LittleWhiteBox/modules/story-summary/vector/retrieval/recall.js

// ═══════════════════════════════════════════════════════════════════════════
// Story Summary - Recall Engine (v7 - Two-Stage: L0 Locate → L1 Evidence)
//
// 命名规范：
// - 存储层用 L0/L1/L2/L3（StateAtom/Chunk/Event/Fact）
// - 召回层用语义名称：anchor/evidence/event/constraint
//
// 架构：
// 阶段 1: Query Build（确定性，无 LLM）
// 阶段 2: Round 1 Dense Retrieval（L0 + L2）
// 阶段 3: Query Refinement（用已命中记忆增强）
// 阶段 4: Round 2 Dense Retrieval（L0 + L2）
// 阶段 5: Lexical Retrieval + L0 Merge
// 阶段 6: L0-only W-RRF Fusion + Rerank ‖ 并发 L1 Cosine 预筛选
// 阶段 7: L1 配对组装（L0 → top-1 AI L1 + top-1 USER L1）
// 阶段 8: Causation Trace
// ═══════════════════════════════════════════════════════════════════════════

import { getAllEventVectors, getChunksByFloors, getMeta, getChunkVectorsByIds } from '../storage/chunk-store.js';
import { getAllStateVectors, getStateAtoms } from '../storage/state-store.js';
import { getEngineFingerprint, embed } from '../utils/embedder.js';
import { xbLog } from '../../../../core/debug-core.js';
import { getContext } from '../../../../../../../extensions.js';
import { buildQueryBundle, refineQueryBundle } from './query-builder.js';
import { getLexicalIndex, searchLexicalIndex } from './lexical-index.js';
import { rerankChunks } from '../llm/reranker.js';
import { createMetrics, calcSimilarityStats } from './metrics.js';

const MODULE_ID = 'recall';

// ═══════════════════════════════════════════════════════════════════════════
// 配置
// ═══════════════════════════════════════════════════════════════════════════

const CONFIG = {
    // 窗口
    LAST_MESSAGES_K: 2,

    // Anchor (L0 StateAtoms)
    ANCHOR_MIN_SIMILARITY: 0.58,

    // Event (L2 Events)
    EVENT_CANDIDATE_MAX: 100,
    EVENT_SELECT_MAX: 50,
    EVENT_MIN_SIMILARITY: 0.55,
    EVENT_MMR_LAMBDA: 0.72,

    // W-RRF 融合（L0-only）
    RRF_K: 60,
    RRF_W_DENSE: 1.0,
    RRF_W_LEX: 0.9,
    FUSION_CAP: 100,

    // Rerank（L0-only）
    RERANK_TOP_N: 50,
    RERANK_MIN_SCORE: 0.15,

    // 因果链
    CAUSAL_CHAIN_MAX_DEPTH: 10,
    CAUSAL_INJECT_MAX: 30,
};

// ═══════════════════════════════════════════════════════════════════════════
// 工具函数
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 计算余弦相似度
 * @param {number[]} a
 * @param {number[]} b
 * @returns {number}
 */
function cosineSimilarity(a, b) {
    if (!a?.length || !b?.length || a.length !== b.length) return 0;
    let dot = 0, nA = 0, nB = 0;
    for (let i = 0; i < a.length; i++) {
        dot += a[i] * b[i];
        nA += a[i] * a[i];
        nB += b[i] * b[i];
    }
    return nA && nB ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
}

/**
 * 标准化字符串
 * @param {string} s
 * @returns {string}
 */
function normalize(s) {
    return String(s || '')
        .normalize('NFKC')
        .replace(/[\u200B-\u200D\uFEFF]/g, '')
        .trim()
        .toLowerCase();
}

/**
 * 获取最近消息
 * @param {object[]} chat
 * @param {number} count
 * @param {boolean} excludeLastAi
 * @returns {object[]}
 */
function getLastMessages(chat, count = 2, excludeLastAi = false) {
    if (!chat?.length) return [];
    let messages = [...chat];
    if (excludeLastAi && messages.length > 0 && !messages[messages.length - 1]?.is_user) {
        messages = messages.slice(0, -1);
    }
    return messages.slice(-count);
}

// ═══════════════════════════════════════════════════════════════════════════
// MMR 选择算法
// ═══════════════════════════════════════════════════════════════════════════

