Structured Outputs

A model produces tokens. Your code needs a record: an order with a numeric quantity, a ticket with one of four priorities, a date that a database will accept. Somewhere between those two facts sits the single most common source of breakage in an LLM application — output that looks right to a human and fails JSON.parse.

There are three layers. Prompt and parse states the contract but relies on the model following it. Constrained decoding enforces a supported grammar or schema during generation. For a successful, complete generation within that supported subset, malformed structure can be ruled out; refusals, truncation, unsupported schema features, and provider failures still need handling. Validate and repair checks the result against your own contract and chooses what to do on failure.

Combine explicit instructions, constrained decoding where available, and validation. Structural enforcement does not remove semantic errors. A schema that says quantity: number may accept -4, and status: string accepts "probably shipped". Meaning and business policy remain yours.