Question

You are building a Databricks application that uses a Foundation Model API chat endpoint for two distinct tasks: (1) extracting structured invoice fields (dates, amounts, vendor IDs) that must be identical across repeated runs of the same document, and (2) generating varied marketing taglines where diversity is desirable. You call the same underlying model for both. What is the most appropriate way to configure the model inference parameters for each task?

Accepted Answer

Set temperature near 0 for the invoice extraction task and a higher temperature (e.g., 0.8) for the tagline generation task

Answer

Set temperature to 1.0 for both tasks, since the model's training determines determinism regardless of parameters

Answer

Set temperature near 0 for both tasks and instead vary the max_tokens parameter to control diversity in the tagline task

Answer

Use a high temperature for extraction to capture all possible field variants and a low temperature for taglines to keep them concise