/**
 * Maximal Marginal Relevance 选择
 * @param {object[]} candidates
 * @param {number} k
 * @param {number} lambda
 * @param {Function} getVector
 * @param {Function} getScore
 * @returns {object[]}
 */
function mmrSelect(candidates, k, lambda, getVector, getScore) {
    const selected = [];
    const ids = new Set();

    while (selected.length < k && candidates.length) {
        let best = null;
        let bestScore = -Infinity;

        for (const c of candidates) {
            if (ids.has(c._id)) continue;

            const rel = getScore(c);
            let div = 0;

            if (selected.length) {
                const vC = getVector(c);
                if (vC?.length) {
                    for (const s of selected) {
                        const sim = cosineSimilarity(vC, getVector(s));
                        if (sim > div) div = sim;
                    }
                }
            }

            const score = lambda * rel - (1 - lambda) * div;
            if (score > bestScore) {
                bestScore = score;
                best = c;
            }
        }

        if (!best) break;
        selected.push(best);
        ids.add(best._id);
    }

    return selected;
}

// ═══════════════════════════════════════════════════════════════════════════
// [Anchors] L0 StateAtoms 检索
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 检索语义锚点
 * @param {number[]} queryVector
 * @param {object} vectorConfig
 * @param {object|null} metrics
 * @returns {Promise<{hits: object[], floors: Set<number>}>}
 */
async function recallAnchors(queryVector, vectorConfig, metrics) {
    const { chatId } = getContext();
    if (!chatId || !queryVector?.length) {
        return { hits: [], floors: new Set() };
    }

    const meta = await getMeta(chatId);
    const fp = getEngineFingerprint(vectorConfig);
    if (meta.fingerprint && meta.fingerprint !== fp) {
        xbLog.warn(MODULE_ID, 'Anchor fingerprint 不匹配');
        return { hits: [], floors: new Set() };
    }

    const stateVectors = await getAllStateVectors(chatId);
    if (!stateVectors.length) {
        return { hits: [], floors: new Set() };
    }

    const atomsList = getStateAtoms();
    const atomMap = new Map(atomsList.map(a => [a.atomId, a]));

    const scored = stateVectors
        .map(sv => {
            const atom = atomMap.get(sv.atomId);
            if (!atom) return null;
            return {
                atomId: sv.atomId,
                floor: sv.floor,
                similarity: cosineSimilarity(queryVector, sv.vector),
                atom,
            };
        })
        .filter(Boolean)
        .filter(s => s.similarity >= CONFIG.ANCHOR_MIN_SIMILARITY)
        .sort((a, b) => b.similarity - a.similarity);

    const floors = new Set(scored.map(s => s.floor));

    if (metrics) {
        metrics.anchor.matched = scored.length;
        metrics.anchor.floorsHit = floors.size;
        metrics.anchor.topHits = scored.slice(0, 5).map(s => ({
            floor: s.floor,
            semantic: s.atom?.semantic?.slice(0, 50),
            similarity: Math.round(s.similarity * 1000) / 1000,
        }));
    }

    return { hits: scored, floors };
}

// ═══════════════════════════════════════════════════════════════════════════
// [Events] L2 Events 检索
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 检索事件
 * @param {number[]} queryVector
 * @param {object[]} allEvents
 * @param {object} vectorConfig
 * @param {string[]} focusEntities
 * @param {object|null} metrics
 * @returns {Promise<object[]>}
 */
async function recallEvents(queryVector, allEvents, vectorConfig, focusEntities, metrics) {
    const { chatId } = getContext();
    if (!chatId || !queryVector?.length || !allEvents?.length) {
        return [];
    }

    const meta = await getMeta(chatId);
    const fp = getEngineFingerprint(vectorConfig);
    if (meta.fingerprint && meta.fingerprint !== fp) {
        xbLog.warn(MODULE_ID, 'Event fingerprint 不匹配');
        return [];
    }

    const eventVectors = await getAllEventVectors(chatId);
    const vectorMap = new Map(eventVectors.map(v => [v.eventId, v.vector]));

    if (!vectorMap.size) {
        return [];
    }

    const focusSet = new Set((focusEntities || []).map(normalize));

    const scored = allEvents.map(event => {
        const v = vectorMap.get(event.id);
        const baseSim = v ? cosineSimilarity(queryVector, v) : 0;

        const participants = (event.participants || []).map(p => normalize(p));
        const hasEntityMatch = participants.some(p => focusSet.has(p));

        return {
            _id: event.id,
            event,
            similarity: baseSim,
            _hasEntityMatch: hasEntityMatch,
            vector: v,
        };
    });

    if (metrics) {
        metrics.event.inStore = allEvents.length;
    }

    let candidates = scored
        .filter(s => s.similarity >= CONFIG.EVENT_MIN_SIMILARITY)
        .sort((a, b) => b.similarity - a.similarity)
        .slice(0, CONFIG.EVENT_CANDIDATE_MAX);

    if (metrics) {
        metrics.event.considered = candidates.length;
    }

    // 实体过滤
    if (focusSet.size > 0) {
        const beforeFilter = candidates.length;

        candidates = candidates.filter(c => {
            if (c.similarity >= 0.85) return true;
            return c._hasEntityMatch;
        });

        if (metrics) {
            metrics.event.entityFilter = {
                focusEntities: focusEntities || [],
                before: beforeFilter,
                after: candidates.length,
                filtered: beforeFilter - candidates.length,
            };
        }
    }

