X/LittleWhiteBox
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update retrieval, rerank, and indexing changes

Browse Source
This commit is contained in:
bielie
2026-02-11 13:55:19 +08:00
parent 8d062d39b5
commit 297cc03770
7 changed files with 501 additions and 287 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
modules/story-summary/vector/retrieval/metrics.js
View File

@@ -8,7 +8,7 @@
// 架构变更（v3 → v4）：
// - evidence 区块反映 L0-only 融合 + L1 按楼层拉取的两阶段架构
// - 删除 mergedByType / selectedByType（不再有混合池）
// - 新增 l0Candidates / l0Selected / l1Pulled / l1Attached / l1CosineTime
// - 新增 floorCandidates / floorsSelected / l0Collected / l1Pulled / l1Attached / l1CosineTime
// - fusion 区块明确标注 L0-only（删除 anchorCount）
// - quality.chunkRealRatio → quality.l1AttachRate
// ═══════════════════════════════════════════════════════════════════════════
@@ -48,10 +48,10 @@ export function createMetrics() {
searchTime: 0,
},
// Fusion (W-RRF, L0-only) - 多路融合
// Fusion (W-RRF, floor-level) - 多路融合
fusion: {
denseCount: 0,
lexCount: 0,
denseFloors: 0,
lexFloors: 0,
totalUnique: 0,
afterCap: 0,
time: 0,
@@ -80,25 +80,27 @@ export function createMetrics() {
entityNames: [],
},
// Evidence (Two-Stage: L0 rerank → L1 pull) - 原文证据
// Evidence (Two-Stage: Floor rerank → L1 pull) - 原文证据
evidence: {
// Stage 1: L0
l0Candidates: 0, // W-RRF 融合后的 L0 候选数
l0Selected: 0, // rerank 后选中的 L0 数
// Stage 1: Floor
floorCandidates: 0, // W-RRF 融合后的 floor 候选数
floorsSelected: 0, // rerank 后选中的 floor 数
l0Collected: 0, // 选中 floor 中收集的 L0 atom 总数
rerankApplied: false,
rerankFailed: false,
beforeRerank: 0,
afterRerank: 0,
rerankTime: 0,
rerankScores: null,
rerankDocAvgLength: 0, // rerank document 平均字符数
// Stage 2: L1
l1Pulled: 0, // 从 DB 拉取的 L1 chunk 总数
l1Attached: 0, // 实际挂载的 L1 数（top-1 × 楼层 × 2侧）
l1Attached: 0, // 实际挂载的 L1 数（top-1 × floor × 2侧）
l1CosineTime: 0, // L1 cosine 打分耗时
// 装配
contextPairsAdded: 0, // 保留兼容（= l1Attached 中 USER 侧数量）
contextPairsAdded: 0, // USER 侧挂载数量
tokens: 0,
assemblyTime: 0,
},
@@ -149,7 +151,7 @@ export function createMetrics() {
quality: {
constraintCoverage: 100,
eventPrecisionProxy: 0,
l1AttachRate: 0, // 有 L1 挂载的 L0 占比
l1AttachRate: 0, // 有 L1 挂载的 floor 占比
potentialIssues: [],
},
};
@@ -223,10 +225,10 @@ export function formatMetricsLog(metrics) {
lines.push(`└─ search_time: ${m.lexical.searchTime}ms`);
lines.push('');
// Fusion (W-RRF, L0-only)
lines.push('[Fusion] W-RRF (L0-only) - 多路融合');
lines.push(`├─ dense_count: ${m.fusion.denseCount}`);
lines.push(`├─ lex_count: ${m.fusion.lexCount}`);
// Fusion (W-RRF, floor-level)
lines.push('[Fusion] W-RRF (floor-level) - 多路融合');
lines.push(`├─ dense_floors: ${m.fusion.denseFloors}`);
lines.push(`├─ lex_floors: ${m.fusion.lexFloors}`);
lines.push(`├─ total_unique: ${m.fusion.totalUnique}`);
lines.push(`├─ after_cap: ${m.fusion.afterCap}`);
lines.push(`└─ time: ${m.fusion.time}ms`);
@@ -277,28 +279,32 @@ export function formatMetricsLog(metrics) {
lines.push(`└─ entities_used: ${m.event.entitiesUsed} [${(m.event.entityNames || []).join(', ')}]`);
lines.push('');
// Evidence (Two-Stage)
lines.push('[Evidence] Two-Stage: L0 Locate → L1 Pull');
