Files
gbrain/src/core/search/dedup.ts
T
d547a64600 feat: search quality boost — compiled truth ranking + detail parameter (v0.8.1) (#64)
* feat: search quality boost — compiled truth ranking, detail parameter, cosine re-scoring

Compiled truth chunks now rank 2x higher in hybrid search via RRF
normalization + source boost. New --detail flag (low/medium/high)
controls timeline inclusion. Cosine re-scoring blends query-chunk
similarity before dedup for query-specific ranking.

Also: remove DISTINCT ON from keyword search (dedup handles per-page
capping), add chunk_id + chunk_index to SearchResult, add
getEmbeddingsByChunkIds to BrainEngine interface.

Inspired by Ramp Labs' "Latent Briefing" paper (April 2026).

* feat: RRF normalization, source-aware dedup, detail param in operations

RRF scores normalized to 0-1 before 2.0x compiled truth boost.
Source-aware dedup guarantees compiled truth chunk per page.
Detail parameter added to query operation, dedupResults added to
bare search operation. Debug logging via GBRAIN_SEARCH_DEBUG=1.

* chore: bump version and changelog (v0.8.1)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: CJK word count in query expansion

CJK text is not space-delimited. A query like "向量搜索优化" was counted
as 1 word and silently skipped expansion. Now counts characters for CJK
queries instead of space-separated tokens.

Co-Authored-By: YIING99 <yiing99@users.noreply.github.com>

* feat: retrieval evaluation harness — P@k, R@k, MRR, nDCG@k + gbrain eval

Full IR evaluation framework: precisionAtK, recallAtK, mrr, ndcgAtK
metrics with runEval() orchestrator. gbrain eval CLI with single-run
table and A/B comparison mode (--config-a / --config-b) for parameter
tuning. HybridSearchOpts now accepts rrfK and dedupOpts overrides.

Co-Authored-By: 4shut0sh <4shut0sh@users.noreply.github.com>

* test: search quality tests — RRF boost, dedup guarantee, cosine similarity, E2E benchmark

42 new tests across 3 files:
- test/search.test.ts: RRF normalization, compiled truth 2x boost, dedup key
  collision prevention, cosine similarity edge cases, CJK word count detection
- test/dedup.test.ts: source-aware compiled truth guarantee, layer interactions,
  custom maxPerPage, empty/single result edge cases
- test/e2e/search-quality.test.ts: full pipeline against PGLite with basis vector
  embeddings — chunk_id/chunk_index fields, detail parameter filtering,
  getEmbeddingsByChunkIds, keyword multi-chunk, vector ordering

Also: export rrfFusion + cosineSimilarity for unit testing, fix PGLite
getEmbeddingsByChunkIds to parse string vectors from pgvector.

* test: search quality benchmark with A/B comparison (baseline vs PR#64)

Benchmark measures P@1, MRR, nDCG@5, and source accuracy across 8 queries
against 5 seeded pages. Key finding: boost helps entity lookups but
over-corrects temporal queries. Validates the --detail parameter as the
right control mechanism. Output at docs/benchmarks/2026-04-13.md.

* feat: query intent classifier — auto-selects detail level, 100% source accuracy

Zero-latency heuristic classifier detects query intent from text patterns:
- "Who is Pedro?" → entity → detail=low (compiled truth only)
- "When did we last meet?" → temporal → detail=high (no boost, natural ranking)
- "Variant fund announcement" → event → detail=high
- General queries → detail=medium (default with boost)

The key insight: skip the 2.0x compiled truth boost for detail=high queries.
Temporal/event queries want natural ranking where timeline entries can win.

Benchmark results (source accuracy = does the top chunk match expected type):
- Baseline: 100% (already good, no boost needed)
- Boost only: 71.4% (boost over-corrects temporal queries)
- Boost + intent classifier: 100% (best of both worlds)

35 unit tests for the classifier. 590 total tests pass.

* feat: query intent classifier — auto-selects detail level, 100% source accuracy

Heuristic classifier detects query intent from text patterns (zero latency,
no LLM call). Maps temporal queries ("when did we last meet") to detail=high,
entity queries ("who is X") to detail=low, events to detail=high.