    // MMR 选择
    const selected = mmrSelect(
        candidates,
        CONFIG.EVENT_SELECT_MAX,
        CONFIG.EVENT_MMR_LAMBDA,
        c => c.vector,
        c => c.similarity
    );

    let directCount = 0;
    let relatedCount = 0;

    const results = selected.map(s => {
        const recallType = s._hasEntityMatch ? 'DIRECT' : 'RELATED';
        if (recallType === 'DIRECT') directCount++;
        else relatedCount++;

        return {
            event: s.event,
            similarity: s.similarity,
            _recallType: recallType,
        };
    });

    if (metrics) {
        metrics.event.selected = results.length;
        metrics.event.byRecallType = { direct: directCount, related: relatedCount, causal: 0, lexical: 0 };
        metrics.event.similarityDistribution = calcSimilarityStats(results.map(r => r.similarity));
    }

    return results;
}

// ═══════════════════════════════════════════════════════════════════════════
// [Causation] 因果链追溯
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 构建事件索引
 * @param {object[]} allEvents
 * @returns {Map<string, object>}
 */
function buildEventIndex(allEvents) {
    const map = new Map();
    for (const e of allEvents || []) {
        if (e?.id) map.set(e.id, e);
    }
    return map;
}

/**
 * 追溯因果链
 * @param {object[]} eventHits
 * @param {Map<string, object>} eventIndex
 * @param {number} maxDepth
 * @returns {{results: object[], maxDepth: number}}
 */
function traceCausation(eventHits, eventIndex, maxDepth = CONFIG.CAUSAL_CHAIN_MAX_DEPTH) {
    const out = new Map();
    const idRe = /^evt-\d+$/;
    let maxActualDepth = 0;

    function visit(parentId, depth, chainFrom) {
        if (depth > maxDepth) return;
        if (!idRe.test(parentId)) return;

        const ev = eventIndex.get(parentId);
        if (!ev) return;

        if (depth > maxActualDepth) maxActualDepth = depth;

        const existed = out.get(parentId);
        if (!existed) {
            out.set(parentId, { event: ev, depth, chainFrom: [chainFrom] });
        } else {
            if (depth < existed.depth) existed.depth = depth;
            if (!existed.chainFrom.includes(chainFrom)) existed.chainFrom.push(chainFrom);
        }

        for (const next of (ev.causedBy || [])) {
            visit(String(next || '').trim(), depth + 1, chainFrom);
        }
    }

    for (const r of eventHits || []) {
        const rid = r?.event?.id;
        if (!rid) continue;
        for (const cid of (r.event?.causedBy || [])) {
            visit(String(cid || '').trim(), 1, rid);
        }
    }

    const results = Array.from(out.values())
        .sort((a, b) => {
            const refDiff = b.chainFrom.length - a.chainFrom.length;
            if (refDiff !== 0) return refDiff;
            return a.depth - b.depth;
        })
        .slice(0, CONFIG.CAUSAL_INJECT_MAX);

    return { results, maxDepth: maxActualDepth };
}

// ═══════════════════════════════════════════════════════════════════════════
// [W-RRF] 加权倒数排名融合（L0-only）
// ═══════════════════════════════════════════════════════════════════════════

/**
 * @typedef {object} RankedItem
 * @property {string} id - 唯一标识符
 * @property {number} score - 该路的原始分数
 */

/**
 * W-RRF 融合两路 L0 候选（dense + lexical）
 *
 * @param {RankedItem[]} denseRank  - Dense 路（cosine 降序）
 * @param {RankedItem[]} lexRank    - Lexical 路（MiniSearch score 降序）
 * @param {number} cap - 输出上限
 * @returns {{top: {id: string, fusionScore: number}[], totalUnique: number}}
 */
function fuseL0Candidates(denseRank, lexRank, cap = CONFIG.FUSION_CAP) {
    const k = CONFIG.RRF_K;
    const wD = CONFIG.RRF_W_DENSE;
    const wL = CONFIG.RRF_W_LEX;

    const buildRankMap = (ranked) => {
        const map = new Map();
        for (let i = 0; i < ranked.length; i++) {
            const id = ranked[i].id;
            if (!map.has(id)) map.set(id, i);
        }
        return map;
    };

    const denseMap = buildRankMap(denseRank || []);
    const lexMap = buildRankMap(lexRank || []);

    const allIds = new Set([
        ...denseMap.keys(),
        ...lexMap.keys(),
    ]);

    const totalUnique = allIds.size;

    const scored = [];
    for (const id of allIds) {
        let score = 0;
        if (denseMap.has(id)) {
            score += wD / (k + denseMap.get(id));
        }
        if (lexMap.has(id)) {
            score += wL / (k + lexMap.get(id));
        }
        scored.push({ id, fusionScore: score });
    }

    scored.sort((a, b) => b.fusionScore - a.fusionScore);

    return {
        top: scored.slice(0, cap),
        totalUnique,
    };
}

// ═══════════════════════════════════════════════════════════════════════════
// [Stage 6] L0-only 融合 + Rerank ‖ 并发 L1 Cosine 预筛选
// ═══════════════════════════════════════════════════════════════════════════

