825 lines
31 KiB
JavaScript
825 lines
31 KiB
JavaScript
// ═══════════════════════════════════════════════════════════════════════════
|
||
// Story Summary - Recall Engine (v5 - 统一命名)
|
||
//
|
||
// 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
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
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 { 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,
|
||
|
||
// Anchor (L0 StateAtoms) 配置
|
||
ANCHOR_MIN_SIMILARITY: 0.58,
|
||
|
||
// Evidence (L1 Chunks) 粗筛配置
|
||
EVIDENCE_COARSE_MAX: 100,
|
||
|
||
// Event (L2 Events) 配置
|
||
EVENT_CANDIDATE_MAX: 100,
|
||
EVENT_SELECT_MAX: 50,
|
||
EVENT_MIN_SIMILARITY: 0.55,
|
||
EVENT_MMR_LAMBDA: 0.72,
|
||
|
||
// Rerank 配置
|
||
RERANK_THRESHOLD: 80,
|
||
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 - 向量A
|
||
* @param {number[]} b - 向量B
|
||
* @returns {number} 相似度 [0, 1]
|
||
*/
|
||
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（用于实体匹配）
|
||
* @param {string} s - 输入字符串
|
||
* @returns {string} 标准化后的字符串
|
||
*/
|
||
function normalize(s) {
|
||
return String(s || '')
|
||
.normalize('NFKC')
|
||
.replace(/[\u200B-\u200D\uFEFF]/g, '')
|
||
.trim()
|
||
.toLowerCase();
|
||
}
|
||
|
||
/**
|
||
* 清理文本用于召回
|
||
* @param {string} text - 原始文本
|
||
* @returns {string} 清理后的文本
|
||
*/
|
||
function cleanForRecall(text) {
|
||
return filterText(text).replace(/\[tts:[^\]]*\]/gi, '').trim();
|
||
}
|
||
|
||
/**
|
||
* 从 focus entities 中移除用户名
|
||
* @param {string[]} focusEntities - 焦点实体列表
|
||
* @param {string} userName - 用户名
|
||
* @returns {string[]} 过滤后的实体列表
|
||
*/
|
||
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 rerank 查询文本
|
||
* @param {object} expansion - query expansion 结果
|
||
* @param {object[]} lastMessages - 最近消息
|
||
* @param {string} pendingUserMessage - 待发送的用户消息
|
||
* @returns {string} 查询文本
|
||
*/
|
||
function buildRerankQuery(expansion, lastMessages, pendingUserMessage) {
|
||
const parts = [];
|
||
|
||
if (expansion?.focus?.length) {
|
||
parts.push(expansion.focus.join(' '));
|
||
}
|
||
|
||
if (expansion?.queries?.length) {
|
||
parts.push(...expansion.queries.slice(0, 3));
|
||
}
|
||
|
||
const recentTexts = (lastMessages || [])
|
||
.slice(-2)
|
||
.map(m => cleanForRecall(m.mes || '').slice(0, 150))
|
||
.filter(Boolean);
|
||
|
||
if (recentTexts.length) {
|
||
parts.push(...recentTexts);
|
||
}
|
||
|
||
if (pendingUserMessage) {
|
||
parts.push(cleanForRecall(pendingUserMessage).slice(0, 200));
|
||
}
|
||
|
||
return parts.filter(Boolean).join('\n').slice(0, 1500);
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// MMR 选择算法
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* Maximal Marginal Relevance 选择
|
||
* @param {object[]} candidates - 候选项
|
||
* @param {number} k - 选择数量
|
||
* @param {number} lambda - 相关性/多样性权衡参数
|
||
* @param {Function} getVector - 获取向量的函数
|
||
* @param {Function} getScore - 获取分数的函数
|
||
* @returns {object[]} 选中的候选项
|
||
*/
|
||
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;
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// [Anchors] L0 StateAtoms 检索
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* 检索语义锚点（L0 StateAtoms）
|
||
* @param {number[]} queryVector - 查询向量
|
||
* @param {object} vectorConfig - 向量配置
|
||
* @param {object} metrics - 指标对象
|
||
* @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]));
|
||
|
||
// 按阈值过滤，不设硬上限
|
||
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 };
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// [Evidence] L1 Chunks 拉取 + 粗筛 + Rerank
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* 统计 evidence 类型构成
|
||
* @param {object[]} chunks - chunk 列表
|
||
