diff --git a/modules/story-summary/generate/prompt.js b/modules/story-summary/generate/prompt.js index 2d078b2..c1ed29c 100644 --- a/modules/story-summary/generate/prompt.js +++ b/modules/story-summary/generate/prompt.js @@ -128,12 +128,12 @@ function normalize(s) { /** * 收集 L0 的实体集合（用于背景证据实体过滤） - * 支持新结构 who/edges，也兼容旧结构 subject/object。 + * 使用 edges.s/edges.t。 * @param {object} l0 * @returns {Set} */ function collectL0Entities(l0) { - const atom = l0?.atom || l0?._atom || {}; + const atom = l0?.atom || {}; const set = new Set(); const add = (v) => { @@ -141,16 +141,11 @@ function collectL0Entities(l0) { if (n) set.add(n); }; - for (const w of (atom.who || [])) add(w); for (const e of (atom.edges || [])) { add(e?.s); add(e?.t); } - // 兼容旧数据 - add(atom.subject); - add(atom.object); - return set; } @@ -159,8 +154,7 @@ function collectL0Entities(l0) { * 规则： * 1) 无焦点实体：保留 * 2) similarity >= 0.70：保留（旁通） - * 3) who/edges 命中焦点实体：保留 - * 4) 兼容旧数据：semantic 文本包含焦点实体：保留 + * 3) edges 命中焦点实体：保留 * 否则过滤。 * @param {object} l0 * @param {Set} focusSet @@ -175,11 +169,6 @@ function shouldKeepEvidenceL0(l0, focusSet) { if (entities.has(f)) return true; } - // 旧数据兜底：从 semantic 文本里做包含匹配 - const textNorm = normalize(l0?.atom?.semantic || l0?.text || ''); - for (const f of focusSet) { - if (f && textNorm.includes(f)) return true; - } return false; } diff --git a/modules/story-summary/vector/llm/atom-extraction.js b/modules/story-summary/vector/llm/atom-extraction.js index 6481778..4f6e3be 100644 --- a/modules/story-summary/vector/llm/atom-extraction.js +++ b/modules/story-summary/vector/llm/atom-extraction.js @@ -3,8 +3,6 @@ // // 设计依据： // - BGE-M3 (BAAI, 2024): 自然语言段落检索精度最高 → semantic = 纯自然语言 -// - Interpersonal Circumplex (Kiesler, 1983): 权力轴+情感轴 → dynamics 枚举 -// - Labov Narrative Structure (1972): 叙事功能轴 → dynamics 枚举补充 // - TransE (Bordes, 2013): s/t/r 三元组方向性 → edges 格式 // // 每楼层 1-2 个场景锚点（非碎片原子），60-100 字场景摘要 @@ -32,26 +30,6 @@ export function isBatchCancelled() { return batchCancelled; } -// ============================================================================ -// dynamics 封闭枚举（8 个标签，两轴四象限 + 叙事轴） -// ============================================================================ - -const VALID_DYNAMICS = new Set([ - // 权力轴 (Interpersonal Circumplex: Dominance-Submission) - '支配', // 控制、命令、审视、威慑、主导 - '让渡', // 顺从、服从、屈服、被动、配合 - - // 情感轴 (Interpersonal Circumplex: Hostility-Friendliness) - '亲密', // 温柔、关怀、依赖、信任、连接 - '敌意', // 对抗、拒绝、攻击、嘲讽、排斥 - - // 叙事轴 (Labov Narrative Structure) - '揭示', // 真相、发现、告白、暴露、秘密 - '决意', // 选择、承诺、放弃、宣言、转折 - '张力', // 悬念、对峙、暗涌、不安、等待 - '丧失', // 分离、死亡、破碎、遗憾、崩塌 -]); - // ============================================================================ // L0 提取 Prompt // ============================================================================ @@ -68,9 +46,7 @@ const SYSTEM_PROMPT = `你是场景摘要器。从一轮对话中提取1-2个场 {"anchors":[ { "scene": "60-100字完整场景描述", - "who": ["角色名1","角色名2"], "edges": [{"s":"施事方","t":"受事方","r":"互动行为"}], - "dynamics": ["标签"], "where": "地点" } ]} @@ -81,20 +57,12 @@ const SYSTEM_PROMPT = `你是场景摘要器。