@jjh32456/dsh-longlive 1.6.7 → 1.8.0

package/test/embedding-recall.test.mjs

@@ -1,347 +0,0 @@
-/**
- * A3/A4 自测：向量索引（tokenizer / 增量缓存 / 融合打分）与 autoRecall 的
- * recall-log 记账。只依赖临时目录与测试 embedder，不加载真模型、不碰 profile。
- */
-
-import test from 'node:test';
-import assert from 'node:assert/strict';
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
-
-import {
-  EMBEDDING_VERSION,
-  buildVectorIndex,
-  cardChunkTexts,
-  createEmbeddingRuntime,
-  loadTokenizer,
-  normalizeEmbeddingConfig,
-} from '../src/recall/embedding.mjs';
-import { chooseRecall } from '../src/recall/recall-core.mjs';
-import { apply as applyAutoRecall } from '../src/recall/index.js';
-import { normalizeFeatureConfig } from '../src/features.js';
-import { PreStepBus, makeUserMessageFactory, textMessage } from './fixtures/pre-step-container.mjs';
-
-function tmpRoot(t) {
-  const root = fs.mkdtempSync(path.join(os.tmpdir(), 'longlive-embed-'));
-  t.after(() => fs.rmSync(root, { recursive: true, force: true }));
-  return root;
-}
-
-function writeFile(root, rel, text) {
-  const file = path.join(root, rel);
-  fs.mkdirSync(path.dirname(file), { recursive: true });
-  fs.writeFileSync(file, text, 'utf8');
-  return file;
-}
-
-function makeEmbedder({ fingerprint = 'test-fp', vector = [1, 0], dim = 2 } = {}) {
-  const docTexts = [];
-  const embedDocument = async (text) => {
-    docTexts.push(text);
-    return Float32Array.from(vector);
-  };
-  return {
-    dim,
-    modelFingerprint: fingerprint,
-    docTexts,
-    embedQuery: async () => Float32Array.from(vector),
-    embedDocument,
-    embedMany: async (texts) => Promise.all(texts.map((text) => embedDocument(text))),
-  };
-}
-
-function makeCtx(bus) {
-  const warns = [];
-  return {
-    warns,
-    ctx: {
-      on: bus.on.bind(bus),
-      logger: {
-        warn: (...args) => warns.push(args.join(' ')),
-        info: () => {},
-      },
-    },
-  };
-}
-
-test('embedding 默认关闭，权重/阈值/缓存目录可配置', () => {
-  const config = normalizeEmbeddingConfig({}, { memoryRoot: 'X:/mem' });
-  assert.equal(config.enabled, false);
-  assert.equal(config.keywordWeight, 0.6);
-  assert.equal(config.vectorWeight, 0.4);
-  assert.equal(config.semanticAdmit, 0.55);
-  assert.equal(config.minFusedScore, 0.2);
-  assert.equal(config.maxSeqLen, 256);
-  assert.equal(config.cacheDir, path.join('X:/mem', '.longlive', 'embeddings'));
-
-  const custom = normalizeEmbeddingConfig(
-    { enabled: true, keywordWeight: 0.7, vectorWeight: 0.3, semanticAdmit: 0.6 },
-    { memoryRoot: 'X:/mem' },
-  );
-  assert.equal(custom.enabled, true);
-  assert.equal(custom.keywordWeight, 0.7);
-  assert.equal(custom.vectorWeight, 0.3);
-  assert.equal(custom.semanticAdmit, 0.6);
-});
-
-test('autoRecall 注入预算硬上限 2000 token（A4）', () => {
-  const normalized = normalizeFeatureConfig({ autoRecall: { enabled: true, budgetTokens: 50000 } });
-  assert.equal(normalized.autoRecall.budgetTokens, 2000);
-  const small = normalizeFeatureConfig({ autoRecall: { enabled: true, budgetTokens: 400 } });
-  assert.equal(small.autoRecall.budgetTokens, 400);
-});
-
-test('tokenizer.json 自解析：中文单字 + 英文 WordPiece + maxLen 截断', (t) => {
-  const root = tmpRoot(t);
-  const vocab = {
-    '[PAD]': 0,
-    '[UNK]': 1,
-    '[CLS]': 2,
-    '[SEP]': 3,
-    '[MASK]': 4,