* @returns {{anchorVirtual: number, chunkReal: number}}
|
||
*/
|
||
function countEvidenceByType(chunks) {
|
||
let anchorVirtual = 0;
|
||
let chunkReal = 0;
|
||
|
||
for (const c of chunks || []) {
|
||
if (c.isAnchorVirtual) {
|
||
anchorVirtual++;
|
||
} else {
|
||
chunkReal++;
|
||
}
|
||
}
|
||
|
||
return { anchorVirtual, chunkReal };
|
||
}
|
||
|
||
/**
|
||
* 根据锚点命中楼层拉取证据（L1 Chunks）
|
||
* @param {Set<number>} anchorFloors - 锚点命中的楼层
|
||
* @param {object[]} anchorHits - 锚点命中结果
|
||
* @param {number[]} queryVector - 查询向量
|
||
* @param {string} queryText - rerank 查询文本
|
||
* @param {object} metrics - 指标对象
|
||
* @returns {Promise<object[]>} 证据 chunks
|
||
*/
|
||
async function pullEvidenceByFloors(anchorFloors, anchorHits, queryVector, queryText, metrics) {
|
||
const { chatId } = getContext();
|
||
if (!chatId || !anchorFloors.size) {
|
||
return [];
|
||
}
|
||
|
||
const floorArray = Array.from(anchorFloors);
|
||
|
||
// 1. 构建锚点虚拟 chunks（来自 L0 StateAtoms）
|
||
const anchorVirtualChunks = (anchorHits || []).map(a => ({
|
||
chunkId: `anchor-${a.atomId}`,
|
||
floor: a.floor,
|
||
chunkIdx: -1,
|
||
speaker: '📌',
|
||
isUser: false,
|
||
text: a.atom?.semantic || '',
|
||
similarity: a.similarity,
|
||
isAnchorVirtual: true,
|
||
_atom: a.atom,
|
||
}));
|
||
|
||
// 2. 拉取真实 chunks（来自 L1）
|
||
let dbChunks = [];
|
||
try {
|
||
dbChunks = await getChunksByFloors(chatId, floorArray);
|
||
} catch (e) {
|
||
xbLog.warn(MODULE_ID, '从 DB 拉取 chunks 失败', e);
|
||
}
|
||
|
||
// 3. L1 向量粗筛
|
||
let coarseFiltered = [];
|
||
if (dbChunks.length > 0 && queryVector?.length) {
|
||
const chunkIds = dbChunks.map(c => c.chunkId);
|
||
let chunkVectors = [];
|
||
try {
|
||
chunkVectors = await getChunkVectorsByIds(chatId, chunkIds);
|
||
} catch (e) {
|
||
xbLog.warn(MODULE_ID, 'L1 向量获取失败', e);
|
||
}
|
||
|
||
const vectorMap = new Map(chunkVectors.map(v => [v.chunkId, v.vector]));
|
||
|
||
coarseFiltered = dbChunks
|
||
.map(c => {
|
||
const vec = vectorMap.get(c.chunkId);
|
||
if (!vec?.length) return null;
|
||
|
||
return {
|
||
...c,
|
||
isAnchorVirtual: false,
|
||
similarity: cosineSimilarity(queryVector, vec),
|
||
};
|
||
})
|
||
.filter(Boolean)
|
||
.sort((a, b) => b.similarity - a.similarity)
|
||
.slice(0, CONFIG.EVIDENCE_COARSE_MAX);
|
||
}
|
||
|
||
// 4. 合并
|
||
const allEvidence = [...anchorVirtualChunks, ...coarseFiltered];
|
||
|
||
// 更新 metrics
|
||
if (metrics) {
|
||
metrics.evidence.floorsFromAnchors = floorArray.length;
|
||
metrics.evidence.chunkTotal = dbChunks.length;
|
||
metrics.evidence.chunkAfterCoarse = coarseFiltered.length;
|
||
metrics.evidence.merged = allEvidence.length;
|
||
metrics.evidence.mergedByType = countEvidenceByType(allEvidence);
|
||
}
|
||
|
||
// 5. 是否需要 Rerank
|
||
if (allEvidence.length <= CONFIG.RERANK_THRESHOLD) {
|
||
if (metrics) {
|
||
metrics.evidence.rerankApplied = false;
|
||
metrics.evidence.selected = allEvidence.length;
|
||
metrics.evidence.selectedByType = countEvidenceByType(allEvidence);
|
||
}
|
||
return allEvidence;
|
||
}
|
||
|
||
// 6. Rerank 精排
|
||
const T_Rerank_Start = performance.now();
|
||
|
||
const reranked = await rerankChunks(queryText, allEvidence, {
|
||
topN: CONFIG.RERANK_TOP_N,
|
||
minScore: CONFIG.RERANK_MIN_SCORE,
|
||
});
|
||
|
||
const rerankTime = Math.round(performance.now() - T_Rerank_Start);
|
||
|
||
if (metrics) {
|
||
metrics.evidence.rerankApplied = true;
|
||
metrics.evidence.beforeRerank = allEvidence.length;
|
||
metrics.evidence.afterRerank = reranked.length;
|
||
metrics.evidence.selected = reranked.length;
|
||
metrics.evidence.selectedByType = countEvidenceByType(reranked);
|
||
metrics.evidence.rerankTime = rerankTime;
|
||
