Jev 1.14 can look at photos, point at things and read long books

What it does

Caparas checked 24 screen targets against their browser-measured positions. He reports median box overlap of 0.50 on the first pass and 0.88 after enlarging crops, with 23 centres landing on target. A missing object still gets a box. Questions joining two planted book facts failed even at 24,000 tokens; separate questions helped. These opt-in features carry TypeSafe's warning against production use. His small seven-language comparison found no clear improvement over 1.13.

How you can use it

With a TypeSafe API access key, follow the photo script and switch images on for that request. Put two questions in it: where is the object, and how likely is it to be there? Use the second answer to decide whether to keep the box. For a tiny button, try the enlarged-crop example.

The book scripts let you compare a question needing two clues with a question about each clue on its own. Keep the combining step in your code. Before changing model versions, run cases from your own task whose answers you know; planted sentences in a novel are much tidier than a real contract.