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

// ═══════════════════════════════════════════════════════════════════════════
// Story Summary - Recall Engine (v3 - L0 作为 L3 索引 + Rerank 精排)
//
// 架构：
// - Query Expansion → L0（主索引）→ L3（按楼层拉取）→ Rerank（精排）
// - Query Expansion → L2（独立检索）
// - L0 和 L2 不在同一抽象层，分开处理
// ═══════════════════════════════════════════════════════════════════════════

import { getAllEventVectors, getChunksByFloors, getMeta } 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 { filterText } from '../utils/text-filter.js';
import { expandQueryCached, buildSearchText } from '../llm/query-expansion.js';
import { rerankChunks } from '../llm/reranker.js';
import { createMetrics, calcSimilarityStats } from './metrics.js';

const MODULE_ID = 'recall';

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

const CONFIG = {
    // Query Expansion
    QUERY_EXPANSION_TIMEOUT: 6000,

    // L0 配置
    L0_MAX_RESULTS: 30,
    L0_MIN_SIMILARITY: 0.50,

    // L2 配置
    L2_CANDIDATE_MAX: 100,
    L2_SELECT_MAX: 50,
    L2_MIN_SIMILARITY: 0.55,
    L2_MMR_LAMBDA: 0.72,

    // L3 配置（从 L0 楼层拉取）
    L3_MAX_CHUNKS_PER_FLOOR: 3,
    L3_MAX_TOTAL_CHUNKS: 60,

    // Rerank 配置
    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
// ═══════════════════════════════════════════════════════════════════════════

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;
}

function normalize(s) {
    return String(s || '')
        .normalize('NFKC')
        .replace(/[\u200B-\u200D\uFEFF]/g, '')
        .trim()
        .toLowerCase();
}

function cleanForRecall(text) {
    return filterText(text).replace(/\[tts:[^\]]*\]/gi, '').trim();
}

/**
 * 从 focusEntities 中移除用户名
 * @param {Array} focusEntities - 焦点实体
 * @param {string} userName - 用户名
 * @returns {Array} 过滤后的实体
 */
function removeUserNameFromFocus(focusEntities, userName) {
    const u = normalize(userName);
    if (!u) return Array.isArray(focusEntities) ? focusEntities : [];

    return (focusEntities || [])
        .map(e => String(e || '').trim())
        .filter(Boolean)
        .filter(e => normalize(e) !== u);
}

/**
 * 构建用于 Rerank 的查询文本
 * 综合 Query Expansion 结果和最近对话
 * @param {object} expansion - Query Expansion 结果
 * @param {Array} lastMessages - 最近的消息
 * @param {string} pendingUserMessage - 待发送的用户消息
 * @returns {string} Rerank 用的查询文本
 */
function buildRerankQuery(expansion, lastMessages, pendingUserMessage) {
    const parts = [];

    // 1. focus entities
    if (expansion?.focus?.length) {
        parts.push(expansion.focus.join(' '));
    }

    // 2. DSL queries（取前3个）
    if (expansion?.queries?.length) {
        parts.push(...expansion.queries.slice(0, 3));
    }

    // 3. 最近对话的关键内容
    const recentTexts = (lastMessages || [])
        .slice(-2)
        .map(m => cleanForRecall(m.mes || '').slice(0, 150))
        .filter(Boolean);

    if (recentTexts.length) {
        parts.push(...recentTexts);
    }

    // 4. 待发送消息
    if (pendingUserMessage) {
        parts.push(cleanForRecall(pendingUserMessage).slice(0, 200));
    }

    return parts.filter(Boolean).join('\n').slice(0, 1500);
}

// ═══════════════════════════════════════════════════════════════════════════
// MMR 选择
// ═══════════════════════════════════════════════════════════════════════════

/**
 * MMR 多样性选择
 * @param {Array} candidates - 候选项
 * @param {number} k - 选择数量
 * @param {number} lambda - MMR 参数
 * @param {Function} getVector - 获取向量函数
 * @param {Function} getScore - 获取分数函数
 * @returns {Array} 选中的项
 */
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;
}

// ═══════════════════════════════════════════════════════════════════════════
// L0 检索：Query → L0 → 楼层集合
// ═══════════════════════════════════════════════════════════════════════════

/**
 * L0 向量检索
 * @param {Array} queryVector - 查询向量
 * @param {object} vectorConfig - 向量配置
 * @param {object} metrics - 指标对象
 * @returns {Promise<object>} {atoms, floors}
 */
async function searchL0(queryVector, vectorConfig, metrics) {
    const { chatId } = getContext();
    if (!chatId || !queryVector?.length) {
        return { atoms: [], floors: new Set() };
    }