metrics.timing.evidenceRerank = rerankTime;
|
||
|
||
const scores = reranked.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)),
|
||
};
|
||
}
|
||
}
|
||
|
||
xbLog.info(MODULE_ID, `Evidence: ${dbChunks.length} L1 → ${coarseFiltered.length} coarse → ${reranked.length} rerank (${rerankTime}ms)`);
|
||
|
||
return reranked;
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// [Events] L2 Events 检索
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* 检索事件（L2 Events）
|
||
* @param {number[]} queryVector - 查询向量
|
||
* @param {object[]} allEvents - 所有事件
|
||
* @param {object} vectorConfig - 向量配置
|
||
* @param {string[]} focusEntities - 焦点实体
|
||
* @param {object} metrics - 指标对象
|
||
* @returns {Promise<object[]>} 事件命中结果
|
||
*/
|
||
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));
|
||
|
||
const bonus = hasEntityMatch ? 0.05 : 0;
|
||
|
||
return {
|
||
_id: event.id,
|
||
event,
|
||
similarity: baseSim + bonus,
|
||
_baseSim: 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,
|
||
_baseSim: s._baseSim,
|
||
};
|
||
});
|
||
|
||
if (metrics) {
|
||
metrics.event.selected = results.length;
|
||
metrics.event.byRecallType = { direct: directCount, related: relatedCount, causal: 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 - 所有事件
|
||
* @returns {Map<string, object>} 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;
|
||
}
|
||
|
||
/**
|
||
* 追溯因果链
|
||
* @param {object[]} eventHits - 事件命中结果
|
||
* @param {Map<string, object>} eventIndex - 事件索引
|
||
* @param {number} maxDepth - 最大深度
|
||
* @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 };
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// 辅助函数
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* 获取最近消息
|
||
* @param {object[]} chat - 聊天记录
|
||
* @param {number} count - 消息数量
|
||
* @param {boolean} excludeLastAi - 是否排除最后的 AI 消息
|
||
* @returns {object[]} 最近消息
|
||
*/
|
||
function getLastMessages(chat, count = 4, 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);
|
||
}
|
||
|
||
/**
|
||
* 构建查询文本
|
||
* @param {object[]} chat - 聊天记录
|
||
* @param {number} count - 消息数量
|
||
* @param {boolean} excludeLastAi - 是否排除最后的 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}: ${text.slice(0, 500)}`;
|
||
}).filter(Boolean).join('\n');
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
// 主函数
|
||
// ═══════════════════════════════════════════════════════════════════════════
|
||
|
||
/**
|
||
* 执行记忆召回
|
||
* @param {string} queryText - 查询文本
|
||
* @param {object[]} allEvents - 所有事件（L2）
|
||
* @param {object} vectorConfig - 向量配置
|
||
* @param {object} options - 选项
|
||
* @returns {Promise<object>} 召回结果
|
||
*/
|
||
export async function recallMemory(queryText, allEvents, vectorConfig, options = {}) {
|
||
const T0 = performance.now();
|
||
const { chat, name1 } = getContext();
|
||
const { pendingUserMessage = null, excludeLastAi = false } = options;
|
||
|
||
const metrics = createMetrics();
|
||
|
||
if (!allEvents?.length) {
|
||
metrics.anchor.needRecall = false;
|
||
return {
|
||
events: [],
|
||
evidenceChunks: [],
|
||
causalChain: [],
|
||
focusEntities: [],
|
||
elapsed: 0,
|
||
logText: 'No events.',
|
||
metrics,
|
||
};
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
// Step 1: Query Expansion
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
|
||
const T_QE_Start = performance.now();
|
||
|
||
const lastMessages = getLastMessages(chat, 4, excludeLastAi);
|
||
|
||
let expansion = { focus: [], queries: [] };
|
||
try {
|
||
expansion = await expandQueryCached(lastMessages, {
|
||
pendingUserMessage,
|
||
timeout: CONFIG.QUERY_EXPANSION_TIMEOUT,
|
||
});
|
||
