PRODUCTION SCENARIO
A Foundry prompt lists six ticket categories in its system message, but the model still labels the same ticket differently across runs and returns category names that are not on the list.
What should the developer add to the request?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Example user and assistant messages after the system message
Few-shot examples condition the model to respond as desired for the current inference, and with the Chat Completion API they go in the messages array as example user and assistant interactions after the initial system message. A lower temperature steadies run-to-run variance but does nothing about invented labels. Forbidding off-list names is an instruction, not a demonstration. Repeating the list can help with recency bias, but it still only tells the model the labels rather than showing it how to pick one.