Jev on Workers AI

No TypeSafe key appears anywhere in the examples. You authenticate to the platform instead, with env.AI.run('typesafe/jev', ...) inside a Worker or POST to /ai/run on the REST side using your Cloudflare account ID and an AI API token as a Bearer credential. The model is flagged third-party, and the card lists zero data retention.

The card states a 32,000-token context window and nothing more about quality inside it. There are no accuracy figures, no calibration numbers and no latency benchmarks, because this is a platform model listing rather than an evaluation. The worked examples are small, around 426 input tokens for a support message with three questions, so nothing on the page tells you how the model behaves near the limit.

Both. The structured refund example passes a nested JSON object as state, with a ticket holding subject and message, an order holding an array of charge records, and a refund_policy string, then asks whether the policy supports the refund given those charges. One call evaluates all three question types together, so a question can reference fields of the state rather than a flattened prompt.

Answers keyed by question id, plus the resolved model id such as jev-1.13.0 and input and output token counts. A noul gives a single probability, a choice gives the picked option, per-option probabilities and a separate confidence, and a score gives a fractional value, its rubric legend and probabilities, which is why the sample returns 1.04 on a three-level scale. Choosing thresholds stays your code's job, and confidence is not the same quantity as a noul probability.

The model page lists $0.042 per million input tokens and $0.00 for output and cached input, and marks the model third-party with zero data retention. It publishes no accuracy or calibration numbers, because this is a platform model listing rather than an evaluation.

Inside a Worker, env.AI.run('typesafe/jev', { state, questions }); from anywhere else, POST to the account-scoped /ai/run endpoint with a Bearer token and a body of { model: 'typesafe/jev', input: { state, questions } }.

A duplicate-charge refund review that passes a nested state with an order and a refund policy, a support request routed to the right department, and account risk scoring with escalation. The refund example returns noul 0.99 for refund_requested and 0.98 for policy_supports_refund.