson T8.1". None of that is repeated here. The lesson starts at the Foundry services and at weighing several requirements against each other.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 1.1-01 | Answers about the terms of travel name their source (R-02, R-03). | F | 95 % of answers with a verifiable source | Must | Business unit |
| 1.1-02 | Simple questions (booking status, opening hours) come quickly. | Q | median to the answer under 2 s | Should | Operations |
| 1.1-03 | Hotel photos get alt texts (R-04, R-05). | F | model understands images | Should | Editorial team |
| 1.1-04 | The assistant remembers a customer's language across conversations and can forget it (R-01). | F | deletion within 24 h | Could | Data protection |
| 1.1-05 | The assistant queries booking data through an API, read-only. | F | 1 tool to start with | Must | Operations |
| 1.1-06 | Inputs are processed in the EU (R-06). | R | deployment type with an EU boundary | Must | Data protection |
| 1.1-07 | The model changes by configuration (R-08). | Q | 0 code changes | Must | Development |
Cost frame. Cost drivers are tokens per million (input and output separate, per model), the tools (according to Learn, file search has additional fees besides the tokens), the search service per hour and tier and the memory store (price: open, preview). Prices are on the pricing pages that are linked under Further reading.