/**
 * L0 融合 + rerank，并发拉取 L1 并 cosine 打分
 *
 * @param {object[]} anchorHits - L0 dense 命中（Round 2）
 * @param {Set<number>} anchorFloors - L0 命中楼层（含 lexical 扩展）
 * @param {number[]} queryVector - 查询向量（v1）
 * @param {string} rerankQuery - rerank 查询文本
 * @param {object} lexicalResult - 词法检索结果
 * @param {object} metrics
 * @returns {Promise<{l0Selected: object[], l1ByFloor: Map<number, {aiTop1: object|null, userTop1: object|null}>}>}
 */
async function locateAndPullEvidence(anchorHits, anchorFloors, queryVector, rerankQuery, lexicalResult, metrics) {
    const { chatId, chat } = getContext();
    if (!chatId) return { l0Selected: [], l1ByFloor: new Map() };

    const T_Start = performance.now();

    // ─────────────────────────────────────────────────────────────────
    // 6a. 构建 L0 候选对象（用于 rerank）
    //
    // 重要：支持 lexical-only 的 L0（atom）进入候选池。
    // 否则 hybrid 会退化为 dense-only：lexical 命中的 atom 若未被 dense 命中会被直接丢弃。
    // ─────────────────────────────────────────────────────────────────

    const l0ObjectMap = new Map();
    for (const a of (anchorHits || [])) {
        const id = `anchor-${a.atomId}`;
        l0ObjectMap.set(id, {
            id,
            atomId: a.atomId,
            floor: a.floor,
            similarity: a.similarity,
            atom: a.atom,
            text: a.atom?.semantic || '',
        });
    }

    // lexical-only atoms：从全量 StateAtoms 补齐（similarity 记为 0，靠 lex rank 贡献 W-RRF）
    const lexAtomIds = lexicalResult?.atomIds || [];
    if (lexAtomIds.length > 0) {
        const atomsList = getStateAtoms();
        const atomMap = new Map(atomsList.map(a => [a.atomId, a]));

        for (const atomId of lexAtomIds) {
            const id = `anchor-${atomId}`;
            if (l0ObjectMap.has(id)) continue;

            const atom = atomMap.get(atomId);
            if (!atom) continue;
            if (typeof atom.floor !== 'number' || atom.floor < 0) continue;

            l0ObjectMap.set(id, {
                id,
                atomId,
                floor: atom.floor,
                similarity: 0,
                atom,
                text: atom.semantic || '',
            });
        }
    }

    // ─────────────────────────────────────────────────────────────────
    // 6b. 构建两路排名（L0-only）
    // ─────────────────────────────────────────────────────────────────

    // Dense 路：anchorHits 按 similarity 排序
    const denseRank = (anchorHits || [])
        .map(a => ({ id: `anchor-${a.atomId}`, score: a.similarity }))
        .sort((a, b) => b.score - a.score);

    // Lexical 路：从 lexicalResult.atomIds 构建排名（允许 lexical-only）
    // atomIds 已按 MiniSearch score 排序（searchLexicalIndex 返回顺序）；W-RRF 依赖 rank，score 为占位
    const lexRank = (lexAtomIds || [])
        .map(atomId => ({ id: `anchor-${atomId}`, score: 1 }))
        .filter(item => l0ObjectMap.has(item.id));

    // ─────────────────────────────────────────────────────────────────
    // 6c. W-RRF 融合（L0-only）
    // ─────────────────────────────────────────────────────────────────

    const T_Fusion_Start = performance.now();

    const { top: fusionResult, totalUnique } = fuseL0Candidates(denseRank, lexRank, CONFIG.FUSION_CAP);

    const fusionTime = Math.round(performance.now() - T_Fusion_Start);

    // 构建 rerank 候选列表
    const rerankCandidates = fusionResult
        .map(f => l0ObjectMap.get(f.id))
        .filter(Boolean);

    if (metrics) {
        metrics.fusion.denseCount = denseRank.length;
        metrics.fusion.lexCount = lexRank.length;
        metrics.fusion.totalUnique = totalUnique;
        metrics.fusion.afterCap = rerankCandidates.length;
        metrics.fusion.time = fusionTime;
        metrics.evidence.l0Candidates = rerankCandidates.length;
    }

    if (rerankCandidates.length === 0) {
        if (metrics) {
            metrics.evidence.l0Selected = 0;
            metrics.evidence.l1Pulled = 0;
            metrics.evidence.l1Attached = 0;
            metrics.evidence.l1CosineTime = 0;
            metrics.evidence.rerankApplied = false;
        }
        return { l0Selected: [], l1ByFloor: new Map() };
    }

