The group-travel team of Nordlicht Reisen GmbH makes up to 20 model calls per request (lesson 0.2), and the requests arrive through the queue anfragen at the Container App app-gruppenreisen (M-10). 25 requests per week are expected (M-01), but a load spike multiplies the calls quickly. The lesson asks: What may a run reuse without saying something wrong or mixing tenants? And: How many runs start at the same time, and who limits that?

Prior knowledge. "AI track, lesson T3.4" explains the prompt cache of the providers in detail: stable start first, cache killers, measuring through the cached input tokens, keys and breakpoints for newer model families. "T2.4" shows middleware around the chat client, including UseDistributedCache for responses. "AI-103, lesson 1.3" calculates quota and throughput. None of that is repeated here. New is the multi-agent situation: seven instructions instead of one, branches that run at the same time, several tenants in one application and a queue in front of it all.

New in this lesson. Assumptions of the task: The result of lieferanten/verfuegbarkeit_lesen may be cached for 2 minutes; lieferanten/vorreservieren (A2) is never cached. A replica handles at most 4 runs at the same time; there are at most 6 replicas; a run has at most 2 concurrent branches (unterkunft, transport, lesson 2.8). A semantic cache does not appear in the case.

The requirement

All numbers are assumptions of the task.

No. Requirement Type Measure Weight Source
2.10-01 No cache delivers data of one tenant to another. R 0 hits across tenant boundaries Must M-07
2.10-02 Calls that change things and price calculation never run from a cache. R 0 hits for A2 and rechner Must M-05, M-03
2.10-03 A replica accepts at most 4 runs at the same time. R 4 Must Quota
2.10-04 The highest number of concurrent model calls is below the quota of the deployment chat. R 48 (calculation below) against quota Must R-08, M-11
2.10-05 Requests are not lost during a load spike. R 0 lost messages Must M-10