Question

You are building a text analysis pipeline that ingests free-form customer support emails and must produce a strictly typed JSON object for each email containing fields for issue_category (from a fixed list), urgency (integer 1-5), and a one-sentence summary. Downstream systems reject any output that is not valid JSON matching the expected schema. Using an Azure OpenAI chat model in Azure AI Foundry, which approach best guarantees the model returns parseable, schema-conformant output?

Accepted Answer

Enable structured outputs (JSON schema / response_format) that constrains generation to the defined schema, and provide the enum values for issue_category in the schema.

Answer

Set temperature to 1.0 and instruct the model in the system prompt to 'try to return JSON', then parse the response with a lenient regex fallback.

Answer

Append the raw email to a prompt asking for a summary only, then run a second Azure Translator call to normalize the text into JSON.

Answer

Use the Azure AI Language key phrase extraction API and map the top three key phrases into the JSON fields.