Customers of Nordlicht Reisen GmbH upload photos and scans of receipts: hotel invoices, restaurant receipts, tickets (R-04). What is needed are merchant, date, amount and receipt type as fields, not as running text. According to E-07, Content Understanding takes this over; the travel terms stay with the indexer and skillset (5.1, 5.3) and are only compared here. All receipts in the exercises are made-up samples (R-12), never real ones.
Prior knowledge. Structured output from text is shown in "AI track, lesson T2.3", foreign text as fenced-in data in "T6.2", hard rules in the code in "T6.4". That Content Understanding is not a tool of the Agent Service and that you wrap a function around the extraction is in lesson 2.4; images and video are covered by 3.4, audio by 4.3 and 4.4.
New in this lesson (does not change R-08, but extends the case). The region list of Content Understanding does not contain germanywestcentral. For the receipts there is therefore a second Foundry resource ai-nordlicht-cu in a listed EU region (swedencentral or westeurope). It carries the model deployments that Content Understanding needs; ai-nordlicht stays in germanywestcentral.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 5.4-01 | From receipts come merchant, date, total amount and receipt type as typed fields. | F | 4 fields | Must | R-04 |
| 5.4-02 | Uncertain fields go to a person. | R | confidence below 0.8 or missing | Must | R-02 |
| 5.4-03 | Processing and storage stay in the EU. | R | listed EU region, processing location set | Must | R-06 |
| 5.4-04 | Access needs no key. | R | 0 keys | Must | R-07 |
| 5.4-05 | Pages and tokens per receipt are traceable. | Q | usage information stored per call | Should | R-09 |
| 5.4-06 | A PDF can be read as Markdown with tables (comparison). | Q | tables stay tables | Could | E-07 |