Agent 工具
基于Jev相关性过滤和可审计token指标的Codex代码搜索插件。
Codex code-search plugin using Jev relevance filtering with auditable token metrics
const response = await fetchImpl('https://api.typesafe.ai/v1/systemone', {链接指向我们抓取当天的那个 commit，行号是准的。
下面是从这个项目源码里原样取出来的 question 组合。
relevance选项Does candidate contain useful evidence for user_question or retrieval_objective?
Yes — The passage provides useful evidence, a relevant dependency, background, a test, or counterevidence.No — The passage does not provide useful evidence for this information need; shared keywords alone are insufficient.import { experimental_evaluate as evaluate } from 'ai';
const { answers } = await evaluate({
model: 'typesafe-ai/jev',
state,
questions: {
relevance: { type: 'choice', instructions: 'Does candidate contain useful evidence for user_question or retrieval_objective?',
criteria: { Yes: 'The passage provides useful evidence, a relevant dependency, background, a test, or counterevidence.', No: 'The passage does not provide useful evidence for this information need; shared keywords alone are insufficient.' } },
},
});