ChoiceOrganize a product catalog
Give Choice your own product categories, including an “other” option. Test mixed scenes and unusual products before routing images automatically.
JEV 1.14 / VISION LAB
Text playground Turn an image into a decision.
Explore Jev image recognition. Bring a photo, ask a focused question, and get a category, probability, score, or bounding box your application can use.
Experimental vision
TypeSafe marks images and Box as experimental. Use this lab to test ideas and evaluate results on your own examples.
01 / Your image
Sample image · ChelseaPrepared image · 451 × 300 · 49 KiB
Your image is prepared in this browser. Large photos are resized; small or narrow images are padded. It is sent to OpenRouter and TypeSafe only when you run a decision.
02 / Your question
03 / The decision
Your result will appear here after a run. The lab shows actual model output, including uncertainty.
Inspect the request
The preview abbreviates image data. Copy or download includes the full image and a fresh request ID, ready for POST /api/jev/run with your account API key. Reuse that ID when retrying the same request.
{
"model": "typesafe/jev-1.14",
"state": {
"photo": {
"type": "image_url",
"image_url": {
"url": "data:image/jpeg;base64,<image data>"
}
},
"context": ""
},
"questions": {
"vision": {
"type": "choice",
"instructions": "Which category best describes the main subject in `photo`?",
"criteria": {
"product": null,
"food": null,
"animal": null,
"landscape": null,
"other": null
}
}
}
}From pixels to useful software decisions
Start with a small question that has a clear output. Compare results across real examples, then decide how your application should handle uncertainty.
NoulCheck a visual requirement
Use Noul to ask whether a photo is clear enough or a required object is visible. Keep ambiguous cases in a review queue.
BoxFind the part that matters
Use Box to locate one described object, then pass the region to a crop or annotation workflow. Check presence before accepting the coordinates.
Jev image recognition, explained
What changed in Jev 1.14?
Released on October 14, 2026, Jev 1.14 adds the Box primitive for locating an object. Images also work with Choice, Noul, and Score. This lab pins the OpenRouter model typesafe/jev-1.14.
Can Jev generate images or write an image description?
Jev returns structured decisions. Choose categories for classification, a statement for a yes/no probability, ordered levels for a score, or a target for a box. Use a generative model when you need a new image or free-form prose.
How are images handled in this playground?
The browser converts one PNG or JPEG to an upright JPEG, with a maximum side of 1,568 pixels and padding where needed. Our run history stores structured results and a request fingerprint; it does not store the image. Running sends the image through OpenRouter to TypeSafe for processing.
How does vision billing work?
Image and text input usage share your existing balance. Classification, checks, and scores use one question; locating uses two because it also checks presence. The displayed hold is a ceiling, not an estimated charge. Actual usage appears with the result.
Can I use the same request in my application?
Yes. Download the full request and send it to /api/jev/run with an account API key. Keep credentials on your server. The API guide covers authentication and retries; the upstream references describe Jev’s image and Box formats.
Verified against official documentation · October 14, 2026