Nordlicht Reisen GmbH (R-01) asks what its travel assistant is to be as a whole. Every lesson answered one question: model, index, agent, image, speech, receipt. Now the answers lie side by side, and in some places they do not fit together. This project arranges them into a decision folder, resolves the conflicts, draws one architecture overview and writes one Bicep sketch. It stays paper: Nothing is deployed, no command runs against Azure.
Prior knowledge. All lessons of modules I1 to I5, plus the case R-01 to R-12 and E-01 to E-08 from lesson 0.1. They are not repeated here: The folder refers to them by number. Recording decisions is taught in "K4, lesson 4.1.3", presenting an architecture in "K4, lesson 4.1.4". The sister version for Azure architecture is "AZ-305, lesson 7.1".
New in this lesson. (1) There are three resources: ai-nordlicht (location, preferably germanywestcentral), ai-nordlicht-cu (Content Understanding, 5.4) and ai-nordlicht-medien (only bild and video), the last two in swedencentral. The name ai-nordlicht-medien is new; it resolves K-2 below. (2) The practice group is called rg-ai103-abschluss. (3) The sheet A-01 to A-08 and the ranking below hold only for this project.
The decisions of the lessons
The column "changes" names where a lesson deliberately changes an earlier assumption.
| No. | Area | Decision | Lesson | changes |
|---|---|---|---|---|
| D-01 | 1 | chat, chat-klein, embed as Data Zone Standard; models only in the parameter file |
1.1, 1.2 | |
| D-02 | 1 | Bicep per environment; a new agent version per run, live only after a green evaluation | 1.2 | |
| D-03 | 1 | Distributing instead of inflating: simple questions to chat-klein; smoothing in the client; alert on 429 |
1.3 | |
| D-04 | 1 | disableLocalAuth, Foundry User, policy against GlobalStandard; private endpoint in production |
1.4 | E-08 |
| D-05 | 1 | Guardrail reiseassistent-streng on models and agents; rules stay code |
1.5 | |
| D-06 | 2 | Direct calls through IChatClient, agents through the project client; package series 2.x |
2.1 | E-04 |
| D-07 | 2 | RAG: search tool of the Agent Service, citation check in the code | 2.2 | K-4 |
| D-08 | 2 | Prompt agent reiseassistent; version only after the test set; no memory (Could) |
2.3, 1.1 | |
| D-09 | 2 | buchung_api (OpenAPI, read-only), stornieren (MCP, with approval) |
2.4 | K-3 |
| D-10 | 2 | Orchestration in code with Agent Framework, pruefer, no hosted agents; packages 3.0.0-beta |
2.5 | K-1 |
| D-11 | 2 | Test set locally and as an evaluation against the version; ongoing sampling | 2.6 | |
| D-12 | 2 | First prompt, then parameters, then one revision per answer | 2.7 | 2.5-05 |
| D-13 | 2 | Switch Mensch/Klein/Gross; path as a tag on the span |
2.8 | K-5 |
| D-14 | 3 | bild in a batch from the code, video with a queue; swedencentral |
3.1, 3.2 | R-08 |
| D-15 | 3 | Alt text from chat (E-05); prebuilt-imageSearch, prebuilt-videoSearch |
3.3, 3.4 | |
| D-16 | 3 | Three layers for images: check, text in the image as foreign text, rules for actions | 3.5 | |
| D-17 | 4 | Model with schema, PII through Language, Content Safety, path to a person as a rule | 4.1 | |
| D-18 | 4 | Translator as the default; Danish only after approval by a person | 4.2 | |
| D-19 | 4 | Fast transcription and batch; word list before custom speech | 4.3 | |
| D-20 | 4 | reiseassistent-telefon (Voice Live, preview) as a prototype, own agent |
4.4 | E-06 |
| D-21 | 5 | Indexer hourly; hybrid with semantic ranking; skillset ss-reise |
5.1 to 5.3 | E-02, E-07 |
| D-22 | 5 | Receipts with Content Understanding in ai-nordlicht-cu |
5.4 | R-08 |
| D-23 | 5 | In the workflow the retrieval is a step of the application with source ledger | 5.5 | D-07 |