Question

A developer is building a support-ticket categorization feature using an Amazon Bedrock foundation model invoked through the Converse API. The model must classify each ticket into one of 12 fixed categories, but early tests show it frequently invents categories that are not in the approved list and formats output inconsistently. The team wants the fastest, lowest-effort change that reliably constrains the model to the exact 12 categories without training a custom model or adding external infrastructure. Which approach should the developer implement?

Accepted Answer

Add a few-shot prompt template that enumerates all 12 categories with a strict instruction to return only one category from the list, and include several labeled example inputs and outputs demonstrating the exact expected format.

Answer

Fine-tune the foundation model on thousands of labeled tickets so it learns the 12 categories, then deploy it with provisioned throughput.

Answer

Raise the temperature and top-p sampling parameters so the model explores more category options before selecting one.

Answer

Wrap the model behind a Bedrock Agent with an action group that validates the output, retrying invocation until a valid category is returned.