fix: extraction, multi-hop traversal, and aggregate result shape (BR-ADV-FEATURES-BUN)

Three advanced-API correctness fixes, all reproducible on Node (not Bun-specific):

- Entity/concept extraction returned []. SmartExtractor combined agreeing signals
  with a weighted sum compared against an absolute 0.60 gate, so a confident
  low-weight signal lost selection to a mediocre high-weight one that then failed
  the gate, dropping the whole result. Select and gate on a normalized weighted
  average instead. The public `confidence` option now controls the threshold (was
  a dead hardcoded 0.60), and the embedding-signal timeout is raised 100ms -> 2000ms
  so the neural signal is not silently dropped on slower runtimes.

- Multi-hop find({ connected }) returned only the 1-hop neighbour. executeGraphSearch
  ignored depth/via; it now delegates to the depth-aware neighbors() BFS.

- find({ aggregate }) hid groupKey/metrics/count under .metadata, so callers
  expecting AggregateResult saw empty rows. Expose those fields at the top level.

Adds real-embedding regression tests in tests/integration/advanced-apis-regression.test.ts.
This commit is contained in:
David Snelling 2026-05-26 11:32:46 -07:00
parent 07754d135a
commit 0a9d1d9a17
8 changed files with 297 additions and 74 deletions

View file

@ -211,12 +211,18 @@ export class SmartExtractor {
formatContext?: FormatContext
allTerms?: string[]
metadata?: any
}
},
minConfidence?: number
): Promise<ExtractionResult | null> {
this.stats.calls++
// Per-call confidence threshold (falls back to the instance default). This lets callers
// such as brain.extractEntities({ confidence }) actually loosen or tighten the gate; it is
// part of the cache key below so results computed at one threshold are not reused at another.
const threshold = minConfidence ?? this.options.minConfidence
// Check cache first
const cacheKey = this.getCacheKey(candidate, context)
const cacheKey = this.getCacheKey(candidate, context, threshold)
const cached = this.getFromCache(cacheKey)
if (cached !== undefined) {
this.stats.cacheHits++
@ -282,8 +288,8 @@ export class SmartExtractor {
// Combine using ensemble or best signal
const result = this.options.enableEnsemble
? this.combineEnsemble(signalResults, formatHints, context?.formatContext)
: this.selectBestSignal(signalResults, formatHints, context?.formatContext)
? this.combineEnsemble(signalResults, formatHints, context?.formatContext, threshold)
: this.selectBestSignal(signalResults, formatHints, context?.formatContext, threshold)
// Cache result (including nulls to avoid recomputation)
this.addToCache(cacheKey, result)
@ -509,7 +515,8 @@ export class SmartExtractor {
private combineEnsemble(
signalResults: SignalResult[],
formatHints: string[],
formatContext?: FormatContext
formatContext?: FormatContext,
minConfidence: number = this.options.minConfidence
): ExtractionResult | null {
// Filter out null results
const validResults = signalResults.filter(r => r.type !== null)
@ -518,58 +525,63 @@ export class SmartExtractor {
return null
}
// Count votes by type with weighted confidence
const typeScores = new Map<NounType, { score: number; signals: SignalResult[] }>()
// Group the signals by the type they voted for
const typeScores = new Map<NounType, SignalResult[]>()
for (const result of validResults) {
if (!result.type) continue
const weighted = result.confidence * result.weight
const existing = typeScores.get(result.type)
if (existing) {
existing.score += weighted
existing.signals.push(result)
existing.push(result)
} else {
typeScores.set(result.type, { score: weighted, signals: [result] })
typeScores.set(result.type, [result])
}
}
// Find best type
// Score each candidate type by a NORMALIZED, weighted-average confidence
// — Σ(confidence·weight) / Σ(weight of its signals) — plus a small boost when
// multiple signals concur, then select the type with the highest such confidence.
//
// Why an average and not a sum (this was the bug): a weighted *sum* lives on the
// signal-weight scale, so two signals agreeing on a fresh brain summed to ≈0.37 —
// below the 0.60 gate — meaning agreement was effectively *penalized*. Worse, selecting
