Ask AI to extract

Asks a model to pull structured values out of text, shaped by a schema.

Ask AI to extract prompt into into

Details #

Gives the model your prompt (including the text to read) and a schema describing the fields you want, and returns an object that matches it.

Parameter	Type	Required	What it does
`prompt`	text	Yes	Required. The instruction and the text to read.
`schema`	record	Yes	Required. A JSON Schema object with `"type": "object"` and the fields under `properties`. The answer is checked against it. A bare `{"date": "string"}` is not a schema and fails.
`dryRunUsesModel`	yes / no		`true` lets a dry run call a cloud model.
`hud`	yes / no		`false` turns off the "Thinking…" HUD with the AI badge that shows while the model works. On by default; a dry run never shows it.
`into`	text		The variable for the extracted object.
`model`	text		`haiku` or `sonnet`, for the cloud; or a model name when `provider` is a provider id.
`preset`	text		An AI preset from Settings › AI (its id), or `default` for the default preset. Pick one in the Inspector's Model menu; renaming a preset changes nothing here, and a deleted one runs on the default preset with a warning. Without it the step behaves as before.
`provider`	text		`auto`, `device` or `cloud`; or a provider id (`anthropic`, `openai`, `gemini`, or an OpenAI-compatible server's id) together with `model`.
`sample`	any value		What a dry run answers when no model runs. Give it an object shaped like the schema.

Every step also accepts label, onFailure, timeout and dryRunSafe.

Pull a date and total out of text

Open in Switchboard to try it

Where it runs. provider is auto (the default), device or cloud. auto uses Apple Intelligence on this Mac when it is on and the request fits its small window (about 4,000 tokens), otherwise your cloud model (your own Anthropic key first, then your Switchboard plan). device never leaves the Mac and fails if it is off or the text is too long. cloud needs a model set up in Settings > AI. model (haiku, the default, or sonnet) only matters for the cloud.
Output. The answer goes into into. The step's own output also carries value, provider (device, own-key or plan), model, ms, tokensIn, tokensOut and cents, plus repaired: true when the first answer was rejected.
A bad answer. If the answer does not fit, the model gets one second try with the errors. If that fails too, the step fails and follows its onFailure.
Time limit. These steps get 60 s by default, not 10 s.
Dry run. These steps are safe in a dry run. The on-device model runs for real. A cloud model does not (it costs money): the step answers with sample, or an empty string, unless you set dryRunUsesModel to true. If no model is set up, a dry run uses sample too.
Text you send to a cloud model leaves your Mac. Do not send secrets or private data to it.
Read a field later as {{found.total}}.