    // 检查 fingerprint
    const meta = await getMeta(chatId);
    const fp = getEngineFingerprint(vectorConfig);
    if (meta.fingerprint && meta.fingerprint !== fp) {
        xbLog.warn(MODULE_ID, 'L0 fingerprint 不匹配');
        return { atoms: [], floors: new Set() };
    }

    // 获取向量
    const stateVectors = await getAllStateVectors(chatId);
    if (!stateVectors.length) {
        return { atoms: [], floors: new Set() };
    }

    // 获取 atoms 元数据
    const atomsList = getStateAtoms();
    const atomMap = new Map(atomsList.map(a => [a.atomId, a]));

    // 计算相似度
    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.L0_MIN_SIMILARITY)
        .sort((a, b) => b.similarity - a.similarity)
        .slice(0, CONFIG.L0_MAX_RESULTS);

    // 收集楼层
    const floors = new Set(scored.map(s => s.floor));

    // 更新 metrics
    if (metrics) {
        metrics.l0.atomsMatched = scored.length;
        metrics.l0.floorsHit = floors.size;
        metrics.l0.topAtoms = scored.slice(0, 5).map(s => ({
            floor: s.floor,
            semantic: s.atom?.semantic?.slice(0, 50),
            similarity: Math.round(s.similarity * 1000) / 1000,
        }));
    }

    return { atoms: scored, floors };
}

// ═══════════════════════════════════════════════════════════════════════════
// L3 拉取：L0 楼层 → Chunks（带 Rerank 精排）
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 按楼层稀疏去重
 * 每楼层最多保留 limit 个 chunk，优先保留分数高的
 * @param {Array} chunks - chunk 列表（假设已按分数排序）
 * @param {number} limit - 每楼层上限
 * @returns {Array} 去重后的 chunks
 */
function sparseByFloor(chunks, limit = 3) {
    const byFloor = new Map();

    for (const c of chunks) {
        const arr = byFloor.get(c.floor) || [];
        if (arr.length < limit) {
            arr.push(c);
            byFloor.set(c.floor, arr);
        }
    }

    const result = [];
    const seen = new Set();

    for (const c of chunks) {
        if (!seen.has(c.chunkId)) {
            const arr = byFloor.get(c.floor);
            if (arr?.includes(c)) {
                result.push(c);
                seen.add(c.chunkId);
            }
        }
    }

    return result;
}

/**
 * 统计 chunks 的类型构成
 * @param {Array} chunks - chunk 列表
 * @returns {object} {l0Virtual, l1Real}
 */
function countChunksByType(chunks) {
    let l0Virtual = 0;
    let l1Real = 0;

    for (const c of chunks || []) {
        if (c.isL0) {
            l0Virtual++;
        } else {
            l1Real++;
        }
    }

    return { l0Virtual, l1Real };
}

/**
 * 从 L0 命中楼层拉取 chunks，并用 Reranker 精排
 * @param {Set} l0Floors - L0 命中的楼层
 * @param {Array} l0Atoms - L0 atoms（用于构建虚拟 chunks）
 * @param {string} queryText - 查询文本（用于 rerank）
 * @param {object} metrics - 指标对象
 * @returns {Promise<Array>} chunks 列表
 */
async function getChunksFromL0Floors(l0Floors, l0Atoms, queryText, metrics) {
    const { chatId } = getContext();
    if (!chatId || !l0Floors.size) {
        return [];
    }

    const floorArray = Array.from(l0Floors);

    // 从 DB 拉取 chunks
    let dbChunks = [];
    try {
        dbChunks = await getChunksByFloors(chatId, floorArray);
    } catch (e) {
        xbLog.warn(MODULE_ID, '从 DB 拉取 chunks 失败', e);
    }

    // 构建 L0 虚拟 chunks
    const l0VirtualChunks = (l0Atoms || []).map(a => ({
        chunkId: `state-${a.atomId}`,
        floor: a.floor,
        chunkIdx: -1,
        speaker: '📌',
        isUser: false,
        text: a.atom?.semantic || '',
        similarity: a.similarity,
        isL0: true,
        _atom: a.atom,
    }));

    // 合并所有 chunks
    const allChunks = [...l0VirtualChunks, ...dbChunks.map(c => ({
        ...c,
        isL0: false,
        similarity: 0.5,
    }))];

