Skip to main content
The questions object maps names you choose to questions. The answers object uses the same names, so your application can retrieve a result without relying on its position. All three types can return a refusal instead of an answer. Examples on this page show illustrative outputs.

​
Predicates

Use type: "noul" for a condition such as whether a product supports a feature. Provide instructions and, optionally, criteria with both true and false descriptions.
The answer at answers.has_cli has this shape:
noul estimates the probability that the statement is true. Do not treat missing evidence as proof that it is false. If your application needs an explicit insufficient-evidence outcome, use a choice question with an unknown option.

​
Choices

Use type: "choice" and a criteria object with 2–255 options. Each key is a possible output value; each description explains when that option applies.
The answer at answers.cli_availability includes the selected key and the option probabilities:

​
Scores

Use type: "score" and an ordered criteria list with 2–10 levels. Give each level a concrete description.
The answer at answers.cli_fit includes the score, a legend, and probabilities:
Levels are indexed from zero. The score is their probability-weighted mean, so it can be fractional: 0 × 0.1 + 1 × 0.1 + 2 × 0.8 = 1.7. It ranges from zero to the number of levels minus one. The legend and probabilities use string indices.

​
Refusals

For any question type, an unanswered question can have this answer:
A downstream provider can return a refusal, including because of safety filtering. A refusal does not contain a value or probabilities. Handle it as a separate outcome; do not turn it into a negative predicate, the first choice, or a zero score. Other questions in the same response can still have answers.

​
Multiple Questions

Add more named questions to the same questions object to mix types. Each request supports 1–100 questions over one shared evidence set. Group questions about a single entity or related topic, and send separate requests for unrelated entities. See Best Practices for input limits and evidence handling, or return to the Quickstart.