从一轮对话中提取1-2个场 - 读者只看 scene 就能复原这一幕 - 60-100字，信息密集但流畅 -## who -- 参与互动的角色正式名称，不用代词或别称 -- 只从正文内容中识别角色名，不要把标签名（如 user、assistant）当作角色 - ## edges（关系三元组） - s=施事方 t=受事方 r=互动行为（10-15字） +- s/t 必须是参与互动的角色正式名称，不用代词或别称 +- 只从正文内容中识别角色名，不要把标签名（如 user、assistant）当作角色 - 每个锚点 1-3 条 -## dynamics（封闭枚举，选0-2个） -权力轴：支配（控制/命令/审视） | 让渡（顺从/服从/屈服） -情感轴：亲密（温柔/信任/连接） | 敌意（对抗/拒绝/攻击） -叙事轴：揭示（真相/秘密） | 决意（选择/承诺） | 张力（对峙/不安） | 丧失（分离/破碎） -纯日常无明显模式时 dynamics 为 [] - ## where - 场景地点，无明确地点时空字符串 @@ -107,7 +75,7 @@ const SYSTEM_PROMPT = `你是场景摘要器。从一轮对话中提取1-2个场 ## 示例 输入：艾拉在火山口举起圣剑刺穿古龙心脏，龙血溅满她的铠甲，她跪倒在地痛哭 输出： -{"anchors":[{"scene":"火山口上艾拉举起圣剑刺穿古龙的心脏，龙血溅满铠甲，古龙轰然倒地，艾拉跪倒在滚烫的岩石上痛哭，完成了她不得不做的弑杀","who":["艾拉","古龙"],"edges":[{"s":"艾拉","t":"古龙","r":"以圣剑刺穿心脏"}],"dynamics":["决意","丧失"],"where":"火山口"}]}`; +{"anchors":[{"scene":"火山口上艾拉举起圣剑刺穿古龙的心脏，龙血溅满铠甲，古龙轰然倒地，艾拉跪倒在滚烫的岩石上痛哭，完成了她不得不做的弑杀","edges":[{"s":"艾拉","t":"古龙","r":"以圣剑刺穿心脏"}],"where":"火山口"}]}`; const JSON_PREFILL = '{"anchors":['; @@ -121,19 +89,6 @@ const sleep = (ms) => new Promise(r => setTimeout(r, ms)); // 清洗与构建 // ============================================================================ -/** - * 清洗 dynamics 标签，只保留合法枚举值 - * @param {string[]} raw - * @returns {string[]} - */ -function sanitizeDynamics(raw) { - if (!Array.isArray(raw)) return []; - return raw - .map(d => String(d || '').trim()) - .filter(d => VALID_DYNAMICS.has(d)) - .slice(0, 2); -} - /** * 清洗 edges 三元组 * @param {object[]} raw @@ -148,26 +103,8 @@ function sanitizeEdges(raw) { t: String(e.t || '').trim(), r: String(e.r || '').trim().slice(0, 30), })) - .filter(e => e.s && e.t && e.r) - .slice(0, 3); -} - -/** - * 清洗 who 列表 - * @param {string[]} raw - * @returns {string[]} - */ -function sanitizeWho(raw) { - if (!Array.isArray(raw)) return []; - const seen = new Set(); - return raw - .map(w => String(w || '').trim()) - .filter(w => { - if (!w || w.length < 1 || seen.has(w)) return false; - seen.add(w); - return true; - }) - .slice(0, 6); + .filter(e => e.s && e.t && e.r) + .slice(0, 3); } /** @@ -186,10 +123,7 @@ function anchorToAtom(anchor, aiFloor, idx) { // scene 过短（< 15 字）可能是噪音 if (scene.length < 15) return null; - - const who = sanitizeWho(anchor.who); const edges = sanitizeEdges(anchor.edges); - const dynamics = sanitizeDynamics(anchor.dynamics); const where = String(anchor.where || '').trim(); return { @@ -201,9 +135,7 @@ function anchorToAtom(anchor, aiFloor, idx) { semantic: scene, // ═══ 图结构层（扩散的 key） ═══ - who, edges, - dynamics, where, }; } @@ -268,7 +200,7 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op } // 兼容：优先 anchors，回退 atoms - const rawAnchors = parsed?.anchors || parsed?.atoms; + const rawAnchors = parsed?.anchors; if (!rawAnchors || !Array.isArray(rawAnchors)) { if (attempt < RETRY_COUNT) { await sleep(RETRY_DELAY); @@ -395,3 +327,4 @@ export async function batchExtractAtoms(chat, onProgress) { return allAtoms; } + diff --git a/modules/story-summary/vector/retrieval/diffusion.js b/modules/story-summary/vector/retrieval/diffusion.js index 6fb55a3..9aec33a 100644 --- a/modules/story-summary/vector/retrieval/diffusion.js +++ b/modules/story-summary/vector/retrieval/diffusion.js @@ -32,6 +32,7 @@ import { xbLog } from '../../../../core/debug-core.js'; import { getContext } from '../../../../../../../extensions.js'; +import { tokenizeForIndex } from '../utils/tokenizer.js'; const MODULE_ID = 'diffusion'; @@ -46,12 +47,11 @@ const CONFIG = { MAX_ITER: 50, // hard iteration cap (typically converges in 15-25) // Edge weight channel coefficients - // Rationale: Rimmon-Kenan (2002) hierarchy: characters > events > setting > themes + // No standalone WHO channel: rely on interaction/action/location only. GAMMA: { - who: 0.50, // entity co-occurrence — Jaccard - what: 0.25, // directed pair overlap — Szymkiewicz-Simpson - where: 0.15, // location exact match — binary - how: 0.10, // dynamics tag co-occurrence — Jaccard + what: 0.55, // interaction pair overlap — Szymkiewicz-Simpson + where: 0.15, // location exact match — binary + how: 0.30, // action-term co-occurrence — Jaccard }, // Post-verification (Cosine Gate) @@ -94,17 +94,13 @@ function cosineSimilarity(a, b) { // ═══════════════════════════════════════════════════════════════════════════ /** - * WHO channel: entity set = who ∪ edges.s ∪ edges.t + * Endpoint entity set from edges.s/edges.t (used for candidate pair generation). * @param {object} atom * @param {Set} excludeEntities - entities to exclude (e.g. name1) * @returns {Set} */ function extractEntities(atom, excludeEntities = new Set()) { const set = new Set(); - for (const w of (atom.who || [])) { - const n = normalize(w); - if (n && !excludeEntities.has(n)) set.add(n); - } for (const e of (atom.edges || [])) { const s = normalize(e?.s); const t = normalize(e?.t); @@ -115,18 +111,19 @@ function extractEntities(atom, excludeEntities = new Set()) { } /** - * WHAT channel: directed interaction pairs "A→B" (strict direction — option A) + * WHAT channel: interaction pairs "A↔B" (direction-insensitive). * @param {object} atom * @param {Set} excludeEntities * @returns {Set} */ -function extractDirectedPairs(atom, excludeEntities = new Set()) { +function extractInteractionPairs(atom, excludeEntities = new Set()) { const set = new Set(); for (const e of (atom.edges || [])) { const s = normalize(e?.s); const t = normalize(e?.t); if (s && t && !excludeEntities.has(s) && !excludeEntities.has(t)) { - set.add(`${s}\u2192${t}`); + const pair = [s, t].sort().join('\u2194'); + set.add(pair); } } return set; @@ -142,15 +139,20 @@ function extractLocation(atom) { } /** - * HOW channel: dynamics tags set + * HOW channel: action terms from edges.r * @param {object} atom + * @param {Set} excludeEntities * @returns {Set} */ -function extractDynamics(atom) { +function extractActionTerms(atom, excludeEntities = new Set()) { const set = new Set(); - for (const d of (atom.dynamics || [])) { - const n = normalize(d); - if (n) set.add(n); + for (const e of (atom.edges || [])) { + const rel = String(e?.r || '').trim(); + if (!rel) continue; + for (const token of tokenizeForIndex(rel)) { + const t = normalize(token); + if (t && !excludeEntities.has(t)) set.add(t); + } } return set; } @@ -198,8 +200,8 @@ function overlapCoefficient(a, b) { // Graph construction // // Candidate pairs discovered via inverted indices on entities and locations. -// Dynamics-only pairs excluded from candidate generation (γ_HOW = 0.10 is -// too weak to justify O(N²) blowup from 8-tag combinatorics). +// HOW-only pairs are still excluded from candidate generation to avoid O(N²); +// all channel weights are