// the best type by that sum let a high-weight signal with mediocre confidence
// (embedding @0.51 · 0.35 = 0.179) outrank a low-weight signal with high confidence
// (pattern @0.82 · 0.20 = 0.164); the wrong type won selection, then failed the threshold
// on its own 0.51 confidence, and the whole extraction returned null. Averaging keeps the
// score on the same 01 scale as the threshold, so selection and the gate agree and the
// most-confident type wins (pattern @0.82 here).
let bestType: NounType | null = null
let bestScore = 0
let finalConfidence = 0
let bestSignals: SignalResult[] = []
for (const [type, data] of typeScores.entries()) {
// Apply agreement boost (multiple signals agree)
let finalScore = data.score
if (data.signals.length > 1) {
const agreementBoost = 0.05 * (data.signals.length - 1)
finalScore += agreementBoost
this.stats.agreementBoosts++
for (const [type, signals] of typeScores.entries()) {
const weightSum = signals.reduce((sum, s) => sum + s.weight, 0)
const weightedConfidence = signals.reduce((sum, s) => sum + s.confidence * s.weight, 0)
let confidence = weightSum > 0 ? weightedConfidence / weightSum : 0
// Reward agreement between independent signals
if (signals.length > 1) {
confidence = Math.min(confidence + 0.05 * (signals.length - 1), 1.0)
}
if (finalScore > bestScore) {
bestScore = finalScore
if (confidence > finalConfidence) {
finalConfidence = confidence
bestType = type
bestSignals = data.signals
bestSignals = signals
}
}
// Determine final confidence score
// FIX: When only one signal matches, use its original confidence instead of weighted score
// The weighted score is too low when only one signal matches (e.g., 0.8 * 0.2 = 0.16 < 0.60 threshold)
let finalConfidence = bestScore
if (bestSignals.length === 1) {
// Single signal: use its original confidence
finalConfidence = bestSignals[0].confidence
}
// Check minimum confidence threshold
if (!bestType || finalConfidence < this.options.minConfidence) {
if (!bestType || finalConfidence < minConfidence) {
return null
}
if (bestSignals.length > 1) {
this.stats.agreementBoosts++
}
// Track signal contributions
const usedSignals = bestSignals.length
this.stats.averageSignalsUsed =
@ -604,13 +616,17 @@ export class SmartExtractor {
private selectBestSignal(
signalResults: SignalResult[],
formatHints: string[],
formatContext?: FormatContext
formatContext?: FormatContext,
minConfidence: number = this.options.minConfidence
): ExtractionResult | null {
// Filter valid results and sort by weighted confidence
// Select by confidence — the same metric the threshold checks below. Sorting by
// weighted score (confidence·weight) instead would let a high-weight, mediocre-confidence
// signal outrank a low-weight, high-confidence one and then fail the gate, dropping the
// result entirely (the same defect fixed in combineEnsemble). Weight matters for ensemble
// voting, not for picking the single best signal.
const validResults = signalResults
.filter(r => r.type !== null)
.map(r => ({ ...r, weightedScore: r.confidence * r.weight }))
.sort((a, b) => b.weightedScore - a.weightedScore)
.sort((a, b) => b.confidence - a.confidence)
if (validResults.length === 0) {
return null
@ -618,9 +634,7 @@ export class SmartExtractor {
const best = validResults[0]
// FIX: Use original confidence, not weighted score for threshold check
// Weighted score is for ranking signals, not for absolute threshold
if (best.confidence < this.options.minConfidence) {
if (best.confidence < minConfidence) {
return null
}
@ -661,11 +675,12 @@ export class SmartExtractor {
/**
* Get cache key from candidate and context
*/
private getCacheKey(candidate: string, context?: any): string {
private getCacheKey(candidate: string, context?: any, minConfidence?: number): string {
const normalized = candidate.toLowerCase().trim()
const defSnippet = context?.definition?.substring(0, 50) || ''
const format = context?.formatContext?.format || ''
return `${normalized}:${defSnippet}:${format}`
const threshold = minConfidence ?? this.options.minConfidence
return `${normalized}:${defSnippet}:${format}:${threshold}`
}
/**