How to Use Jev AI

What You Need Before Using Jev

Start with an accessible TypeSafe account, a development environment and one specific judgment you want to evaluate. A useful first task has an answer you can check manually. Avoid beginning with an entire autonomous workflow.

• A server or local terminal that can keep an API key private.
• A small set of fictional or approved test inputs.
• A written definition of a correct answer and an uncertain answer.
• A fallback path when the service or model cannot answer reliably.

Step 1 — Access Jev

Follow the official quick start to the playground or dashboard. Verify account access there. The launch announcement described early access; do not assume the launch-day access rules still describe every account.

Reference: Official documentation ↗

Step 2 — Understand Jev Inputs

The state field holds the material being evaluated. It may be text or a structured JSON value. Put only relevant context into this field and define the judgment separately in the question instructions.

For a support workflow, begin with the message itself. Add account context only if the decision truly depends on it. Supplying irrelevant fields makes the experiment harder to interpret.

Reference: Official documentation ↗

Step 3 — Send a Request

Open the linked Jev API Guide for complete cURL, JavaScript and Python examples. The starter request asks whether a message requests an invoice copy. Run it locally with your own key, inspect the HTTP status, and save the response for review.

Before changing the question, try several paraphrases and one counterexample. For instance, distinguish a request for an invoice from a statement that an invoice has already arrived.

Open the three-language request examples →

Step 4 — Understand the Structured Output

Choice selects from defined alternatives. Score evaluates ordered criteria. Noul represents the probability of a yes answer. These outputs let code consume a judgment without parsing a free-form essay.

Keep the original numeric value in logs instead of immediately discarding it with a threshold. Label a small evaluation set yourself, measure mistakes, then choose a threshold based on the consequences of the action.

Reference: Official documentation ↗

Step 5 — Integrate Jev Into an Application

Wrap the API call in a backend function with a timeout and explicit error handling. Validate the output before using it. Separate evaluation from execution so that a model result cannot bypass normal authorization or business rules.

• Develop against a saved response fixture before connecting the live service.
• Test normal, ambiguous, adversarial and missing-data cases.
• Run in observation mode first: record proposed actions without executing them.
• Monitor outcome quality and costs after changing the model or instructions.
• Treat retries separately from side effects to avoid executing an action twice.

Common Use Cases

Classification and routing are natural patterns to evaluate with bounded choices. A scoring rubric may support a review queue. These are integration ideas, not a claim of validated accuracy for your workload. Risk scoring requires a domain-specific evaluation; a numeric answer alone is not a validated risk model.

Reference: Official documentation ↗

Common Problems

A 401 response usually means the key or authorization header needs attention. A 422 response indicates request validation failed. For 429 or 529 responses, use bounded backoff rather than a tight retry loop.

For poor judgments, inspect the actual state and question before adding complexity. Keep arithmetic in code and test hostile input. The provider’s known-limitations page is linked in the sources.

Reference: Official documentation ↗

How to Use Jev AI: Check Account Access

The official quick start separates trying a question in the Playground from calling the HTTP API. The Playground requires login; the dashboard supplies the API key used by a server-side request. A successful browser session therefore does not show that your local process has received its credential. For a first integration, use the documented POST endpoint with JSON and a Bearer authorization header. Account-specific quotas, permissions and credit allowances are not confirmed in this guide. Check those in your account rather than treating the public quick-start example as a promise of access or a free allowance. Primary source · Reviewed Sep 26, 2026.

Reference: Official documentation ↗

How to Use Jev AI Question Types

Choose the question type from the answer your code needs. Choice uses a criteria object whose keys identify alternatives and whose values explain them. Score uses an ordered criteria array describing levels. Noul asks a yes/no question and can omit criteria. Each question also needs its type and complete instructions; its identifier alone does not communicate the question to the model. For an original support example, define a Choice between invoice and other, a Score describing increasingly urgent wording, and a Noul asking whether a copy is requested. These are design examples, not measured results. Primary source · Reviewed Sep 26, 2026.

All three questions may share one questions object and one state. Their responses appear under your chosen identifiers, but the judgments are independent. Do not assume the Score secretly incorporates the Choice result. If a later question needs an earlier answer, make that dependency explicit in a subsequent request. No new API example here has been run against a paid account. Primary source · Reviewed Sep 26, 2026.

Reference: Official documentation ↗

Prepare Jev AI State for Debugging

State is the material to evaluate, while questions specify the judgments. The documented input can be a string, an object or an array containing text. For a simple invoice request, a message string is sufficient. If a judgment also depends on a policy, an object can keep the message and policy in separately named fields. This gives you a concrete debugging comparison: did the request include both pieces of information, and did the instructions refer to the intended field? Jev's documented state input is text-based; attaching an image, audio file or video is not a supported substitute for that material. Primary source · Reviewed Sep 26, 2026.

Reference: Official documentation ↗

Interpret Jev AI Confidence Carefully

Confidence and the selected answer describe different things. For Choice and Score, the documentation derives confidence from the distribution over the available alternatives. Inspect that distribution alongside the chosen option or score when reviewing a questionable result. A concentrated distribution is not an independent guarantee that the judgment matches your business rule. Keep uncertainty handling separate from the action your application might take. For example, a review queue can retain the original answer fields so a person can inspect them. This guide does not claim a universally safe confidence threshold; no threshold has been validated here for your workload. Primary source · Reviewed Sep 26, 2026.

Reference: Official documentation ↗

Troubleshoot Jev AI HTTP Errors

Use the HTTP response to choose the next debugging step. The reference associates 401 with a missing or invalid credential, so inspect how the Authorization header is assembled without exposing the secret in logs. For 422, inspect the validation error body and repair the indicated request field before retrying. The documented 429 and 529 conditions mean rate limiting and temporary overload respectively; the reference recommends exponential backoff. A malformed body and an overloaded service therefore need different handling. This guide does not establish your account's numerical rate limit or promise that retrying a request will succeed. Primary source · Reviewed Sep 26, 2026.

Reference: Official documentation ↗

Frequently asked questions

Keep exploring