evaluated for the entity/location candidate set. // All four channels evaluated for every candidate pair. // ═══════════════════════════════════════════════════════════════════════════ @@ -207,14 +209,14 @@ function overlapCoefficient(a, b) { * Pre-extract features for all atoms * @param {object[]} allAtoms * @param {Set} excludeEntities - * @returns {object[]} feature objects with entities/directedPairs/location/dynamics + * @returns {object[]} feature objects with entities/interactionPairs/location/actionTerms */ function extractAllFeatures(allAtoms, excludeEntities = new Set()) { return allAtoms.map(atom => ({ entities: extractEntities(atom, excludeEntities), - directedPairs: extractDirectedPairs(atom, excludeEntities), + interactionPairs: extractInteractionPairs(atom, excludeEntities), location: extractLocation(atom), - dynamics: extractDynamics(atom), + actionTerms: extractActionTerms(atom, excludeEntities), })); } @@ -279,10 +281,10 @@ function buildGraph(allAtoms, excludeEntities = new Set()) { collectPairsFromIndex(entityIndex, pairSet, N); collectPairsFromIndex(locationIndex, pairSet, N); - // Compute four-channel edge weights for all candidates + // Compute three-channel edge weights for all candidates const neighbors = Array.from({ length: N }, () => []); let edgeCount = 0; - const channelStats = { who: 0, what: 0, where: 0, how: 0 }; + const channelStats = { what: 0, where: 0, how: 0 }; for (const packed of pairSet) { const i = Math.floor(packed / N); @@ -291,13 +293,11 @@ function buildGraph(allAtoms, excludeEntities = new Set()) { const fi = features[i]; const fj = features[j]; - const wWho = jaccard(fi.entities, fj.entities); - const wWhat = overlapCoefficient(fi.directedPairs, fj.directedPairs); + const wWhat = overlapCoefficient(fi.interactionPairs, fj.interactionPairs); const wWhere = (fi.location && fi.location === fj.location) ? 1.0 : 0.0; - const wHow = jaccard(fi.dynamics, fj.dynamics); + const wHow = jaccard(fi.actionTerms, fj.actionTerms); const weight = - CONFIG.GAMMA.who * wWho + CONFIG.GAMMA.what * wWhat + CONFIG.GAMMA.where * wWhere + CONFIG.GAMMA.how * wHow; @@ -307,7 +307,6 @@ function buildGraph(allAtoms, excludeEntities = new Set()) { neighbors[j].push({ target: i, weight }); edgeCount++; - if (wWho > 0) channelStats.who++; if (wWhat > 0) channelStats.what++; if (wWhere > 0) channelStats.where++; if (wHow > 0) channelStats.how++; @@ -318,8 +317,7 @@ function buildGraph(allAtoms, excludeEntities = new Set()) { xbLog.info(MODULE_ID, `Graph: ${N} nodes, ${edgeCount} edges ` + - `(who=${channelStats.who} what=${channelStats.what} ` + - `where=${channelStats.where} how=${channelStats.how}) ` + + `(what=${channelStats.what} where=${channelStats.where} how=${channelStats.how}) ` + `(${buildTime}ms)` ); @@ -593,7 +591,7 @@ function postVerify(pi, atomIds, atomById, seedAtomIds, vectorMap, queryVector) * @param {object[]} seeds - l0Selected from recall Stage 6 * Each: { atomId, rerankScore, similarity, atom, ... } * @param {object[]} allAtoms - getStateAtoms() result - * Each: { atomId, floor, semantic, who, edges, dynamics, where } + * Each: { atomId, floor, semantic, edges, where } * @param {object[]} stateVectors - getAllStateVectors() result * Each: { atomId, floor, vector: Float32Array } * @param {Float32Array|number[]} queryVector - R2 weighted query vector @@ -760,7 +758,7 @@ function fillMetricsEmpty(metrics) { cosineGateNoVector: 0, finalCount: 0, scoreDistribution: { min: 0, max: 0, mean: 0 }, - byChannel: { who: 0, what: 0, where: 0, how: 0 }, + byChannel: { what: 0, where: 0, how: 0 }, time: 0, }; } @@ -782,7 +780,7 @@ function fillMetrics(metrics, data) { cosineGateNoVector: data.cosineGateNoVector || 0, finalCount: data.finalCount || 0, scoreDistribution: data.scoreDistribution || { min: 0, max: 0, mean: 0 }, - byChannel: data.channelStats || { who: 0, what: 0, where: 0, how: 0 }, + byChannel: data.channelStats || { what: 0, where: 0, how: 0 }, time: data.time || 0, }; } diff --git a/modules/story-summary/vector/retrieval/entity-lexicon.js b/modules/story-summary/vector/retrieval/entity-lexicon.js index a9af031..ddfd65a 100644 --- a/modules/story-summary/vector/retrieval/entity-lexicon.js +++ b/modules/story-summary/vector/retrieval/entity-lexicon.js @@ -71,11 +71,12 @@ export function buildEntityLexicon(store, context) { add(f.s); } - // 5. L0 atoms 的 who（新角色在 L2 总结前即可进入词典） + // 5. L0 atoms 的 edges.s/edges.t（新角色在 L2 总结前即可进入词典） const atoms = getStateAtoms(); for (const atom of atoms) { - for (const name of (atom.who || [])) { - add(name); + for (const e of (atom.edges || [])) { + add(e?.s); + add(e?.t); } } @@ -122,11 +123,12 @@ export function buildDisplayNameMap(store, context) { if (!f.retracted) register(f.s); } - // 5. L0 atoms 的 who + // 5. L0 atoms 的 edges.s/edges.t const atoms = getStateAtoms(); for (const atom of atoms) { - for (const name of (atom.who || [])) { - register(name); + for (const e of (atom.edges || [])) { + register(e?.s); + register(e?.t); } } diff --git a/modules/story-summary/vector/retrieval/metrics.js b/modules/story-summary/vector/retrieval/metrics.js index fb6764d..c75a794 100644 --- a/modules/story-summary/vector/retrieval/metrics.js +++ b/modules/story-summary/vector/retrieval/metrics.js @@ -125,7 +125,7 @@ export function createMetrics() { cosineGateNoVector: 0, finalCount: 0, scoreDistribution: { min: 0, max: 0, mean: 0 }, - byChannel: { who: 0, what: 0, where: 0, how: 0 }, + byChannel: { what: 0, where: 0, how: 0 }, time: 0, }, @@ -366,7 +366,7 @@ export function formatMetricsLog(metrics) { lines.push(`├─ graph: ${m.diffusion.graphNodes} nodes, ${m.diffusion.graphEdges} edges`); if (m.diffusion.graphEdges > 0) { const ch = m.diffusion.byChannel || {}; - lines.push(`│ └─ by_channel: who=${ch.who || 0}, what=${ch.what || 0}, where=${ch.where || 0}, how=${ch.how || 0}`); + lines.push(`│ └─ by_channel: what=${ch.what || 0}, where=${ch.where || 0}, how=${ch.how || 0}`); } if (m.diffusion.iterations > 0) { lines.push(`├─ ppr: ${m.diffusion.iterations} iterations, ε=${Number(m.diffusion.convergenceError).toExponential(1)}`); @@ -630,7 +630,7 @@ export function detectIssues(metrics) { // ───────────────────────────────────────────────────────────────── if (m.diffusion.graphEdges === 0 && m.diffusion.seedCount > 0) { - issues.push('No diffusion graph edges - atoms may lack who/edges fields'); + issues.push('No diffusion graph edges - atoms may