    // ─────────────────────────────────────────────────────────────────
    // 6d. 收集所有候选 L0 的楼层（用于并发拉取 L1）
    //     包含 AI 楼层本身 + 上方 USER 楼层
    // ─────────────────────────────────────────────────────────────────

    const candidateFloors = new Set();
    for (const c of rerankCandidates) {
        candidateFloors.add(c.floor);
        // 上方 USER 楼层
        const userFloor = c.floor - 1;
        if (userFloor >= 0 && chat?.[userFloor]?.is_user) {
            candidateFloors.add(userFloor);
        }
    }

    // ─────────────────────────────────────────────────────────────────
    // 6e. 并发：rerank L0 ‖ 拉取 L1 chunks + 向量 + cosine 打分
    // ─────────────────────────────────────────────────────────────────

    const T_Rerank_Start = performance.now();

    // 并发任务 1：rerank L0
    const rerankPromise = rerankChunks(rerankQuery, rerankCandidates, {
        topN: CONFIG.RERANK_TOP_N,
        minScore: CONFIG.RERANK_MIN_SCORE,
    });

    // 并发任务 2：拉取 L1 chunks + 向量 → cosine 打分
    const l1Promise = pullAndScoreL1(chatId, Array.from(candidateFloors), queryVector, chat);

    // 等待两个任务完成
    const [rerankedL0, l1ScoredByFloor] = await Promise.all([rerankPromise, l1Promise]);

    const rerankTime = Math.round(performance.now() - T_Rerank_Start);

    // ─────────────────────────────────────────────────────────────────
    // 6f. 记录 rerank metrics
    // ─────────────────────────────────────────────────────────────────

    if (metrics) {
        metrics.evidence.rerankApplied = true;
        metrics.evidence.beforeRerank = rerankCandidates.length;
        metrics.evidence.afterRerank = rerankedL0.length;
        metrics.evidence.l0Selected = rerankedL0.length;
        metrics.evidence.rerankTime = rerankTime;
        metrics.timing.evidenceRerank = rerankTime;

        const scores = rerankedL0.map(c => c._rerankScore || 0).filter(s => s > 0);
        if (scores.length > 0) {
            scores.sort((a, b) => a - b);
            metrics.evidence.rerankScores = {
                min: Number(scores[0].toFixed(3)),
                max: Number(scores[scores.length - 1].toFixed(3)),
                mean: Number((scores.reduce((a, b) => a + b, 0) / scores.length).toFixed(3)),
            };
        }
    }

    // ─────────────────────────────────────────────────────────────────
    // 6g. 构建最终 l0Selected + l1ByFloor
    // ─────────────────────────────────────────────────────────────────

    const l0Selected = rerankedL0.map(item => ({
        id: item.id,
        atomId: item.atomId,
        floor: item.floor,
        similarity: item.similarity,
        rerankScore: item._rerankScore || 0,
        atom: item.atom,
        text: item.text,
    }));

    // 为每个选中的 L0 楼层组装 top-1 L1 配对
    const selectedFloors = new Set(l0Selected.map(l => l.floor));
    const l1ByFloor = new Map();
    let contextPairsAdded = 0;

    for (const floor of selectedFloors) {
        const aiChunks = l1ScoredByFloor.get(floor) || [];
        const userFloor = floor - 1;
        const userChunks = (userFloor >= 0 && chat?.[userFloor]?.is_user)
            ? (l1ScoredByFloor.get(userFloor) || [])
            : [];

        // top-1：取 cosine 最高的
        const aiTop1 = aiChunks.length > 0
            ? aiChunks.reduce((best, c) => (c._cosineScore > best._cosineScore ? c : best))
            : null;

        const userTop1 = userChunks.length > 0
            ? userChunks.reduce((best, c) => (c._cosineScore > best._cosineScore ? c : best))
            : null;

        // context pair = 上方 USER 楼层成功挂载（用于 metrics）
        if (userTop1) contextPairsAdded++;

        l1ByFloor.set(floor, { aiTop1, userTop1 });
    }

    // ─────────────────────────────────────────────────────────────────
    // 6h. L1 metrics
    // ─────────────────────────────────────────────────────────────────

    if (metrics) {
        let totalPulled = 0;
        let totalAttached = 0;

        for (const [, scored] of l1ScoredByFloor) {
            totalPulled += scored.length;
        }

        for (const [, pair] of l1ByFloor) {
            if (pair.aiTop1) totalAttached++;
            if (pair.userTop1) totalAttached++;
        }

        metrics.evidence.l1Pulled = totalPulled;
        metrics.evidence.l1Attached = totalAttached;
        metrics.evidence.contextPairsAdded = contextPairsAdded;
        metrics.evidence.l1CosineTime = l1ScoredByFloor._cosineTime || 0;
    }

    const totalTime = Math.round(performance.now() - T_Start);
    if (metrics) {
        metrics.timing.evidenceRetrieval = Math.max(0, totalTime - fusionTime - rerankTime);
    }

    xbLog.info(MODULE_ID,
        `Evidence: ${anchorHits?.length || 0} L0 dense → fusion=${rerankCandidates.length} → rerank=${rerankedL0.length} → L1 attached=${metrics?.evidence?.l1Attached || 0} (${totalTime}ms)`
    );