Benchmark results (29 pages, 20 queries, graded relevance):
- Baseline: P@1=0.947, MRR=0.974, source accuracy=89.5%
- Boost only: P@1=0.895, MRR=0.939, source accuracy=63.2% (over-correction)
- Boost + intent: P@1=0.947, MRR=0.974, source accuracy=89.5% (fully recovered)

The intent classifier eliminates the boost's over-correction on temporal queries
while preserving its benefits for entity lookups. 35 unit tests for the classifier.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test: search quality benchmark with A/B comparison (baseline vs PR#64)

Rich benchmark: 29 pages, 58 chunks, 20 queries with graded relevance.
Now measures CHUNK-LEVEL quality, not just page-level retrieval.

Key findings (C. Boost+Intent vs A. Baseline):
- Unique pages in top-10: 7.2 → 8.7 (+21% broader coverage)
- Compiled truth ratio: 51.6% → 66.8% (+15pp more signal)
- CT-first rate: 100% (compiled truth leads for entity queries)
- Timeline accessible: 100% (temporal queries still find dates)
- Source accuracy: 89.5% maintained (intent classifier prevents regression)

The boost alone (B) causes -26pp source accuracy regression.
Intent classifier (C) recovers it fully.

* docs: clean benchmark report — ELI10 search quality analysis for PR#64

Replaces two drafts with one clean report. Explains what changed, why it
matters, and what the numbers mean. All fictional data, no private info.

Key findings: 21% more page coverage per query, 29% more compiled truth
in results. Intent classifier prevents boost from burying timeline for
temporal queries. Full per-query breakdown with before/after comparison.

* chore: remove auto-generated benchmark file (clean version is 2026-04-14-search-quality.md)

* docs: update project documentation for search quality boost

CLAUDE.md: added search/intent.ts, search/eval.ts, commands/eval.ts to key
files. Added 5 new test files (search, dedup, intent, eval, e2e/search-quality).
Updated test count from 23+4 to 28+5. Added docs/benchmarks/ to key files.

README.md: updated search pipeline diagram with intent classifier, RRF
normalization, compiled truth boost, cosine re-scoring, and 5-layer dedup.
Added --detail flag explanation and benchmark instructions.

CHANGELOG.md: added search quality entries to v0.9.3 (intent classifier,
--detail flag, gbrain eval, CJK fix). Credited @4shut0sh and @YIING99.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: headline benchmark gains in changelog

* docs: add community attribution rule to CHANGELOG voice section

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: YIING99 <yiing99@users.noreply.github.com>
Co-authored-by: 4shut0sh <4shut0sh@users.noreply.github.com>
2026-04-13 21:03:40 -10:00

