更适合 Agent 的评估。

1import { defineEval, defineJudge } from "niceeval";
2import { pattern } from "niceeval/expect";
3
4const imageContext = defineJudge({ name: "image-context", rubric: "最后一轮回复是否基于先前图片？" });
5
6export default defineEval({
7  description: "评估 agent 在多轮对话中多模态的能力",
8
9  async test(t) {
    const first = await t.sendFile("evals/sample.png", "这张图片里有什么？");
    await first.succeeded().orStop();
    first.usedNoTools();
    const second = await t.send("图片里的背景是什么颜色？");
    t.check(second.message, pattern(/蓝|blue|白|方块|square/i));
    const third = await t.send("中间那个形状是什么颜色的？");
16
17    await t.group("后续追问能联系图片上下文", () => {
      t.check(third.message, pattern(/白|white/i));
19    });
20
    t.judge({ earlierImageFacts: ["蓝色背景", "白色方块"], question: third.input, answer: third.message }, imageContext).gate(0.7);≥ 0.7
22  },
23});