    return { l0Selected, l1ByFloor };
}

// ═══════════════════════════════════════════════════════════════════════════
// [L1] 拉取 + Cosine 打分（并发子任务）
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 从 IndexedDB 拉取指定楼层的 L1 chunks + 向量，用 queryVector cosine 打分
 *
 * @param {string} chatId
 * @param {number[]} floors - 需要拉取的楼层列表
 * @param {number[]} queryVector - 查询向量（v1）
 * @param {object[]} chat - 聊天消息数组
 * @returns {Promise<Map<number, object[]>>} floor → scored chunks（带 _cosineScore）
 */
async function pullAndScoreL1(chatId, floors, queryVector, chat) {
    const T0 = performance.now();

    /** @type {Map<number, object[]>} */
    const result = new Map();

    if (!chatId || !floors?.length || !queryVector?.length) {
        result._cosineTime = 0;
        return result;
    }

    // 拉取 chunks
    let dbChunks = [];
    try {
        dbChunks = await getChunksByFloors(chatId, floors);
    } catch (e) {
        xbLog.warn(MODULE_ID, 'L1 chunks 拉取失败', e);
        result._cosineTime = Math.round(performance.now() - T0);
        return result;
    }

    if (!dbChunks.length) {
        result._cosineTime = Math.round(performance.now() - T0);
        return result;
    }

    // 拉取向量
    const chunkIds = dbChunks.map(c => c.chunkId);
    let chunkVectors = [];
    try {
        chunkVectors = await getChunkVectorsByIds(chatId, chunkIds);
    } catch (e) {
        xbLog.warn(MODULE_ID, 'L1 向量拉取失败', e);
        result._cosineTime = Math.round(performance.now() - T0);
        return result;
    }

    const vectorMap = new Map(chunkVectors.map(v => [v.chunkId, v.vector]));

    // Cosine 打分 + 按楼层分组
    for (const chunk of dbChunks) {
        const vec = vectorMap.get(chunk.chunkId);
        const cosineScore = vec?.length ? cosineSimilarity(queryVector, vec) : 0;

        const scored = {
            chunkId: chunk.chunkId,
            floor: chunk.floor,
            chunkIdx: chunk.chunkIdx,
            speaker: chunk.speaker,
            isUser: chunk.isUser,
            text: chunk.text,
            _cosineScore: cosineScore,
        };

        if (!result.has(chunk.floor)) {
            result.set(chunk.floor, []);
        }
        result.get(chunk.floor).push(scored);
    }

    // 每楼层按 cosine 降序排序
    for (const [, chunks] of result) {
        chunks.sort((a, b) => b._cosineScore - a._cosineScore);
    }

    result._cosineTime = Math.round(performance.now() - T0);

    xbLog.info(MODULE_ID,
        `L1 pull: ${floors.length} floors → ${dbChunks.length} chunks → scored (${result._cosineTime}ms)`
    );

    return result;
}

// ═══════════════════════════════════════════════════════════════════════════
// 主函数
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 执行记忆召回
 *
 * @param {object[]} allEvents - 所有事件（L2）
 * @param {object} vectorConfig - 向量配置
 * @param {object} options
 * @param {boolean} options.excludeLastAi
 * @param {string|null} options.pendingUserMessage
 * @returns {Promise<object>}
 */
export async function recallMemory(allEvents, vectorConfig, options = {}) {
    const T0 = performance.now();
    const { chat } = getContext();
    const { pendingUserMessage = null, excludeLastAi = false } = options;

    const metrics = createMetrics();

    if (!allEvents?.length) {
        metrics.anchor.needRecall = false;
        metrics.timing.total = Math.round(performance.now() - T0);
        return {
            events: [],
            l0Selected: [],
            l1ByFloor: new Map(),
            causalChain: [],
            focusEntities: [],
            elapsed: metrics.timing.total,
            logText: 'No events.',
            metrics,
        };
    }

    metrics.anchor.needRecall = true;

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 1: Query Build
    // ═══════════════════════════════════════════════════════════════════

    const T_Build_Start = performance.now();

    const lastMessages = getLastMessages(chat, CONFIG.LAST_MESSAGES_K, excludeLastAi);

    const bundle = buildQueryBundle(lastMessages, pendingUserMessage);

    metrics.query.buildTime = Math.round(performance.now() - T_Build_Start);
    metrics.anchor.focusEntities = bundle.focusEntities;

    if (metrics.query?.lengths) {
        metrics.query.lengths.v0Chars = String(bundle.queryText_v0 || '').length;
        metrics.query.lengths.v1Chars = null;
        metrics.query.lengths.rerankChars = String(bundle.rerankQuery || bundle.queryText_v0 || '').length;
    }

    xbLog.info(MODULE_ID,
        `Query Build: focus=[${bundle.focusEntities.join(',')}] lexTerms=[${bundle.lexicalTerms.slice(0, 5).join(',')}]`
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 2: Round 1 Dense Retrieval
    // ═══════════════════════════════════════════════════════════════════

