Question

You are refining a customer support summarization prompt in Azure AI Foundry prompt flow. Business stakeholders complain that summaries are inconsistent and occasionally invent details not present in the source ticket. You want to systematically test whether reducing model randomness improves factual consistency while keeping all other prompt wording identical, and you need to compare the results before committing a change. What is the most appropriate first step?

Accepted Answer

Create prompt variants of the same node that differ only in the temperature and top_p sampling parameters, then run a batch evaluation to compare their outputs against a groundedness metric.

Answer

Increase the provisioned throughput units on the deployment so the model has more capacity to produce accurate summaries.

Answer

Deploy each candidate prompt as a separate serverless API endpoint and split production traffic 50/50 to compare live results.

Answer

Rewrite the system prompt to add stricter instructions and immediately merge it to the main branch after a quick manual spot check.