lack edges fields'); } if (m.diffusion.pprActivated > 0 && m.diffusion.cosineGatePassed === 0) { diff --git a/modules/story-summary/vector/retrieval/recall.js b/modules/story-summary/vector/retrieval/recall.js index ca2ca88..75382f0 100644 --- a/modules/story-summary/vector/retrieval/recall.js +++ b/modules/story-summary/vector/retrieval/recall.js @@ -20,7 +20,9 @@ // 阶段 5: Lexical Retrieval + Dense-Gated Event Merge // 阶段 6: Floor W-RRF Fusion + Rerank + L1 配对 // 阶段 7: L1 配对组装（L0 → top-1 AI L1 + top-1 USER L1） -// 阶段 8: Causation Trace +// 阶段 7.5: PPR Diffusion +// 阶段 8: L0 → L2 反向查找（后置，基于最终 l0Selected） +// 阶段 9: Causation Trace // ═══════════════════════════════════════════════════════════════════════════ import { getAllEventVectors, getChunksByFloors, getMeta, getChunkVectorsByIds } from '../storage/chunk-store.js'; @@ -1114,6 +1116,7 @@ export async function recallMemory(allEvents, vectorConfig, options = {}) { const eventIndex = buildEventIndex(allEvents); let lexicalEventCount = 0; let lexicalEventFilteredByDense = 0; + let l0LinkedCount = 0; const focusSetForLexical = new Set((bundle.focusEntities || []).map(normalize)); for (const eid of lexicalResult.eventIds) { @@ -1149,46 +1152,6 @@ export async function recallMemory(allEvents, vectorConfig, options = {}) { lexicalEventCount++; } - // ═══════════════════════════════════════════════════════════════════ - // 阶段 5.5: L0 → L2 反向查找 - // 已召回的 L0 楼层落在某 L2 事件范围内，但该 L2 自身未被召回 - // ═══════════════════════════════════════════════════════════════════ - - const recalledL0Floors = new Set(anchorHits.map(h => h.floor)); - let l0LinkedCount = 0; - - for (const event of allEvents) { - if (existingEventIds.has(event.id)) continue; - - const range = parseFloorRange(event.summary); - if (!range) continue; - - let hasOverlap = false; - for (const floor of recalledL0Floors) { - if (floor >= range.start && floor <= range.end) { - hasOverlap = true; - break; - } - } - if (!hasOverlap) continue; - - // 实体分类：与所有路径统一标准 - const participants = (event.participants || []).map(p => normalize(p)); - const hasEntityMatch = focusSetForLexical.size > 0 - && participants.some(p => focusSetForLexical.has(p)); - - const evVec = eventVectorMap.get(event.id); - const sim = evVec?.length ? cosineSimilarity(queryVector_v1, evVec) : 0; - - eventHits.push({ - event, - similarity: sim, - _recallType: hasEntityMatch ? 'DIRECT' : 'RELATED', - }); - existingEventIds.add(event.id); - l0LinkedCount++; - } - if (metrics) { metrics.lexical.eventFilteredByDense = lexicalEventFilteredByDense; @@ -1196,14 +1159,10 @@ export async function recallMemory(allEvents, vectorConfig, options = {}) { metrics.event.byRecallType.lexical = lexicalEventCount; metrics.event.selected += lexicalEventCount; } - if (l0LinkedCount > 0) { - metrics.event.byRecallType.l0Linked = l0LinkedCount; - metrics.event.selected += l0LinkedCount; - } } xbLog.info(MODULE_ID, - `Lexical: chunks=${lexicalResult.chunkIds.length} events=${lexicalResult.eventIds.length} mergedEvents=+${lexicalEventCount} filteredByDense=${lexicalEventFilteredByDense} l0Linked=+${l0LinkedCount} (${lexTime}ms)` + `Lexical: chunks=${lexicalResult.chunkIds.length} events=${lexicalResult.eventIds.length} mergedEvents=+${lexicalEventCount} filteredByDense=${lexicalEventFilteredByDense} floorFiltered=${metrics.lexical.floorFilteredByDense || 0} (${lexTime}ms)` ); // ═══════════════════════════════════════════════════════════════════ @@ -1248,7 +1207,56 @@ export async function recallMemory(allEvents, vectorConfig, options = {}) { metrics.timing.diffusion = metrics.diffusion?.time || 0; // ═══════════════════════════════════════════════════════════════════ - // 阶段 7: Causation Trace + // Stage 8: L0 → L2 反向查找（后置，基于最终 l0Selected） + // ═══════════════════════════════════════════════════════════════════ + + const recalledL0Floors = new Set(l0Selected.map(x => x.floor)); + + for (const event of allEvents) { + if (existingEventIds.has(event.id)) continue; + + const range = parseFloorRange(event.summary); + if (!range) continue; + + let hasOverlap = false; + for (const floor of recalledL0Floors) { + if (floor >= range.start && floor <= range.end) { + hasOverlap = true; + break; + } + } + if (!hasOverlap) continue; + + // Dense similarity 门槛（与 Lexical Event 对齐） + const evVec = eventVectorMap.get(event.id); + const sim = evVec?.length ? cosineSimilarity(queryVector_v1, evVec) : 0; + if (sim < CONFIG.LEXICAL_EVENT_DENSE_MIN) continue; + + // 实体分类：与所有路径统一标准 + const participants = (event.participants || []).map(p => normalize(p)); + const hasEntityMatch = focusSetForLexical.size > 0 + && participants.some(p => focusSetForLexical.has(p)); + + eventHits.push({ + event, + similarity: sim, + _recallType: hasEntityMatch ? 'DIRECT' : 'RELATED', + }); + existingEventIds.add(event.id); + l0LinkedCount++; + } + + if (metrics && l0LinkedCount > 0) { + metrics.event.byRecallType.l0Linked = l0LinkedCount; + metrics.event.selected += l0LinkedCount; + } + + xbLog.info(MODULE_ID, + `L0-linked events: ${recalledL0Floors.size} floors → ${l0LinkedCount} events linked (sim≥${CONFIG.LEXICAL_EVENT_DENSE_MIN})` + ); + + // ═══════════════════════════════════════════════════════════════════ + // 阶段 9: Causation Trace // ═══════════════════════════════════════════════════════════════════ const { results: causalMap, maxDepth: causalMaxDepth } = traceCausation(eventHits, eventIndex); @@ -1288,7 +1296,7 @@ export async function recallMemory(allEvents, vectorConfig, options = {}) { console.log(`Fusion (floor, weighted): dense=${metrics.fusion.denseFloors} lex=${metrics.fusion.lexFloors} → cap=${metrics.fusion.afterCap} (${metrics.fusion.time}ms)`); console.log(`Floor Rerank: ${metrics.evidence.beforeRerank || 0} → ${metrics.evidence.floorsSelected || 0} floors → L0=${metrics.evidence.l0Collected || 0} (${metrics.evidence.rerankTime || 0}ms)`); console.log(`L1: ${metrics.evidence.l1Pulled || 0} pulled → ${metrics.evidence.l1Attached || 0} attached (${metrics.evidence.l1CosineTime || 0}ms)`); - console.log(`Events: ${eventHits.length} hits, ${causalChain.length} causal`); + console.log(`Events: ${eventHits.length} hits (l0Linked=+${l0LinkedCount}), ${causalChain.length} causal`); console.log(`Diffusion: ${metrics.diffusion?.seedCount || 0} seeds → ${metrics.diffusion?.pprActivated || 0} activated → ${metrics.diffusion?.finalCount || 0} final (${metrics.diffusion?.time || 0}ms)`); console.groupEnd();