-    你: 5,
-    好: 6,
-    play: 7,
-    '##ing': 8,
-    世: 9,
-    界: 10,
-  };
-  const file = writeFile(root, 'tokenizer.json', JSON.stringify({
-    normalizer: { type: 'BertNormalizer', clean_text: true, handle_chinese_chars: true, lowercase: false },
-    pre_tokenizer: { type: 'BertPreTokenizer' },
-    model: { type: 'WordPiece', unk_token: '[UNK]', continuing_subword_prefix: '##', max_input_chars_per_word: 100, vocab },
-    added_tokens: [
-      { id: 0, content: '[PAD]' },
-      { id: 1, content: '[UNK]' },
-      { id: 2, content: '[CLS]' },
-      { id: 3, content: '[SEP]' },
-      { id: 4, content: '[MASK]' },
-    ],
-  }));
-  const tokenizer = loadTokenizer(file);
-  assert.equal(tokenizer.vocabSize, 11);
-  assert.deepEqual(tokenizer.encode('你好 playing'), [2, 5, 6, 7, 8, 3]);
-  assert.deepEqual(tokenizer.encode('你好 playing', { maxLen: 4 }), [2, 5, 6, 3]);
-  assert.deepEqual(tokenizer.tokensOf('playing'), ['play', '##ing']);
-  assert.equal(tokenizer.decode([2, 5, 6, 3]), '[CLS] 你 好 [SEP]');
-});
-
-test('向量缓存按卡增量：未改复用、改一张只重算一张、指纹变全量', async (t) => {
-  const root = tmpRoot(t);
-  const cachePath = path.join(root, 'index.json');
-  const config = normalizeEmbeddingConfig({}, { memoryRoot: root });
-  const cards = [
-    { path: 'topics/a.md', title: 'A', scopeIn: ['甲'], aliases: [], body: '甲卡的正文。' },
-    { path: 'topics/b.md', title: 'B', scopeIn: ['乙'], aliases: [], body: '乙卡的正文。' },
-  ];
-  const embedder = makeEmbedder();
-
-  const first = await buildVectorIndex({ cards, config, embedder, cachePath });
-  assert.equal(first.stats.embeddedCards, 2);
-  assert.equal(first.stats.reusedCards, 0);
-  assert.equal(first.stats.written, true);
-  assert.equal(first.cardVectors.get('topics/a.md')[0].length, 2);
-  assert.equal(fs.existsSync(cachePath), true);
-  const cache = JSON.parse(fs.readFileSync(cachePath, 'utf8'));
-  assert.equal(cache.version, EMBEDDING_VERSION);
-  assert.equal(cache.modelFingerprint, 'test-fp');
-  assert.deepEqual(Object.keys(cache.cards).sort(), ['topics/a.md', 'topics/b.md']);
-
-  const second = await buildVectorIndex({ cards, config, embedder, cachePath });
-  assert.equal(second.stats.embeddedCards, 0);
-  assert.equal(second.stats.reusedCards, 2);
-
-  cards[1] = { ...cards[1], body: '乙卡的正文改了。' };
-  const third = await buildVectorIndex({ cards, config, embedder, cachePath });
-  assert.equal(third.stats.embeddedCards, 1);
-  assert.equal(third.stats.reusedCards, 1);
-  assert.equal(embedder.docTexts.filter((text) => text.includes('乙卡')).length, 2);
-
-  const other = makeEmbedder({ fingerprint: 'test-fp-2' });
-  const fourth = await buildVectorIndex({ cards, config, embedder: other, cachePath });
-  assert.equal(fourth.stats.embeddedCards, 2);
-  assert.equal(fourth.stats.reusedCards, 0);
-});
-
-test('融合打分：关键词+向量、纯语义准入、静默不误召、权重可调', () => {
-  const cards = [
-    { path: 'topics/a.md', title: '沙箱卡', scopeIn: ['沙箱', 'pwsh'], aliases: [], body: '沙箱里 pwsh 起不来。' },
-    { path: 'topics/b.md', title: '语义卡', scopeIn: ['量子纠缠'], aliases: [], body: '只靠语义进来的卡。' },
-  ];
-  const queryText = '沙箱 pwsh 起不来怎么办';
-  const queryVector = Float32Array.from([1, 0]);
-  const vectors = new Map([
-    ['topics/a.md', [Float32Array.from([1, 0])]],
-    ['topics/b.md', [Float32Array.from([0.9, 0.1])]],