lines.push(`├─ Stage 1 (L0):`);
lines.push(`│ ├─ candidates (post-fusion): ${m.evidence.l0Candidates}`);
// Evidence (Two-Stage: Floor Rerank → L1 Pull)
lines.push('[Evidence] Two-Stage: Floor Rerank → L1 Pull');
lines.push(`├─ Stage 1 (Floor Rerank):`);
lines.push(`│ ├─ floor_candidates (post-fusion): ${m.evidence.floorCandidates}`);
if (m.evidence.rerankApplied) {
lines.push(`│ ├─ rerank_applied: true`);
if (m.evidence.rerankFailed) {
lines.push(`│ ├─ rerank_failed: ⚠ YES (using fusion order)`);
lines.push(`│ │ ⚠ rerank_failed: using fusion order`);
}
lines.push(`│ │ ├─ before: ${m.evidence.beforeRerank}`);
lines.push(`│ │ ├─ after: ${m.evidence.afterRerank}`);
lines.push(`│ │ ├─ before: ${m.evidence.beforeRerank} floors`);
lines.push(`│ │ ├─ after: ${m.evidence.afterRerank} floors`);
lines.push(`│ │ └─ time: ${m.evidence.rerankTime}ms`);
if (m.evidence.rerankScores) {
const rs = m.evidence.rerankScores;
lines.push(`│ ├─ rerank_scores: min=${rs.min}, max=${rs.max}, mean=${rs.mean}`);
}
if (m.evidence.rerankDocAvgLength > 0) {
lines.push(`│ ├─ rerank_doc_avg_length: ${m.evidence.rerankDocAvgLength} chars`);
}
} else {
lines.push(`│ ├─ rerank_applied: false`);
}
lines.push(`│ └─ selected: ${m.evidence.l0Selected}`);
lines.push(`│ ├─ floors_selected: ${m.evidence.floorsSelected}`);
lines.push(`│ └─ l0_atoms_collected: ${m.evidence.l0Collected}`);
lines.push(`├─ Stage 2 (L1):`);
lines.push(`│ ├─ pulled: ${m.evidence.l1Pulled}`);
lines.push(`│ ├─ attached: ${m.evidence.l1Attached}`);
@@ -345,9 +351,7 @@ export function formatMetricsLog(metrics) {
lines.push(`├─ constraint_filter: ${m.timing.constraintFilter}ms`);
lines.push(`├─ event_retrieval: ${m.timing.eventRetrieval}ms`);
lines.push(`├─ evidence_retrieval: ${m.timing.evidenceRetrieval}ms`);
if (m.timing.evidenceRerank > 0) {
lines.push(`├─ evidence_rerank: ${m.timing.evidenceRerank}ms`);
}
lines.push(`├─ floor_rerank: ${m.timing.evidenceRerank || 0}ms`);
lines.push(`├─ l1_cosine: ${m.evidence.l1CosineTime}ms`);
lines.push(`├─ evidence_assembly: ${m.timing.evidenceAssembly}ms`);
lines.push(`├─ formatting: ${m.timing.formatting}ms`);
@@ -406,20 +410,20 @@ export function detectIssues(metrics) {
// 词法检索问题
// ─────────────────────────────────────────────────────────────────
if ((m.lexical.terms || []).length > 0 && m.lexical.atomHits === 0 && m.lexical.chunkHits === 0 && m.lexical.eventHits === 0) {
if ((m.lexical.terms || []).length > 0 && m.lexical.chunkHits === 0 && m.lexical.eventHits === 0) {
issues.push('Lexical search returned zero hits - terms may not match any indexed content');
}
// ─────────────────────────────────────────────────────────────────
// 融合问题（L0-only）
// 融合问题（floor-level）
// ─────────────────────────────────────────────────────────────────
if (m.fusion.lexCount === 0 && m.fusion.denseCount > 0) {
issues.push('No lexical L0 candidates in fusion - hybrid retrieval not contributing');
if (m.fusion.lexFloors === 0 && m.fusion.denseFloors > 0) {
issues.push('No lexical floors in fusion - hybrid retrieval not contributing');
}
if (m.fusion.afterCap === 0) {
issues.push('Fusion produced zero L0 candidates - all retrieval paths may have failed');
