---
title: 'Structured Output'
description: 'Force the AI to return a JSON object matching a schema you define, validated before it flows downstream.'
---
# Structured Output

Sends data to an LLM and forces the response into a **JSON object matching a
schema you define** (property → type). The result is validated against that
schema before it flows downstream, so nodes after it can rely on the shape.

## When to use

- Pull a fixed set of fields out of unstructured text — e.g. extract
  `{ first_name, last_name, dob, doc_number }` from a KYC submission.
- Any time you want the AI's answer as **predictable JSON** rather than free
  text, and you want a validation signal you can branch on.

For free-form text (summaries, rewrites) use **AI Transform**. For picking one of
several categories use **AI Classify**.

## Configuration

| Field | Required | What it does |
|---|---|---|
| `instruction` | Yes | What to produce. Supports `{{...}}` refs. |
| `inputField` | No | The data to work on. Blank = the upstream node's output. |
| `schema` | Yes | The target shape as a property → type map. Types: `string`, `number`, `boolean`, `object`, `array`, `any`. |
| `provider` | No | `groq` (default), `gemini`, or `openrouter`. |
| `model` | No | Provider model ID. Default `llama-3.3-70b-versatile`. |

### Example schema

```json
{
  "first_name": "string",
  "last_name": "string",
  "date_of_birth": "string",
  "document_number": "string",
  "verified": "boolean"
}
```

## What it outputs

```
{
  result: { first_name: "…", last_name: "…", verified: true },  // the parsed object
  valid: true,          // did it match the schema?
  errors: [],           // list of mismatches when valid = false
  model: "llama-3.3-70b-versatile",
  provider: "groq"
}
```

Downstream nodes reference `{{structured_output.data.result.first_name}}`. Branch
on `{{structured_output.data.valid}}` with an **If / Else** to handle the case
where the model couldn't produce a conforming object.

## Gotchas

- **Validation is shape + type, not business rules.** `valid: true` means every
  declared property is present with the right JSON type — it does not check that
  a date is real or an amount is positive. Add downstream checks for that.
- **Keep the schema flat where you can.** Deeply nested schemas are harder for
  the model to satisfy reliably.
- **Non-determinism still applies** — the values can vary run to run even though
  the shape is fixed.