-  ]);
-
-  const fused = chooseRecall({
-    queryText, cards, queryVector, vectors, budget: 1200,
-    options: { maxItems: 3, trace: true },
-  });
-  assert.equal(fused.items.length, 2);
-  assert.equal(fused.items[0].path, 'topics/a.md');
-  assert.equal(fused.items[0].reason, 'keyword+vector');
-  assert.equal(fused.items[1].path, 'topics/b.md');
-  assert.equal(fused.items[1].reason, 'vector');
-  assert.ok(fused.items[1].score > 0.39 && fused.items[1].score < 0.4);
-  assert.equal(fused.hits.semanticAdmitted, 2);
-
-  // 静默样本：没有关键词命中，余弦也低于 semanticAdmit → 一张都不注入。
-  const silent = chooseRecall({
-    queryText: '今天天气怎么样',
-    cards,
-    queryVector,
-    vectors: new Map([
-      ['topics/a.md', [Float32Array.from([0.3, 0.9539])]],
-      ['topics/b.md', [Float32Array.from([0.3, 0.9539])]],
-    ]),
-    budget: 1200,
-  });
-  assert.deepEqual(silent.items, []);
-  assert.equal(silent.hits.semanticAdmitted, 0);
-
-  // 权重可调：把向量权重压到 0，纯语义卡过不了 minFusedScore。
-  const keywordOnly = chooseRecall({
-    queryText, cards, queryVector, vectors, budget: 1200,
-    options: { maxItems: 3, fusion: { keywordWeight: 1, vectorWeight: 0 } },
-  });
-  assert.deepEqual(keywordOnly.items.map((item) => item.path), ['topics/a.md']);
-});
-
-test('autoRecall：融合注入 + recall-log 行格式（含 blocked）', async (t) => {
-  const root = tmpRoot(t);
-  writeFile(root, 'topics/a.md', '---\ntitle: 沙箱卡\nscope_in: ["沙箱", "pwsh"]\n---\n# 沙箱卡\n- 沙箱里的 pwsh 起不来。\n');
-  writeFile(root, 'topics/b.md', '---\ntitle: 语义卡甲\nscope_in: ["量子纠缠"]\n---\n# 语义卡甲\n- 只靠语义进来的甲。\n');
-  writeFile(root, 'topics/c.md', '---\ntitle: 语义卡乙\nscope_in: ["黎曼猜想"]\n---\n# 语义卡乙\n- 只靠语义进来的乙。\n');
-
-  const bus = new PreStepBus();
-  const { ctx, warns } = makeCtx(bus);
-  applyAutoRecall(ctx, {
-    memoryRoot: root,
-    dirs: ['topics'],
-    budget: 1200,
-    maxItems: 2,
-    everyNTurns: 6,
-    log: true,
-    createUserMessage: makeUserMessageFactory(),
-    embedding: { enabled: true, embedder: makeEmbedder() },
-  });
-
-  const { decision } = await bus.compose({
-    messages: [textMessage('沙箱 pwsh 起不来怎么办')],
-    agent: { session: { id: 'sess-embed', header: { origin: 'main' } } },
-    turn: 1,
-    step: 1,
-  });
-  assert.equal(decision.messages.length, 2);
-  const injected = decision.messages[1];
-  assert.equal(injected.source?.kind, 'session-reference');
-  const text = injected.content.map((part) => part.text).join('\n');
-  assert.match(text, /【自动想起】/);
-  assert.match(text, /沙箱卡/);
-  assert.match(text, /语义卡甲/);
-  assert.doesNotMatch(text, /语义卡乙/);
-
-  const logPath = path.join(root, '.longlive', 'recall-log.jsonl');
-  const lines = fs.readFileSync(logPath, 'utf8').trim().split(/\r?\n/);
-  const row = JSON.parse(lines.at(-1));
-  assert.deepEqual(Object.keys(row).sort(), ['at', 'blocked', 'injected', 'sessionId', 'turn']);
-  assert.equal(row.sessionId, 'sess-embed');
-  assert.equal(row.turn, 1);
-  assert.equal(Number.isNaN(Date.parse(row.at)), false);
-  assert.deepEqual(row.injected.map((item) => item.id), ['topics/a.md', 'topics/b.md']);
-  for (const item of row.injected) {
-    assert.equal(typeof item.title, 'string');
-    assert.equal(typeof item.score, 'number');
-    assert.match(item.reason, /^(keyword\+vector|vector|keyword)$/);