xbLog.info(MODULE_ID, `Query Expansion: focus=[${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);
|
||
}
|
||
|
||
const searchText = buildSearchText(expansion);
|
||
const finalSearchText = searchText || queryText || lastMessages.map(m => cleanForRecall(m.mes || '').slice(0, 200)).join(' ');
|
||
|
||
const focusEntities = removeUserNameFromFocus(expansion.focus, name1);
|
||
|
||
metrics.anchor.needRecall = true;
|
||
metrics.anchor.focusEntities = focusEntities;
|
||
metrics.anchor.queries = expansion.queries || [];
|
||
metrics.anchor.queryExpansionTime = Math.round(performance.now() - T_QE_Start);
|
||
metrics.timing.queryExpansion = metrics.anchor.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: [],
|
||
evidenceChunks: [],
|
||
causalChain: [],
|
||
focusEntities,
|
||
elapsed: metrics.timing.total,
|
||
logText: 'Embedding failed.',
|
||
metrics,
|
||
};
|
||
}
|
||
|
||
if (!queryVector?.length) {
|
||
metrics.timing.total = Math.round(performance.now() - T0);
|
||
return {
|
||
events: [],
|
||
evidenceChunks: [],
|
||
causalChain: [],
|
||
focusEntities,
|
||
elapsed: metrics.timing.total,
|
||
logText: 'Empty query vector.',
|
||
metrics,
|
||
};
|
||
}
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
// Step 3: Anchor (L0) 检索
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
|
||
const T_Anchor_Start = performance.now();
|
||
|
||
const { hits: anchorHits, floors: anchorFloors } = await recallAnchors(queryVector, vectorConfig, metrics);
|
||
|
||
metrics.timing.anchorSearch = Math.round(performance.now() - T_Anchor_Start);
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
// Step 4: Evidence (L1) 拉取 + 粗筛 + Rerank
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
|
||
const T_Evidence_Start = performance.now();
|
||
|
||
const rerankQuery = buildRerankQuery(expansion, lastMessages, pendingUserMessage);
|
||
const evidenceChunks = await pullEvidenceByFloors(anchorFloors, anchorHits, queryVector, rerankQuery, metrics);
|
||
|
||
metrics.timing.evidenceRetrieval = Math.round(performance.now() - T_Evidence_Start);
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
// Step 5: Event (L2) 独立检索
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
|
||
const T_Event_Start = performance.now();
|
||
|
||
const eventHits = await recallEvents(queryVector, allEvents, vectorConfig, focusEntities, metrics);
|
||
|
||
metrics.timing.eventRetrieval = Math.round(performance.now() - T_Event_Start);
|
||
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
// Step 6: 因果链追溯
|
||
// ═══════════════════════════════════════════════════════════════════════
|
||
|
||
const eventIndex = buildEventIndex(allEvents);
|
||
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 = focusEntities;
|
||
metrics.event.entitiesUsed = focusEntities.length;
|
||
|
||
console.group('%c[Recall v5]', 'color: #7c3aed; font-weight: bold');
|
||
console.log(`Elapsed: ${metrics.timing.total}ms`);
|
||
console.log(`Query Expansion: focus=[${expansion.focus.join(', ')}]`);
|
||
console.log(`Anchors: ${anchorHits.length} hits → ${anchorFloors.size} floors`);
|
||
console.log(`Evidence: ${metrics.evidence.chunkTotal || 0} L1 → ${metrics.evidence.chunkAfterCoarse || 0} coarse → ${evidenceChunks.length} final`);
|
||
if (metrics.evidence.rerankApplied) {
|
||
console.log(`Evidence Rerank: ${metrics.evidence.beforeRerank} → ${metrics.evidence.afterRerank} (${metrics.evidence.rerankTime}ms)`);
|
||
}
|
||
console.log(`Events: ${eventHits.length} hits, ${causalChain.length} causal`);
|
||
console.groupEnd();
|
||
|
||
return {
|
||
events: eventHits,
|
||
causalChain,
|
||
evidenceChunks,
|
||
expansion,
|
||
focusEntities,
|
||
elapsed: metrics.timing.total,
|
||
metrics,
|
||
};
|
||
}
|