Decide (Sarvam Prism)
Answers a set of typed questions about one input and returns a calibrated probability for every allowed option. The model does not generate text, so there is nothing to parse. The order of choice options does not affect the answer.
Requires api-subscription-key, or Authorization: Bearer <key>.
Question types: noul (yes/no), choice (one of your options) and score (a level on your ordered scale).
Repeatability: probabilities can vary slightly between identical requests; the top answer is stable.
Not supported: images and streaming. Generation fields such as temperature, max_tokens, stream and images are refused: image URLs return 400, and the others return 422. Unknown keys inside a question also return 422.
Limits: 32,768 tokens per question (state + question + options), 2–255 options per choice and 2–10 levels per score. There is no cap on the number of questions. The state is not truncated, and if any one question breaks a limit the whole request fails with 422.
Timeout: each request must finish within 60 seconds. A longer request fails with 503 model_overloaded; split very large batches across several requests before retrying.
Errors: every error carries a message, a code and a request_id. Validation errors also carry param, the path to the field at fault, such as questions.urgent.noul.instruction.
Authentication
Bearer authentication of the form Bearer <token>, where token is your auth token.
Request
The decision model to use. Currently sarvam-prism (Sarvam Prism).
The input to decide on: text, or an object or array that is rendered as compact JSON. Required; an empty string is allowed. It is never truncated: if any one question (state + question + options) exceeds 32,768 tokens the request fails with 422.