Extract or produce a JSON response

Use an llm step with output.schema when a support workflow needs an application-specific JSON value, such as a reference and requested action. The schema defines the result shape; it does not make unsupported facts true. For fixed business questions with answerability and issue codes, use a decision.

List extract in config/support_triage/triage/flow.yaml, then save:

# config/support_triage/triage/extract.step.yaml
type: llm
input:
  message:
    pointer: /payload/message
instructions: Extract the invoice reference stated in the email. Use null if absent.
output:
  schema: invoice.schema.json

Place the schema beside the step:

{
  "type": "object",
  "properties": {
    "invoice_reference": {"type": ["string", "null"]}
  },
  "required": ["invoice_reference"],
  "additionalProperties": false
}

For “Please send invoice INV-42,” the completed public record at /flows/triage/steps/extract can contain "result": {"invoice_reference": "INV-42"}. The value type is a JSON object. If no reference is stated, the validated result is {"invoice_reference": null}; it is still a completed LLM step. A downstream policy can decide whether that absence needs review.

input is required and may be empty. Without prompt, the named inputs are sent as compact JSON. If you add a prompt, each {{ name }} placeholder must name a declared input; inserted values are rendered once as JSON, and template syntax inside values is not evaluated. {{{{ and }}}} render literal double braces. Bind an earlier result explicitly, for example pointer: /steps/classify/result/answer/optionId; there is no implicit conversation history or forwarding of the whole envelope.

Use a local schema path or inline schema. Local $ref resources must stay in the bundle; remote references and escaping paths fail compilation. The selected model must declare JSON Schema capability. Set defaults.model on the workflow or model here; see models. A Markdown .step.md file can put configuration in front matter and instructions in the body, with exactly one instruction source.

max_iterations defaults to 4 logical model turns (1–1024), including a final answer turn. A plain extraction normally uses one. Add max_iterations: 1 on this step when you want to disallow additional model turns; provider retries of one request do not count as new iterations. The separate provider-attempt and deadline limits are in execution limits.

Malformed output, schema violations, provider refusal, and truncation fail with invalid_output; timeouts and exhausted budgets are technical failures. This step requests needs_review only if an attached MCP tool asks for caller input. Test extraction accuracy against known messages using task scoring. For model-selected lookups, continue with agent loops; for prose, see text response.