Skip to main content
Text is the base of every decision. Put the material to judge in state and the judgment you want in each question’s instructions. A text-only request omits assets. state accepts a string, a JSON object, or a JSON array. Objects and arrays are serialized as text and read as one shared state, not as a batch of separate requests. A string is enough for a single message. Use an object when the decision depends on several pieces, such as a customer message and the policy it is checked against:
Answers come back in data.answers under the same names. Question types covers each answer’s shape.

Write a state the model can judge

  • Include the content, not a link. URLs in state are read as text; paste the document, page, or transcript you want evaluated.
  • Name the parts. Keys such as message, policy, and history tell the model what each piece is.
  • Keep one decision per request. Questions in a request share the same state and run independently. When a question depends on an earlier answer, ask it in a later request.
  • Watch the context window. Both models read up to 262,144 tokens per model read. See Models.
For spoken content, send the recording as an audio clip, or put the transcript in state. A transcript costs fewer tokens; the recording also carries tone and delivery.