    // ★ 更新 metrics - 候选规模（rerank 前）
    if (metrics) {
        metrics.l3.floorsFromL0 = floorArray.length;
        metrics.l3.chunksInRange = allChunks.length;
        metrics.l3.chunksInRangeByType = {
            l0Virtual: l0VirtualChunks.length,
            l1Real: dbChunks.length,
        };
    }

    // 如果数量不超限，直接按楼层去重返回
    if (allChunks.length <= CONFIG.L3_MAX_TOTAL_CHUNKS) {
        allChunks.sort((a, b) => (b.similarity || 0) - (a.similarity || 0));

        const selected = sparseByFloor(allChunks, CONFIG.L3_MAX_CHUNKS_PER_FLOOR);

        // ★ 更新 metrics - 最终注入规模
        if (metrics) {
            metrics.l3.rerankApplied = false;
            metrics.l3.chunksSelected = selected.length;
            metrics.l3.chunksSelectedByType = countChunksByType(selected);
        }

        return selected;
    }

    // ★ Reranker 精排
    const T_Rerank_Start = performance.now();

    const reranked = await rerankChunks(queryText, allChunks, {
        topN: CONFIG.RERANK_TOP_N,
        minScore: CONFIG.RERANK_MIN_SCORE,
    });

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

    // 按楼层稀疏去重
    const selected = sparseByFloor(reranked, CONFIG.L3_MAX_CHUNKS_PER_FLOOR);

    // ★ 更新 metrics
    if (metrics) {
        metrics.l3.rerankApplied = true;
        metrics.l3.beforeRerank = allChunks.length;
        metrics.l3.afterRerank = reranked.length;
        metrics.l3.chunksSelected = selected.length;
        metrics.l3.chunksSelectedByType = countChunksByType(selected);
        metrics.l3.rerankTime = rerankTime;
        metrics.timing.l3Rerank = rerankTime;

        // rerank 分数分布（基于 selected）
        const scores = selected.map(c => c._rerankScore || 0).filter(s => s > 0);
        if (scores.length > 0) {
            scores.sort((a, b) => a - b);
            metrics.l3.rerankScoreDistribution = {
                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)),
            };
        }
    }

    xbLog.info(MODULE_ID, `L3 Rerank: ${allChunks.length} → ${reranked.length} → ${selected.length} (${rerankTime}ms)`);

    return selected;
}

// ═══════════════════════════════════════════════════════════════════════════
// L2 检索：Query → Events（独立）
// ═══════════════════════════════════════════════════════════════════════════

/**
 * L2 事件向量检索
 * @param {Array} queryVector - 查询向量
 * @param {Array} allEvents - 所有事件
 * @param {object} vectorConfig - 向量配置
 * @param {Array} focusEntities - 焦点实体（用于实体过滤）
 * @param {object} metrics - 指标对象
 * @returns {Promise<Array>} 事件列表
 */
async function searchL2Events(queryVector, allEvents, vectorConfig, focusEntities, metrics) {
    const { chatId } = getContext();
    if (!chatId || !queryVector?.length || !allEvents?.length) {
        return [];
    }

    // 检查 fingerprint
    const meta = await getMeta(chatId);
    const fp = getEngineFingerprint(vectorConfig);
    if (meta.fingerprint && meta.fingerprint !== fp) {
        xbLog.warn(MODULE_ID, 'L2 fingerprint 不匹配');
        return [];
    }

    // 获取事件向量
    const eventVectors = await getAllEventVectors(chatId);
    const vectorMap = new Map(eventVectors.map(v => [v.eventId, v.vector]));

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

    // 实体匹配集合
    const focusSet = new Set((focusEntities || []).map(normalize));

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

        // 实体命中检查
        const participants = (event.participants || []).map(p => normalize(p));
        const hasEntityMatch = participants.some(p => focusSet.has(p));

        // 实体匹配加权
        const bonus = hasEntityMatch ? 0.05 : 0;

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

    // 更新 metrics
    if (metrics) {
        metrics.l2.eventsInStore = allEvents.length;
    }

    // 阈值过滤
    let candidates = scored
        .filter(s => s.similarity >= CONFIG.L2_MIN_SIMILARITY)
        .sort((a, b) => b.similarity - a.similarity)
        .slice(0, CONFIG.L2_CANDIDATE_MAX);

    if (metrics) {
        metrics.l2.eventsConsidered = candidates.length;
    }

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

        candidates = candidates.filter(c => {
            // 高相似度绕过
            if (c.similarity >= 0.85) return true;
            // 有实体匹配的保留
            return c._hasEntityMatch;
        });

        if (metrics) {
            metrics.l2.entityFilterStats = {
                focusEntities: focusEntities || [],
                before: beforeFilter,
                after: candidates.length,
                filtered: beforeFilter - candidates.length,
            };
        }
    }

