📐

Structured Output

OrchestrationReceiving results through a schema instead of free text

Structured output means asking the model for results that fit a schema or object shape instead of unconstrained prose. The goal is not just readability for humans. It is predictable consumption by code.

▶Architecture Diagram

🔄 Process

✍️Prompt

📐Schema

🤖Model

✅Validator

🔁Retry/Fallback

⚙️App Logic

Dashed line animations indicate the flow direction of data or requests

Why do you need it?

If outputs only arrive as natural language, equivalent answers can still break parsers because they vary in wording and structure. A system might understand that "high priority" and "priority: high" mean the same thing, but application code often cannot. Automation becomes fragile when output format is left ambiguous.

Why did this approach emerge?

When LLMs were used mainly for human-facing chat, slight wording variation was acceptable. As they moved into extraction, tool calls, UI generation, and system workflows, output shape had to become an explicit contract rather than an afterthought.

How does it work inside?

The application defines fields, types, required values, and constraints in a schema. The model produces an object that should satisfy that contract, and runtime validation checks whether it actually does. If it fails, the system can retry or fall back in a predictable way instead of relying on ad hoc text parsing.

In Code

Lock the output contract first

{
  "type": "object",
  "properties": {
    "priority": { "type": "string", "enum": ["low", "medium", "high"] },
    "reason": { "type": "string" },
    "needs_review": { "type": "boolean" }
  },
  "required": ["priority", "reason", "needs_review"],
  "additionalProperties": false
}

The point is to hand the model a field contract with explicit types and allowed values, not to hope that prose will look structured enough.

Only pass validated objects downstream

const result = await runModel({ prompt, schema });

if (!result.valid) {
  return retryOrFallback();
}

handleTicket(result.output);

Structured output is not just JSON-looking text. It is a flow where only validated objects are allowed into application logic.

Boundaries & Distinctions

Prompt engineering, structured output, and tool use all connect model output to a system, but they stabilize different parts of that path. If decision criteria are weak, look at prompting. If field shape and required values are weak, look at structured output. If the validated result must query or change another system, look at tool use. A valid schema match does not guarantee that the content itself is correct.

When should you use it?

Teams use this for extraction, classification, UI state generation, and tool argument creation. Once fields are stable, quality measurement also becomes easier because systems can see exactly which field is missing or wrong. If the schema is too loose, the system drifts back toward brittle interpretation.

Field extractionTool argument generationUI state generationClassification pipelines