Jev 1.14
Model id: typesafe/jev-1.14. Jev is a System One decision model from TypeSafe: instead of generating text it returns typed answers (yes/no, choice, score) with calibrated probabilities for routing, cl
Model Overview
Jev is a "System One" decision model from TypeSafe: instead of generating text it returns typed answers — yes/no, choice, score, box — with calibrated probabilities. It is built for routing, classification, triage and other in-app decision points where a fast, predictable verdict matters more than prose, and it answers in well under a second.
Version 1.14 improves across the board, most of all in languages other than English, and adds two experimental capabilities: reading images, and the new box question — where something is in an image. Context window is 262,144 tokens.
This is the pinned id. typesafe/jev currently resolves to this same version, but it moves with each release — pin typesafe/jev-1.14 when you need the behaviour to stay put, and see Jev 1.13 for the previous one.
API Schema
Resolved model version that answered.
POST /v1/decisions HTTP/1.1
Host: api.aimlapi.com
Content-Type: application/json
Accept: */*
Content-Length: 312
{
"model": "typesafe/jev-1.14",
"state": "Help! My payments have been failing for 3 days.",
"questions": {
"ANY_ADDITIONAL_PROPERTY": {
"type": "noul",
"instructions": "text",
"criteria": {
"true": "text",
"false": "text"
}
}
},
"images": [],
"assets": {
"a3f1c2e8-door": {
"type": "image",
"format": "jpg",
"data_base64": "/9j/4AAQSkZJRg..."
}
}
}{
"model": "text",
"answers": {
"ANY_ADDITIONAL_PROPERTY": {
"type": "noul",
"noul": 1
}
},
"usage": {
"input_tokens": 1,
"output_tokens": 1
}
}Code Example
Every key you send under questions comes back under answers with the matching type. The top-level model is the exact build that answered (jev-1.14.0); meta.model is the id you called.
Images and the box question
New in 1.14, and marked experimental by TypeSafe: the model reads images, and a box question returns where something is in one.
Images go in assets, not in images — each entry is a base64-encoded PNG or JPEG, with no data: prefix and no line breaks. Refer to an asset by its key from the question's instructions.
A box answer is four edges, each a fraction of the image from the top-left corner, plus a confidence. There is no "not found" answer, so pair it with a noul that asks whether the target is in the picture at all — exactly as below.
To turn the fractions into pixels, multiply left/right by the image width and top/bottom by its height.
Last updated
Was this helpful?