The travel assistant is meant to answer questions about the travel terms and name its source (R-02, R-05). The knowledge exists as 120 PDF documents (R-03) and sits in the Azure AI Search index reisebedingungen of the search service srch-nordlicht (E-02). This lesson asks only one thing: How do the agent or the application reach this index from the Foundry side? How the index comes about (data source, indexer, skills, chunking, vectors) is taught in lessons 5.1 to 5.3; the search methods themselves in lesson 5.2. Here the index is assumed to exist.
Prior knowledge. Why a model does not know your own documents, how answers are grounded, sources named and questions without a source declined is in "AI track, lessons T5.1 to T5.3"; hybrid search and semantic ranking in "AI track, lesson T5.4"; embeddings and vector stores in "AI track, lessons T4.1 to T4.4"; foreign text as fenced data in "AI track, lesson T6.2", hard rules in code in "T6.4". The Foundry SDK and the project client are in lesson 2.1.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 2.2-01 | Every answer about the travel terms names at least one source with a link. | F | 100% of answers with a source | Must | Business unit (R-05) |
| 2.2-02 | Without a hit the assistant does not answer from the memory of the model but hands over to a person. | R | 0 answers without a hit | Must | R-02 |
| 2.2-03 | Search combines keywords, vectors and semantic ranking. | Q | hybrid, semantic (E-02) | Should | Business unit |
| 2.2-04 | Agent and application access the search service without a key. | R | 0 keys | Must | R-07 |
| 2.2-05 | Documents changed three months later appear without a code change. | Q | only the index changes | Should | Operations (R-03) |