Structured output Ai Questions

Design the schema like a form for a new teammate: clear key names with units, a short description on every field, enums instead of free text where the values are known, and an explicit null or "not found" option so the model never has to invent data. Keep it shallow, order fields so context comes before conclusions, and stay inside the schema features your provider supports. Then measure it with a small set of test inputs instead of guessing.

JSON mode only promises that the reply parses as JSON, not that it has the fields you need. Structured output goes further and constrains the reply to a JSON Schema you send, so the shape is enforced at generation time. Function calling uses the same schema machinery for a different job: the model asks your code to run a tool with typed arguments, instead of answering the user in a fixed shape.

Treat model output like any untrusted API response: parse it against a Zod schema at the server boundary before it reaches React. Use safeParse so a failure becomes a value you can branch on, log the raw text and the Zod issues, and return a typed error state the UI knows how to render. With the AI SDK, Output.object runs this validation for you and throws NoObjectGeneratedError when the object does not match.

You have four main options: wait for the full text, repair the partial text on every chunk (close open strings and brackets, then parse), stream complete elements one at a time, or switch the wire format to one object per line. The AI SDK gives you the repair approach as partialOutputStream on the server and useObject on the client, and elementStream for arrays. Partial objects cannot be schema-validated, so treat every field as possibly missing and validate once at the end.

Decide by the kind of failure. Retry transient errors like timeouts, 429s and 5xx with backoff, because the same request can succeed next time. Repair when the content is close: fix small problems in code first, then re-ask the model once with the validation errors. Fall back when retrying cannot help, such as a refusal, a content filter or repeated failures, and always cap attempts because every retry costs tokens and time.