    let queryVector_v0;
    try {
        const [vec] = await embed([bundle.queryText_v0], vectorConfig, { timeout: 10000 });
        queryVector_v0 = vec;
    } catch (e1) {
        xbLog.warn(MODULE_ID, 'Round 1 向量化失败，500ms 后重试', e1);
        await new Promise(r => setTimeout(r, 500));
        try {
            const [vec] = await embed([bundle.queryText_v0], vectorConfig, { timeout: 15000 });
            queryVector_v0 = vec;
        } catch (e2) {
            xbLog.error(MODULE_ID, 'Round 1 向量化重试仍失败', e2);
            metrics.timing.total = Math.round(performance.now() - T0);
            return {
                events: [], l0Selected: [], l1ByFloor: new Map(), causalChain: [],
                focusEntities: bundle.focusEntities,
                elapsed: metrics.timing.total,
                logText: 'Embedding failed (round 1, after retry).',
                metrics,
            };
        }
    }

    if (!queryVector_v0?.length) {
        metrics.timing.total = Math.round(performance.now() - T0);
        return {
            events: [], l0Selected: [], l1ByFloor: new Map(), causalChain: [],
            focusEntities: bundle.focusEntities,
            elapsed: metrics.timing.total,
            logText: 'Empty query vector (round 1).',
            metrics,
        };
    }

    const T_R1_Anchor_Start = performance.now();
    const { hits: anchorHits_v0 } = await recallAnchors(queryVector_v0, vectorConfig, null);
    const r1AnchorTime = Math.round(performance.now() - T_R1_Anchor_Start);

    const T_R1_Event_Start = performance.now();
    const eventHits_v0 = await recallEvents(queryVector_v0, allEvents, vectorConfig, bundle.focusEntities, null);
    const r1EventTime = Math.round(performance.now() - T_R1_Event_Start);

    xbLog.info(MODULE_ID,
        `Round 1: anchors=${anchorHits_v0.length} events=${eventHits_v0.length} (anchor=${r1AnchorTime}ms event=${r1EventTime}ms)`
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 3: Query Refinement
    // ═══════════════════════════════════════════════════════════════════

    const T_Refine_Start = performance.now();

    refineQueryBundle(bundle, anchorHits_v0, eventHits_v0);

    metrics.query.refineTime = Math.round(performance.now() - T_Refine_Start);
    metrics.anchor.focusEntities = bundle.focusEntities;

    if (metrics.query?.lengths) {
        metrics.query.lengths.v1Chars = bundle.queryText_v1 == null ? null : String(bundle.queryText_v1).length;
        metrics.query.lengths.rerankChars = String(bundle.rerankQuery || bundle.queryText_v1 || bundle.queryText_v0 || '').length;
    }

    xbLog.info(MODULE_ID,
        `Refinement: focus=[${bundle.focusEntities.join(',')}] hasV1=${!!bundle.queryText_v1} (${metrics.query.refineTime}ms)`
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 4: Round 2 Dense Retrieval
    // ═══════════════════════════════════════════════════════════════════

    const queryTextFinal = bundle.queryText_v1 || bundle.queryText_v0;

    let queryVector_v1;
    try {
        const [vec] = await embed([queryTextFinal], vectorConfig, { timeout: 10000 });
        queryVector_v1 = vec;
    } catch (e) {
        xbLog.warn(MODULE_ID, 'Round 2 向量化失败，降级使用 Round 1 向量', e);
        queryVector_v1 = queryVector_v0;
    }

    const T_R2_Anchor_Start = performance.now();
    const { hits: anchorHits, floors: anchorFloors_dense } = await recallAnchors(queryVector_v1, vectorConfig, metrics);
    metrics.timing.anchorSearch = Math.round(performance.now() - T_R2_Anchor_Start);

    const T_R2_Event_Start = performance.now();
    let eventHits = await recallEvents(queryVector_v1, allEvents, vectorConfig, bundle.focusEntities, metrics);
    metrics.timing.eventRetrieval = Math.round(performance.now() - T_R2_Event_Start);

    xbLog.info(MODULE_ID,
        `Round 2: anchors=${anchorHits.length} floors=${anchorFloors_dense.size} events=${eventHits.length}`
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 5: Lexical Retrieval + L0 Merge
    // ═══════════════════════════════════════════════════════════════════

    const T_Lex_Start = performance.now();

    let lexicalResult = {
        atomIds: [], atomFloors: new Set(),
        chunkIds: [], chunkFloors: new Set(),
        eventIds: [], chunkScores: [], searchTime: 0,
    };

    try {
        const index = await getLexicalIndex();
        if (index) {
            lexicalResult = searchLexicalIndex(index, bundle.lexicalTerms);
        }
    } catch (e) {
        xbLog.warn(MODULE_ID, 'Lexical 检索失败', e);
    }

    const lexTime = Math.round(performance.now() - T_Lex_Start);

    if (metrics) {
        metrics.lexical.atomHits = lexicalResult.atomIds.length;
        metrics.lexical.chunkHits = lexicalResult.chunkIds.length;
        metrics.lexical.eventHits = lexicalResult.eventIds.length;
        metrics.lexical.searchTime = lexTime;
        metrics.lexical.terms = bundle.lexicalTerms.slice(0, 10);
    }

