I’m testing DeepSeek for a workflow where each response needs to fit a JSON schema, including required fields and a few allowed values. The tricky part is that I also want the model to reason through ambiguous input before returning the result.

Has anyone found a reliable way to get strict JSON output from a DeepSeek reasoning model without extra text, missing keys, or a different structure from one response to the next? I’d be interested in whether this depends more on the model, the API options, or how the schema is described.