Skip to main content

State

The state is everything Super Decider reads before it answers: a support ticket, a database record, a chat history, a contract, a scanned invoice. Questions are asked about the state.

The model only knows what is in the state and in the questions. If an answer depends on a policy, a date or a customer's history, put it in the state.

Formats​

The state can be:

  • Text: a plain string.
  • JSON: any JSON value. Use objects to keep fields apart, such as ticket, customer and policy.
  • Images: up to 8 images sent alongside the state. See Images.
{
  "model": "super-decider",
  "state": {
    "ticket": "I was charged twice for my March subscription.",
    "customer": {"plan": "Pro", "tenure_months": 26, "open_tickets": 0},
    "policy": "Duplicate charges are refunded automatically within 7 days."
  },
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "What is the customer asking for?",
      "criteria": {"refund": null, "cancel": null, "technical_help": null, "other": null}
    }
  }
}
Chat completions

On the chat completions endpoint the state goes in the user message and must be JSON-encoded, even a plain string. In Python, wrap it with json.dumps(state). Raw text returns a 400 error.

Size​

A request can hold up to 131,072 tokens, state and questions together. A state that is too long returns a 400 error with the reason.

Input tokens are billed, so trim the state to what the questions need. A shorter state is also faster.

Writing good state​

Include the facts the decision depends on. If the answer turns on a refund window, state the window and the delivery date. Don't assume the model knows your rules.

Label each part. With JSON, field names such as policy, customer_message and order tell the model what each part is. With text, use short headings or prefixes such as Policy:.

Keep the same structure from request to request. Answers can shift when the input changes in ways that keep its meaning, such as the order of JSON keys. Build the state the same way every time, for example with a fixed function that writes keys in a fixed order.

Write in English. The model's training data is English-centric.

Do the arithmetic in code where you can. The model is sometimes confidently wrong on multi-step date and money calculations. Compute values such as "days since delivery" or "amount outstanding" before you send the state. When you can't, turn on thinking.

State and span questions​

Span answers point back into the state with character offsets. When the state is a string, offsets count from its start. When the state is JSON, offsets count into its text field, so put the text you want quoted from in a field named text.

Next steps​