    // MMR 去重
    const selected = mmrSelect(
        candidates,
        CONFIG.L2_SELECT_MAX,
        CONFIG.L2_MMR_LAMBDA,
        c => c.vector,
        c => c.similarity
    );

    // 统计召回类型
    let directCount = 0;
    let contextCount = 0;

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

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

    // 更新 metrics
    if (metrics) {
        metrics.l2.eventsSelected = results.length;
        metrics.l2.byRecallType = { direct: directCount, context: contextCount, causal: 0 };
        metrics.l2.similarityDistribution = calcSimilarityStats(results.map(r => r.similarity));
    }

    return results;
}

// ═══════════════════════════════════════════════════════════════════════════
// 因果链追溯
// ═══════════════════════════════════════════════════════════════════════════

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

/**
 * 追溯因果祖先
 * @param {Array} recalledEvents - 召回的事件
 * @param {Map} eventIndex - 事件索引
 * @param {number} maxDepth - 最大深度
 * @returns {object} {results, maxDepth}
 */
function traceCausalAncestors(recalledEvents, 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 recalledEvents || []) {
        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 };
}

// ═══════════════════════════════════════════════════════════════════════════
// 辅助函数
// ═══════════════════════════════════════════════════════════════════════════

/**
 * 获取最近的消息
 * @param {Array} chat - 聊天数组
 * @param {number} count - 消息数量
 * @param {boolean} excludeLastAi - 是否排除最后一条 AI 消息
 * @returns {Array} 消息列表
 */
function getLastMessages(chat, count = 4, excludeLastAi = false) {
    if (!chat?.length) return [];

    let messages = [...chat];

    // 排除最后一条 AI 消息（swipe/regenerate 场景）
    if (excludeLastAi && messages.length > 0 && !messages[messages.length - 1]?.is_user) {
        messages = messages.slice(0, -1);
    }

    return messages.slice(-count);
}

/**
 * 构建查询文本（降级用）
 * @param {Array} chat - 聊天数组
 * @param {number} count - 消息数量
 * @param {boolean} excludeLastAi - 是否排除最后一条 AI 消息
 * @returns {string} 查询文本
 */
export function buildQueryText(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).map(m => {
        const text = cleanForRecall(m.mes);
        const speaker = m.name || (m.is_user ? '用户' : '角色');
        return `${speaker}: ${text.slice(0, 500)}`;
    }).filter(Boolean).join('\n');
}

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

/**
 * 记忆召回主函数
 * @param {string} queryText - 查询文本（降级用）
 * @param {Array} allEvents - 所有事件
 * @param {object} vectorConfig - 向量配置
 * @param {object} options - 选项
 * @returns {Promise<object>} 召回结果
 */
export async function recallMemory(queryText, allEvents, vectorConfig, options = {}) {
    const T0 = performance.now();
    const { chat, name1 } = getContext();
    const { pendingUserMessage = null, excludeLastAi = false } = options;

    const metrics = createMetrics();

    if (!allEvents?.length) {
        metrics.l0.needRecall = false;
        return { events: [], chunks: [], causalEvents: [], focusEntities: [], elapsed: 0, logText: 'No events.', metrics };
    }

    // ═══════════════════════════════════════════════════════════════════════
    // Step 1: Query Expansion
    // ═══════════════════════════════════════════════════════════════════════

    const T_QE_Start = performance.now();

    // 获取最近对话
    const lastMessages = getLastMessages(chat, 4, excludeLastAi);

    let expansion = { focus: [], queries: [] };
    try {
        expansion = await expandQueryCached(lastMessages, {
            pendingUserMessage,
            timeout: CONFIG.QUERY_EXPANSION_TIMEOUT,
        });
        xbLog.info(MODULE_ID, `Query Expansion: focus=[${expansion.focus.join(',')}] queries=${expansion.queries.length}`);
    } catch (e) {
        xbLog.warn(MODULE_ID, 'Query Expansion 失败，降级使用原始文本', e);
    }

    // 构建检索文本
    const searchText = buildSearchText(expansion);
    const finalSearchText = searchText || queryText || lastMessages.map(m => cleanForRecall(m.mes || '').slice(0, 200)).join(' ');

    // focusEntities（移除用户名）
    const focusEntities = removeUserNameFromFocus(expansion.focus, name1);