    // 合并 L0 floors（dense + lexical）
    const anchorFloors = new Set(anchorFloors_dense);
    for (const f of lexicalResult.atomFloors) {
        anchorFloors.add(f);
    }

    // 合并 L2 events（lexical 命中但 dense 未命中的 events）
    const existingEventIds = new Set(eventHits.map(e => e.event?.id).filter(Boolean));
    const eventIndex = buildEventIndex(allEvents);
    let lexicalEventCount = 0;

    for (const eid of lexicalResult.eventIds) {
        if (!existingEventIds.has(eid)) {
            const ev = eventIndex.get(eid);
            if (ev) {
                eventHits.push({
                    event: ev,
                    similarity: 0,
                    _recallType: 'LEXICAL',
                });
                existingEventIds.add(eid);
                lexicalEventCount++;
            }
        }
    }

    if (metrics && lexicalEventCount > 0) {
        metrics.event.byRecallType.lexical = lexicalEventCount;
        metrics.event.selected += lexicalEventCount;
    }

    xbLog.info(MODULE_ID,
        `Lexical: atoms=${lexicalResult.atomIds.length} chunks=${lexicalResult.chunkIds.length} events=${lexicalResult.eventIds.length} mergedFloors=${anchorFloors.size} mergedEvents=+${lexicalEventCount} (${lexTime}ms)`
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 6: L0-only W-RRF Fusion + Rerank ‖ 并发 L1 Cosine
    // ═══════════════════════════════════════════════════════════════════

    const { l0Selected, l1ByFloor } = await locateAndPullEvidence(
        anchorHits,
        anchorFloors,
        queryVector_v1,
        bundle.rerankQuery,
        lexicalResult,
        metrics
    );

    // ═══════════════════════════════════════════════════════════════════
    // 阶段 7: Causation Trace
    // ═══════════════════════════════════════════════════════════════════

    const { results: causalMap, maxDepth: causalMaxDepth } = traceCausation(eventHits, eventIndex);

    const recalledIdSet = new Set(eventHits.map(x => x?.event?.id).filter(Boolean));
    const causalChain = causalMap
        .filter(x => x?.event?.id && !recalledIdSet.has(x.event.id))
        .map(x => ({
            event: x.event,
            similarity: 0,
            _recallType: 'CAUSAL',
            _causalDepth: x.depth,
            chainFrom: x.chainFrom,
        }));

    if (metrics.event.byRecallType) {
        metrics.event.byRecallType.causal = causalChain.length;
    }
    metrics.event.causalChainDepth = causalMaxDepth;
    metrics.event.causalCount = causalChain.length;

    // ═══════════════════════════════════════════════════════════════════
    // 完成
    // ═══════════════════════════════════════════════════════════════════

    metrics.timing.total = Math.round(performance.now() - T0);
    metrics.event.entityNames = bundle.focusEntities;
    metrics.event.entitiesUsed = bundle.focusEntities.length;

    console.group('%c[Recall v7]', 'color: #7c3aed; font-weight: bold');
    console.log(`Total: ${metrics.timing.total}ms`);
    console.log(`Query Build: ${metrics.query.buildTime}ms | Refine: ${metrics.query.refineTime}ms`);
    console.log(`Focus: [${bundle.focusEntities.join(', ')}]`);
    console.log(`Round 2 Anchors: ${anchorHits.length} hits → ${anchorFloors.size} floors`);
    console.log(`Lexical: atoms=${lexicalResult.atomIds.length} chunks=${lexicalResult.chunkIds.length} events=${lexicalResult.eventIds.length}`);
    console.log(`Fusion (L0-only): dense=${metrics.fusion.denseCount} lex=${metrics.fusion.lexCount} → cap=${metrics.fusion.afterCap} (${metrics.fusion.time}ms)`);
    console.log(`L0 Rerank: ${metrics.evidence.beforeRerank || 0} → ${metrics.evidence.l0Selected || 0} (${metrics.evidence.rerankTime || 0}ms)`);
    console.log(`L1 Pull: ${metrics.evidence.l1Pulled || 0} chunks → ${metrics.evidence.l1Attached || 0} attached (${metrics.evidence.l1CosineTime || 0}ms)`);
    console.log(`Events: ${eventHits.length} hits, ${causalChain.length} causal`);
    console.groupEnd();

    return {
        events: eventHits,
        causalChain,
        l0Selected,
        l1ByFloor,
        focusEntities: bundle.focusEntities,
        elapsed: metrics.timing.total,
        metrics,
    };
}