You have a Microsoft Foundry agent that answers questions about products.

Data acquisition: Production examples for high-volume product1 and synthetic examples for sparse product2; Curation: De-identify source records; Labeling: Use active learning to prioritize SME review.

Product1 represents a large share of real traffic, so production examples best preserve its true usage distribution. Product2 appears infrequently but has high escalation impact, making synthetic generation appropriate for filling the coverage gap without waiting for more production traffic. Because the available logs contain names and contact information, those records should be de-identified before they are reused for training or labeling. Finally, the SME budget is limited, so active learning should prioritize the examples where expert labels are most informative instead of reviewing a uniform random sample. Microsoft Foundry guidance supports synthetic fine-tuning data when real examples are sparse, and Microsoft healthcare/privacy tooling supports de-identification of sensitive identifiers. This design balances representativeness, rare-case coverage, privacy, and labeling efficiency. At implementation time, the same rule should be expressed through the framework or service configuration rather than left only as a natural-language convention. That makes the behavior repeatable across runs, easier to test, and less sensitive to model variability.

Official Microsoft reference: Microsoft Foundry - synthetic fine-tuning data generation

============================================================