    // 更新 L0 metrics
    metrics.l0.needRecall = true;
    metrics.l0.focusEntities = focusEntities;
    metrics.l0.queries = expansion.queries || [];
    metrics.l0.queryExpansionTime = Math.round(performance.now() - T_QE_Start);
    metrics.timing.queryExpansion = metrics.l0.queryExpansionTime;

    // ═══════════════════════════════════════════════════════════════════════
    // Step 2: 向量化查询
    // ═══════════════════════════════════════════════════════════════════════

    let queryVector;
    try {
        const [vec] = await embed([finalSearchText], vectorConfig, { timeout: 10000 });
        queryVector = vec;
    } catch (e) {
        xbLog.error(MODULE_ID, '向量化失败', e);
        metrics.timing.total = Math.round(performance.now() - T0);
        return { events: [], chunks: [], causalEvents: [], focusEntities, elapsed: metrics.timing.total, logText: 'Embedding failed.', metrics };
    }

    if (!queryVector?.length) {
        metrics.timing.total = Math.round(performance.now() - T0);
        return { events: [], chunks: [], causalEvents: [], focusEntities, elapsed: metrics.timing.total, logText: 'Empty query vector.', metrics };
    }

    // ═══════════════════════════════════════════════════════════════════════
    // Step 3: L0 检索 → L3 拉取（并行准备）
    // ═══════════════════════════════════════════════════════════════════════

    const T_L0_Start = performance.now();

    const { atoms: l0Atoms, floors: l0Floors } = await searchL0(queryVector, vectorConfig, metrics);

    metrics.timing.l0Search = Math.round(performance.now() - T_L0_Start);

    // ═══════════════════════════════════════════════════════════════════════
    // Step 4: L3 从 L0 楼层拉取（带 Rerank）
    // ═══════════════════════════════════════════════════════════════════════

    const T_L3_Start = performance.now();

    // 构建 rerank 用的查询文本
    const rerankQuery = buildRerankQuery(expansion, lastMessages, pendingUserMessage);

    const chunks = await getChunksFromL0Floors(l0Floors, l0Atoms, rerankQuery, metrics);

    metrics.timing.l3Retrieval = Math.round(performance.now() - T_L3_Start);

    // ═══════════════════════════════════════════════════════════════════════
    // Step 5: L2 独立检索
    // ═══════════════════════════════════════════════════════════════════════

    const T_L2_Start = performance.now();

    const eventResults = await searchL2Events(queryVector, allEvents, vectorConfig, focusEntities, metrics);

    metrics.timing.l2Retrieval = Math.round(performance.now() - T_L2_Start);

    // ═══════════════════════════════════════════════════════════════════════
    // Step 6: 因果链追溯
    // ═══════════════════════════════════════════════════════════════════════

    const eventIndex = buildEventIndex(allEvents);
    const { results: causalMap, maxDepth: causalMaxDepth } = traceCausalAncestors(eventResults, eventIndex);

    const recalledIdSet = new Set(eventResults.map(x => x?.event?.id).filter(Boolean));
    const causalEvents = 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,
        }));

    // 更新因果链 metrics
    if (metrics.l2.byRecallType) {
        metrics.l2.byRecallType.causal = causalEvents.length;
    }
    metrics.l2.causalChainDepth = causalMaxDepth;
    metrics.l2.causalEventsCount = causalEvents.length;

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

    metrics.timing.total = Math.round(performance.now() - T0);

    // 实体信息
    metrics.l2.entityNames = focusEntities;
    metrics.l2.entitiesLoaded = focusEntities.length;

    // 日志
    console.group('%c[Recall v3]', 'color: #7c3aed; font-weight: bold');
    console.log(`Elapsed: ${metrics.timing.total}ms`);
    console.log(`Query Expansion: focus=[${expansion.focus.join(', ')}]`);
    console.log(`L0: ${l0Atoms.length} atoms → ${l0Floors.size} floors`);
    console.log(`L3: ${chunks.length} chunks (L0=${metrics.l3.chunksSelectedByType?.l0Virtual || 0}, DB=${metrics.l3.chunksSelectedByType?.l1Real || 0})`);
    if (metrics.l3.rerankApplied) {
        console.log(`L3 Rerank: ${metrics.l3.beforeRerank} → ${metrics.l3.afterRerank} (${metrics.l3.rerankTime}ms)`);
    }
    console.log(`L2: ${eventResults.length} events, ${causalEvents.length} causal`);
    console.groupEnd();

    return {
        events: eventResults,
        causalEvents,
        chunks,
        expansion,
        focusEntities,
        elapsed: metrics.timing.total,
        metrics,
    };
}