-  }
-  assert.equal(row.injected[0].reason, 'keyword+vector');
-  assert.equal(row.injected[1].reason, 'vector');
-  assert.equal(row.blocked.length, 1);
-  assert.equal(row.blocked[0].id, 'topics/c.md');
-  assert.equal(row.blocked[0].reason, 'max-items');
-  assert.deepEqual(warns.filter((line) => /EMBEDDING_FALLBACK/.test(line)), []);
-});
-
-test('autoRecall：模型缺失时静默回退关键词，recall-log 照样写 reason=keyword', async (t) => {
-  const root = tmpRoot(t);
-  writeFile(root, 'topics/a.md', '---\ntitle: 沙箱卡\nscope_in: ["沙箱", "pwsh"]\n---\n# 沙箱卡\n- 沙箱里的 pwsh 起不来。\n');
-
-  const bus = new PreStepBus();
-  const { ctx, warns } = makeCtx(bus);
-  applyAutoRecall(ctx, {
-    memoryRoot: root,
-    dirs: ['topics'],
-    budget: 1200,
-    maxItems: 2,
-    log: true,
-    createUserMessage: makeUserMessageFactory(),
-    embedding: { enabled: true, modelPath: path.join(root, 'no-such-model') },
-  });
-
-  const { decision } = await bus.compose({
-    messages: [textMessage('沙箱 pwsh 起不来怎么办')],
-    agent: { session: { id: 'sess-fallback', header: { origin: 'main' } } },
-    turn: 1,
-    step: 1,
-  });
-  assert.equal(decision.messages.length, 2);
-  assert.match(decision.messages[1].content[0].text, /沙箱卡/);
-  assert.equal(warns.some((line) => /EMBEDDING_FALLBACK/.test(line)), true);
-
-  const row = JSON.parse(fs.readFileSync(path.join(root, '.longlive', 'recall-log.jsonl'), 'utf8').trim());
-  assert.equal(row.injected.length, 1);
-  assert.equal(row.injected[0].id, 'topics/a.md');
-  assert.equal(row.injected[0].reason, 'keyword');
-});
-
-test('autoRecall：embedding 关闭时不建 embeddings 目录，也不读模型', async (t) => {
-  const root = tmpRoot(t);
-  writeFile(root, 'topics/a.md', '---\ntitle: 沙箱卡\nscope_in: ["沙箱"]\n---\n# 沙箱卡\n- 沙箱相关。\n');
-  const bus = new PreStepBus();
-  const { ctx } = makeCtx(bus);
-  applyAutoRecall(ctx, {
-    memoryRoot: root,
-    dirs: ['topics'],
-    budget: 1200,
-    createUserMessage: makeUserMessageFactory(),
-  });
-  const { decision } = await bus.compose({
-    messages: [textMessage('沙箱里怎么运行')],
-    agent: { session: { id: 'sess-off', header: { origin: 'main' } } },
-    turn: 1,
-    step: 1,
-  });
-  assert.equal(decision.messages.length, 2);
-  assert.equal(fs.existsSync(path.join(root, '.longlive', 'embeddings')), false);
-  const row = JSON.parse(fs.readFileSync(path.join(root, '.longlive', 'recall-log.jsonl'), 'utf8').trim());
-  assert.equal(row.injected[0].reason, 'keyword');
-});
-
-test('createEmbeddingRuntime：ensure 失败返回结构化错误而不是抛进宿主', async (t) => {
-  const root = tmpRoot(t);
-  const runtime = createEmbeddingRuntime(
-    { enabled: true, modelPath: path.join(root, 'missing') },
-    { memoryRoot: root },
-  );
-  const result = await runtime.ensure([]);
-  assert.equal(result.ok, false);
-  assert.equal(result.code, 'E_EMBEDDING_MODEL_MISSING');
-  assert.equal(runtime._state().ok, false);
-});
-
-test('cardChunkTexts：B 文本包含标题/关键词/正文三段', () => {
-  const texts = cardChunkTexts({
-    path: 'topics/a.md',
-    title: '标题甲',
-    scopeIn: ['词一'],
-    aliases: ['别称一'],
-    body: '正文第一段。',
-  }, { maxChunkChars: 400, maxChunksPerCard: 6 });
-  assert.equal(texts.length, 1);
-  assert.match(texts[0], /【标题】标题甲/);
-  assert.match(texts[0], /【关键词】词一 别称一/);
-  assert.match(texts[0], /【正文】/);
-});