Skip to main content
POST
Jev

Introduction

Jev is TypeSafeโ€™s System One decision model. It does not generate text. Instead, it answers questions with a defined answer space (questions) about the content you provide (state), returning structured results your code can use directly:
  • Choice (choice): Select one of your options, such as which department should handle a support ticket.
  • Score (score): Rate content against your levels, such as customer anger: 0 calm / 1 dissatisfied / 2 angry.
  • Yes/no (noul): Return the probability of yes, such as whether the customer is requesting a refund.
  • Bounding box (box): Locate a described target in an image and return its four edges, such as scratches on a car.
Ask multiple questions in one request. The content is read once and each question is evaluated independently. Choice, Score, and Box answers include confidence; Noul answers return a probability. Use these decisions for ticket routing, moderation, sentiment scoring, intent detection, agent tool-call checks, and model routing.
The request structure and successful response fields follow TypeSafeโ€™s System One API. The platform may wrap error responses; see Error handling.
Jev does not support conversational endpoints such as OpenAIโ€™s /v1/chat/completions or /v1/responses. Call it through /v1/systemone below. It does not write replies, generate code, or explain its reasoning.

Authentication

string
required
Bearer token, e.g. Bearer sk-xxxxxxxxxx
string
required
application/json, or multipart/form-data for multipart image uploads. See Multipart image uploads.
string
Set to true. The official specification requires this header for requests containing images (assets) or Box questions (box).
string
Set to true to enable long context and raise the context limit from 64k to 262k. See Context length.
Create an API key on the platformโ€™s API keys page. Do not embed keys in frontend code, public repositories, or client installation packages.

Model information

  • jev-latest and jev-preview are aliases that move when TypeSafe releases new versions, so answers may change. The responseโ€™s model field identifies the version that answered.
  • If you have tuned confidence thresholds for a version, use its pinned model ID and switch only after evaluating the new version.

Endpoint

Request parameters

Jev has no sampling parameters such as temperature or max_tokens, and no output length setting is needed.

state

Provide the content to evaluate in any of three forms:
  • String: Text such as a customer message, support ticket, or complete contract.
  • Object: Structured data such as {"subject": "Duplicate charge", "message": "Please help"}. Use it to combine messages, order records, and business rules.
  • Array: Multiple pieces of text, such as a conversation history.

questions

A collection of named questions. The response returns answers under the same names in answers. Each value defines a question:
  • Questions are independent. Each is evaluated against state without seeing the other answers.
  • Specific option and level descriptions make decisions more consistent. Avoid overlapping options.

assets

A collection of images keyed by your placeholders (a new UUID per image is recommended). A placeholder anywhere in a string in state, question instructions, or criteria is replaced with the corresponding image. The following example uses a 280 ร— 280 PNG with a red rectangle on white. It is also used in the image and Box examples below. Do not replace it with a placeholder image below the minimum dimensions.
The official specification requires TypeSafe-Danger-Zone-Images: true for requests with assets. Images and Box questions are early features; request formats, answers, and limits may change between versions.

Image limits

Images must satisfy all of the following official limits. Images outside the dimension limits are not resized automatically; the whole request fails: Every asset must be referenced in state, instructions, or criteria. Undersized images, invalid formats, and unreferenced assets typically return 400. Exceeding file, dimension, or request size limits may return 413.

Multipart image uploads

For large or multiple images, upload the original files with multipart/form-data instead of Base64:
  • The first part must be named request and contain the JSON request body (Content-Type: application/json).
  • Each image follows in its own part. Part names must be unique, nonempty, and referenced by an assetโ€™s part field.

Context length

Every request must fit both limits:
  • Exceeding the limit returns an upstream 400 error with max_tokens_exceeded.
  • For long contracts or conversation histories, send TypeSafe-Danger-Zone-Long-Context: true. Long context has the same per-token price, but is slower and slightly less accurate. Enable it only when needed.

Usage examples

Successful response

Box answer

Response fields

Using confidence: Probabilities and confidence are calibrated and can support automated decision thresholds. For example, automatically process noul โ‰ฅ 0.8, send 0.2 ~ 0.8 for human review, and allow < 0.2. Tune thresholds on a set of manually labeled examples.

Billing

Formula: Cost per request = usage.input_tokens / 1,000,000 ร— input price
  • These are official base prices. Actual charges also apply the multiplier for the API keyโ€™s group. Refer to the prices shown in the model marketplace.
  • Input tokens include state, all question instructions, options, and level descriptions, plus image tokens when images are included.
  • output_tokens is informational and currently not billed.
  • Long context keeps the same per-token price and only permits longer inputs.
Example: If the multiple-question request above returns input_tokens: 443, the cost is 443 / 1,000,000 ร— 0.042 โ‰ˆ USD 0.0000186, or about USD 0.19 for 10,000 such requests.

Error handling

Failed requests return a JSON error object. The platform may wrap upstream errors: message may contain a specific reason or only openai_error. Do not rely on this field alone to identify the cause. Example error format:
Official API validation failures may return 422, while a gateway may return 400 with an error wrapper. Check the HTTP status first, then read the error fields. Do not assume every gateway preserves the official detail structure or specific error messages.
Common issues:

Official references