issues.push('Fusion produced zero floor candidates - all retrieval paths may have failed');
}
// ─────────────────────────────────────────────────────────────────
@@ -463,29 +467,30 @@ export function detectIssues(metrics) {
}
// ─────────────────────────────────────────────────────────────────
// L0 Rerank 问题
// Floor Rerank 问题
// ─────────────────────────────────────────────────────────────────
if (m.evidence.rerankApplied) {
if (m.evidence.beforeRerank > 0 && m.evidence.afterRerank > 0) {
const filterRatio = 1 - (m.evidence.afterRerank / m.evidence.beforeRerank);
if (filterRatio > 0.7) {
issues.push(`High L0 rerank filter ratio (${(filterRatio * 100).toFixed(0)}%) - many irrelevant L0 in fusion output`);
}
}
if (m.evidence.rerankFailed) {
issues.push('Rerank API failed — using fusion rank order as fallback, relevance scores are zero');
}
if (m.evidence.rerankApplied && !m.evidence.rerankFailed) {
if (m.evidence.rerankScores) {
const rs = m.evidence.rerankScores;
if (rs.max < 0.5) {
issues.push(`Low L0 rerank scores (max=${rs.max}) - query may be poorly matched`);
if (rs.max < 0.3) {
issues.push(`Low floor rerank scores (max=${rs.max}) - query-document domain mismatch`);
}
if (rs.mean < 0.3) {
issues.push(`Very low average L0 rerank score (mean=${rs.mean}) - context may be weak`);
if (rs.mean < 0.2) {
issues.push(`Very low average floor rerank score (mean=${rs.mean}) - context may be weak`);
}
}
if (m.evidence.rerankTime > 2000) {
issues.push(`Slow L0 rerank (${m.evidence.rerankTime}ms) - may affect response time`);
if (m.evidence.rerankTime > 3000) {
issues.push(`Slow floor rerank (${m.evidence.rerankTime}ms) - may affect response time`);
}
if (m.evidence.rerankDocAvgLength > 3000) {
issues.push(`Large rerank documents (avg ${m.evidence.rerankDocAvgLength} chars) - may reduce rerank precision`);
}
}
@@ -493,21 +498,17 @@ export function detectIssues(metrics) {
// L1 挂载问题
// ─────────────────────────────────────────────────────────────────
if (m.evidence.rerankFailed) {
issues.push('Rerank API failed — using fusion rank order as fallback, relevance scores are zero');
}
if (m.evidence.l0Selected > 0 && m.evidence.l1Pulled === 0) {
if (m.evidence.floorsSelected > 0 && m.evidence.l1Pulled === 0) {
issues.push('Zero L1 chunks pulled - L1 vectors may not exist or DB read failed');
}
if (m.evidence.l0Selected > 0 && m.evidence.l1Attached === 0 && m.evidence.l1Pulled > 0) {
issues.push('L1 chunks pulled but none attached - cosine scores may be too low or floor mismatch');
if (m.evidence.floorsSelected > 0 && m.evidence.l1Attached === 0 && m.evidence.l1Pulled > 0) {
issues.push('L1 chunks pulled but none attached - cosine scores may be too low');
}
const l1AttachRate = m.quality.l1AttachRate || 0;
if (m.evidence.l0Selected > 5 && l1AttachRate < 20) {
issues.push(`Low L1 attach rate (${l1AttachRate}%) - many L0 lack concrete dialogue evidence`);
if (m.evidence.floorsSelected > 3 && l1AttachRate < 50) {
issues.push(`Low L1 attach rate (${l1AttachRate}%) - selected floors lack L1 chunks`);
}
// ─────────────────────────────────────────────────────────────────