DeepSeek Harness 本地聊天记忆：按来源保存聊天原文，关键词 + 向量检索，可选大模型重排与检索摘要。源码公开，仅限非商用
浏览文件
main
分支
eval.test.js
15 行 · 955 B历史下载| 1 | import test from 'node:test'; |
| 2 | import assert from 'node:assert/strict'; |
| 3 | import { evaluate } from './eval/run.js'; |
| 4 | |
| 5 | // Floors for the lexical (no embedding) pipeline on the synthetic chat corpus. Raise them when |
| 6 | // retrieval improves; a drop means a change made memory worse. Run `node test/eval/run.js` for the report. |
| 7 | test('retrieval quality does not regress', async () => { |
| 8 | const { byKind, newerLast, leaks } = await evaluate(); |
| 9 | assert.equal(leaks, 0, 'another person\'s private chat must never surface'); |
| 10 | assert.ok(byKind.exact.hit3 >= 1, 'exact: every record in the top 3'); |
| 11 | assert.ok(byKind.context.answer >= 1, 'context: the neighbouring turn carries the answer into the evidence'); |
| 12 | assert.ok(newerLast, 'changed facts: the newer statement is shown, after the older one'); |
| 13 | assert.ok(byKind.all.hit3 >= 0.75, `top-3 hit rate ${byKind.all.hit3}`); |
| 14 | assert.ok(byKind.all.answer >= 0.75, `answer-in-evidence rate ${byKind.all.answer}`); |
| 15 | }); |