183 lines
5.1 KiB
TypeScript

/**
* 4-Layer Dedup Pipeline + Compiled Truth Guarantee
* Ported from production Ruby implementation (content_chunk.rb)
*
* 1. By source: top 3 chunks per page by score
* 2. By text similarity: remove chunks >0.85 Jaccard-similar to kept results
* 3. By type: no page type exceeds 60% of results
* 4. By page: max N chunks per page (default 2)
* 5. Compiled truth guarantee: ensure at least 1 compiled_truth chunk per page
*/
import type { SearchResult } from '../types.ts';
const COSINE_DEDUP_THRESHOLD = 0.85;
const MAX_TYPE_RATIO = 0.6;
const MAX_PER_PAGE = 2;
export function dedupResults(
results: SearchResult[],
opts?: {
cosineThreshold?: number;
maxTypeRatio?: number;
maxPerPage?: number;
},
): SearchResult[] {
const threshold = opts?.cosineThreshold ?? COSINE_DEDUP_THRESHOLD;
const maxRatio = opts?.maxTypeRatio ?? MAX_TYPE_RATIO;
const maxPerPage = opts?.maxPerPage ?? MAX_PER_PAGE;
// Preserve pre-dedup input for compiled truth guarantee
const preDedup = results;
let deduped = results;
// Layer 1: Top 3 chunks per page by score
deduped = dedupBySource(deduped);
// Layer 2: Text similarity dedup (Jaccard on word sets)
deduped = dedupByTextSimilarity(deduped, threshold);
// Layer 3: Type diversity (no page type exceeds 60%)
deduped = enforceTypeDiversity(deduped, maxRatio);
// Layer 4: Cap chunks per page
deduped = capPerPage(deduped, maxPerPage);
// Final pass: guarantee compiled_truth representation
deduped = guaranteeCompiledTruth(deduped, preDedup);
return deduped;
}
/**
* Layer 1: Keep top 3 chunks per page.
* Later layers (text similarity, cap per page) handle further reduction.
*/
function dedupBySource(results: SearchResult[]): SearchResult[] {
const byPage = new Map<string, SearchResult[]>();
for (const r of results) {
const existing = byPage.get(r.slug) || [];
existing.push(r);
byPage.set(r.slug, existing);
}
const kept: SearchResult[] = [];
for (const chunks of byPage.values()) {
chunks.sort((a, b) => b.score - a.score);
kept.push(...chunks.slice(0, 3));
}
return kept.sort((a, b) => b.score - a.score);
}
/**
* Layer 2: Remove chunks that are too similar to already-kept results.
* Uses Jaccard similarity on word sets as a proxy for cosine similarity.
*/
function dedupByTextSimilarity(results: SearchResult[], threshold: number): SearchResult[] {
const kept: SearchResult[] = [];
for (const r of results) {
const rWords = new Set(r.chunk_text.toLowerCase().split(/\s+/));
let tooSimilar = false;
for (const k of kept) {
const kWords = new Set(k.chunk_text.toLowerCase().split(/\s+/));
const intersection = new Set([...rWords].filter(w => kWords.has(w)));
const union = new Set([...rWords, ...kWords]);
const jaccard = intersection.size / union.size;
if (jaccard > threshold) {
tooSimilar = true;
break;
}
}
if (!tooSimilar) {
kept.push(r);
}
}
return kept;
}
/**
* Layer 3: No page type exceeds maxRatio of total results.
*/
function enforceTypeDiversity(results: SearchResult[], maxRatio: number): SearchResult[] {
const maxPerType = Math.max(1, Math.ceil(results.length * maxRatio));
const typeCounts = new Map<string, number>();
const kept: SearchResult[] = [];
for (const r of results) {
const count = typeCounts.get(r.type) || 0;
if (count < maxPerType) {
kept.push(r);
typeCounts.set(r.type, count + 1);
}
}
return kept;
}
/**
* Layer 4: Cap chunks per page.
*/
function capPerPage(results: SearchResult[], maxPerPage: number): SearchResult[] {
const pageCounts = new Map<string, number>();
const kept: SearchResult[] = [];
for (const r of results) {
const count = pageCounts.get(r.slug) || 0;
if (count < maxPerPage) {
kept.push(r);
pageCounts.set(r.slug, count + 1);
}
}
return kept;
}
/**
* Final pass: for each page in results that has no compiled_truth chunk,
* swap in the best compiled_truth chunk from the pre-dedup set (if one exists).
*/
function guaranteeCompiledTruth(results: SearchResult[], preDedup: SearchResult[]): SearchResult[] {
// Group results by page
const byPage = new Map<string, SearchResult[]>();
for (const r of results) {
const existing = byPage.get(r.slug) || [];
existing.push(r);
byPage.set(r.slug, existing);
}
const output = [...results];
for (const [slug, pageChunks] of byPage) {
const hasCompiledTruth = pageChunks.some(c => c.chunk_source === 'compiled_truth');
if (hasCompiledTruth) continue;
// Find the best compiled_truth chunk from pre-dedup input for this page
const candidate = preDedup
.filter(r => r.slug === slug && r.chunk_source === 'compiled_truth')
.sort((a, b) => b.score - a.score)[0];
if (!candidate) continue;
// Swap: replace the lowest-scored chunk from this page
const lowestIdx = output.reduce((minIdx, r, idx) => {
if (r.slug !== slug) return minIdx;
if (minIdx === -1) return idx;
return r.score < output[minIdx].score ? idx : minIdx;
}, -1);
if (lowestIdx !== -1) {
output[lowestIdx] = candidate;
}
}
return output;
}