Jev Eyes

It gives the text-only Jev eyes: an image becomes the text state Jev can decide over, produced locally via OCR plus a spatial layout, and returned as a plain dict so you always see exactly what Jev saw.

Alongside OCR text in reading order, the state includes the same words at their true position on a character grid. The README says this lets Jev answer where things are — which region is the main content, whether an error sits next to the form — without a pixel reaching it.

Not for perception: after installing jev-eyes, the command jev-eyes see screenshot.png --show runs with no key at all. ask(image, questions) does perception plus the Jev call when TYPESAFE_API_KEY is set.

One line changes: pass see("screenshot.png") as the state to TypeSafeClient().system_one(...) with Noul and Choice questions. Answers come back as result.answers["decorative"].noul, result.answers["kind"].choice, and a needs_vision noul you can gate on.

It is a gate for escalating to a VLM only when the extracted text is insufficient to describe what the image shows, so most images are handled locally without a vision model call.

The README shows pip install jev-eyes but notes that until the PyPI release lands you should install with pip install git+https://github.com/LeddoEngano/jev-eyes instead.

The repository is MIT licensed and Python-based, and the digest's README excerpt cuts off inside the description of what the state contains, so fields beyond image size and source are not covered here.