Jev Images is a feature of 📝Jev that lets requests to TypeSafe's System One API include PNG and JPEG images, so the model can answer typed questions about what a picture shows and, through Box questions, locate where something appears in it. Before images, Jev decided only over text state, which left camera feeds, screenshots, documents, and product photos outside its fast, calibrated decisions. It is aimed at developers prototyping visual classification, moderation, and detection inside their software; TypeSafe labels it "Dreamer-ready," meaning the request format and limits may still change before production.
Key Capabilities
- Images anywhere in a question — Attach pictures to the state, a question's instructions, or its criteria, and every question type, including Choice, Score, and Noul, accepts them.
- Box localization — A Box question returns where something sits in an image as a box measured in fractions of the image's width and height; it requires Jev 1.14.
- Reference-based assets — Each image is uploaded once in a top-level assets map and cited by a reference string, so one picture can appear in several places.
- Priced like text — Images bill as input tokens at the same per-token rate as text, with no separate image fee, and are counted in the response's input token usage.
- Strict validation — Images are never resized; one oversized, undersized, or animated image fails the whole request, so callers must size images first.
Getting Started
- Install the latest TypeSafe Python or JavaScript SDK.
- Send the
TypeSafe-Danger-Zone-Images: trueheader on any request with an image. - Pass an Image object wherever text would go in the state, instructions, or criteria.
- Use Jev 1.14 for Box questions, or
jev-latestonce it points there. - Keep images PNG or JPEG, 280 to 1,568 px per side, and under 16 MiB each.