开源复现

visual-jev

@andrueandersoncs1Python更新于 2026-09-20可当场跑

基于共享视觉编码和 Qwen3-VL,实现图像原生的类型化决策。

英文原文

Image-native typed decisions with shared visual encoding and Qwen3-VL

andrueandersoncs/visual-jev

它在哪儿调用了 Jev

@app.post("/v1/systemone")

src/visual_jev/api.py:576

链接指向我们抓取当天的那个 commit,行号是准的。

它问 Jev 的问题

下面是从这个项目源码里原样取出来的 question 组合。

取自 tests/test_core.py:29

  1. composition选项

    Is the composition balanced?

    • balancedThe visual weight is balanced
    • left-heavyWeight sits mainly on the left
    • unclearThe image cannot be assessed
  2. legibility是/否

    Is the title legible?

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    composition: { type: 'choice', instructions: 'Is the composition balanced?',
      criteria: { balanced: 'The visual weight is balanced', left-heavy: 'Weight sits mainly on the left', unclear: 'The image cannot be assessed' } },
    legibility: { type: 'boolean', instructions: 'Is the title legible?' },
  },
});

同类的其他项目