From f5a5608a12afbf05d8fd160a94962644b7dcea32 Mon Sep 17 00:00:00 2001 From: Youzini-afk <13153778771cx@gmail.com> Date: Wed, 25 Mar 2026 14:41:44 +0800 Subject: [PATCH] =?UTF-8?q?perf:=20=E8=AE=B0=E5=BF=86=E6=95=B4=E5=90=88?= =?UTF-8?q?=E6=89=B9=E9=87=8F=E5=8C=96=20=E2=80=94=20=E5=B0=86=20N=20?= =?UTF-8?q?=E6=AC=A1=20LLM=20=E5=8E=8B=E7=BC=A9=E4=B8=BA=201=20=E6=AC=A1?= =?UTF-8?q?=E6=89=B9=E9=87=8F=E8=B0=83=E7=94=A8?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 重构 consolidator.js 为 4 阶段架构: Phase 0: 收集有效新节点 Phase 1: 批量 Embed (直连 1 次 embedBatch / 后端逐次) Phase 2: 各节点查近邻 (直连本地 cosine / 后端逐次 query) Phase 3: 单次 LLM 批量判定 (统一 prompt 含所有新节点) Phase 4: 逐个处理结果 - 更新 panel.js DEFAULT_PROMPTS.consolidation 为 results 数组格式 - 直连模式: N embed + N LLM -> 1 embedBatch + 1 LLM (-67~75%) - 后端模式: N query + N LLM -> N query + 1 LLM (-33~50%) - 总体效果: 原始 6 LLM -> 2 LLM (2新节点场景) --- consolidator.js | 540 +++++++++++++++++++++++++++++------------------- panel.js | 17 +- 2 files changed, 333 insertions(+), 224 deletions(-) diff --git a/consolidator.js b/consolidator.js index 98e67d0..925c9a1 100644 --- a/consolidator.js +++ b/consolidator.js @@ -1,12 +1,14 @@ -// ST-BME: 统一记忆整合引擎 +// ST-BME: 统一记忆整合引擎（批量化版） // 合并 Mem0 精确对照 + A-MEM 记忆进化为单一阶段 -// 每个新节点只需 1 次 embed + 1 次 LLM 调用 +// 批量 embed + 批量查近邻 + 单次 LLM 调用 +import { embedBatch, searchSimilar } from './embedding.js'; import { addEdge, createEdge, getActiveNodes, getNode } from './graph.js'; import { callLLMForJSON } from './llm.js'; import { buildNodeVectorText, findSimilarNodesByText, + isDirectVectorConfig, validateVectorConfig, } from './vector-index.js'; @@ -27,8 +29,7 @@ function throwIfAborted(signal) { } /** - * 统一记忆整合系统提示词 - * 同时完成 Mem0 冲突判定 + A-MEM 进化分析 + * 统一记忆整合系统提示词（支持批量输出） */ const CONSOLIDATION_SYSTEM_PROMPT = `你是一个记忆整合分析器。当新记忆加入知识图谱时，你需要同时完成两项任务： @@ -45,24 +46,30 @@ const CONSOLIDATION_SYSTEM_PROMPT = `你是一个记忆整合分析器。当新 输出严格 JSON： { - "action": "keep" | "merge" | "skip", - "merge_target_id": "仅 action=merge 时必填：要合并到的旧节点 ID", - "merged_fields": { "仅 action=merge 时可选：合并后的字段更新" }, - "reason": "判定理由（简述）", - "evolution": { - "should_evolve": true/false, - "connections": ["需要建立链接的旧记忆 ID 列表"], - "neighbor_updates": [ - { - "nodeId": "需更新的旧节点 ID", - "newContext": "基于新信息修正后的描述（不需修改则为 null）", - "newTags": ["更新后的分类标签，不需修改则为 null"] + "results": [ + { + "node_id": "新记忆的节点 ID", + "action": "keep" | "merge" | "skip", + "merge_target_id": "仅 action=merge 时必填：要合并到的旧节点 ID", + "merged_fields": { "仅 action=merge 时可选：合并后的字段更新" }, + "reason": "判定理由（简述）", + "evolution": { + "should_evolve": true/false, + "connections": ["需要建立链接的旧记忆 ID 列表"], + "neighbor_updates": [ + { + "nodeId": "需更新的旧节点 ID", + "newContext": "基于新信息修正后的描述（不需修改则为 null）", + "newTags": ["更新后的分类标签，不需修改则为 null"] + } + ] } - ] - } + } + ] } 整合规则： +- 必须对每条新记忆都给出一个 result 条目 - 当 action=skip 时，evolution 可省略或设 should_evolve=false - 当 action=merge 时，evolution 可省略或设 should_evolve=false - 仅当 action=keep 且新信息确实改变了对旧记忆的理解时，才设 should_evolve=true @@ -72,19 +79,23 @@ const CONSOLIDATION_SYSTEM_PROMPT = `你是一个记忆整合分析器。当新 - neighbor_updates 中每条必须有实际意义的修改`; /** - * 统一记忆整合主函数 + * 统一记忆整合主函数（批量化版） * - * 合并了原先的 mem0ConflictCheck（精确对照）和 evolveMemories（进化）， - * 实现"1 次 embed + 1 次 LLM"完成冲突检测 + 进化分析。 + * 4 阶段架构： + * Phase 0: 收集有效新节点 + * Phase 1: 批量 Embed（直连 1 次 embedBatch / 后端逐次） + * Phase 2: 各节点查近邻（直连本地 cosine / 后端逐次 query） + * Phase 3: 单次 LLM 批量判定 + * Phase 4: 逐个处理结果 * * @param {object} params * @param {object} params.graph - 当前图状态 * @param {string[]} params.newNodeIds - 本次新创建的节点 ID 列表 * @param {object} params.embeddingConfig - Embedding API 配置 * @param {object} [params.options] - * @param {number} [params.options.neighborCount=5] - 近邻搜索数量 - * @param {number} [params.options.conflictThreshold=0.85] - 冲突判定阈值（低于此值跳过冲突检测） - * @param {string} [params.customPrompt] - 自定义提示词 + * @param {number} [params.options.neighborCount=5] + * @param {number} [params.options.conflictThreshold=0.85] + * @param {string} [params.customPrompt] * @param {AbortSignal} [params.signal] * @returns {Promise<{merged: number, skipped: number, kept: number, evolved: number, connections: number, updates: number}>} */ @@ -113,218 +124,200 @@ export async function consolidateMemories({ return stats; } + // ══════════════════════════════════════════════ + // Phase 0: 收集有效新节点 + // ══════════════════════════════════════════════ + const newEntries = []; + for (const id of newNodeIds) { + const node = getNode(graph, id); + if (!node || node.archived) continue; + const text = buildNodeVectorText(node); + if (!text) continue; + newEntries.push({ id, node, text }); + } + + if (newEntries.length === 0) return stats; + const activeNodes = getActiveNodes(graph).filter(n => { const text = buildNodeVectorText(n); return typeof text === 'string' && text.length > 0; }); - if (activeNodes.length < 2) return stats; + if (activeNodes.length < 2) { + // 图中节点不够，全部 keep + stats.kept = newEntries.length; + return stats; + } - for (const newId of newNodeIds) { - throwIfAborted(signal); - const newNode = getNode(graph, newId); - if (!newNode || newNode.archived) continue; + throwIfAborted(signal); + console.log(`[ST-BME] 记忆整合开始: ${newEntries.length} 个新节点`); - const queryText = buildNodeVectorText(newNode); - if (!queryText) continue; + // ══════════════════════════════════════════════ + // Phase 1 + 2: 批量 Embed + 查近邻 + // ══════════════════════════════════════════════ + /** @type {Map>} */ + const neighborsMap = new Map(); - // 排除自身的候选池 - const candidates = activeNodes.filter(n => n.id !== newId); - if (candidates.length === 0) { - stats.kept++; - continue; - } + if (isDirectVectorConfig(embeddingConfig)) { + // ── 直连模式: 1 次 embedBatch + N 次本地 cosine ── + const texts = newEntries.map(e => e.text); + let queryVectors; try { - // ── 1次 Embed：查近邻 ── - const neighbors = await findSimilarNodesByText( - graph, - queryText, - embeddingConfig, - neighborCount, - candidates, - signal, - ); + queryVectors = await embedBatch(texts, embeddingConfig, { signal }); + } catch (e) { + if (isAbortError(e)) throw e; + console.warn('[ST-BME] 批量 embed 失败，回退到逐条:', e.message); + queryVectors = null; + } - if (neighbors.length === 0) { - stats.kept++; - continue; + // 构建候选池（含 embedding 的活跃节点） + const candidatePool = activeNodes + .filter(n => Array.isArray(n.embedding) && n.embedding.length > 0) + .map(n => ({ nodeId: n.id, embedding: n.embedding })); + + for (let i = 0; i < newEntries.length; i++) { + throwIfAborted(signal); + const entry = newEntries[i]; + const candidates = candidatePool.filter(c => c.nodeId !== entry.id); + + if (queryVectors?.[i] && candidates.length > 0) { + // 本地 cosine 搜索（0 API 调用） + const neighbors = searchSimilar(queryVectors[i], candidates, neighborCount); + neighborsMap.set(entry.id, neighbors); + } else { + // fallback: 逐条 embed + try { + const neighbors = await findSimilarNodesByText( + graph, entry.text, embeddingConfig, neighborCount, + activeNodes.filter(n => n.id !== entry.id), signal, + ); + neighborsMap.set(entry.id, neighbors); + } catch (e) { + if (isAbortError(e)) throw e; + console.warn(`[ST-BME] 近邻查询失败 (${entry.id}):`, e.message); + neighborsMap.set(entry.id, []); + } } + } + } else { + // ── 后端模式: 逐条 /api/vector/query ── + for (let i = 0; i < newEntries.length; i++) { + throwIfAborted(signal); + const entry = newEntries[i]; + try { + const neighbors = await findSimilarNodesByText( + graph, entry.text, embeddingConfig, neighborCount, + activeNodes.filter(n => n.id !== entry.id), signal, + ); + neighborsMap.set(entry.id, neighbors); + } catch (e) { + if (isAbortError(e)) throw e; + console.warn(`[ST-BME] 近邻查询失败 (${entry.id}):`, e.message); + neighborsMap.set(entry.id, []); + } + } + } - // 构建近邻描述文本 - const neighborsContext = neighbors.map(n => { + // ══════════════════════════════════════════════ + // Phase 3: 单次 LLM 批量判定 + // ══════════════════════════════════════════════ + throwIfAborted(signal); + + const userPromptSections = []; + userPromptSections.push(`本轮共新增 ${newEntries.length} 条记忆，请逐条分析：\n`); + + for (let i = 0; i < newEntries.length; i++) { + const entry = newEntries[i]; + const neighbors = neighborsMap.get(entry.id) || []; + + const newNodeFieldsStr = Object.entries(entry.node.fields) + .map(([k, v]) => `${k}: ${v}`) + .join(', '); + + // 构建近邻描述 + let neighborText; + if (neighbors.length === 0) { + neighborText = ' (无近邻命中)'; + } else { + neighborText = neighbors.map(n => { const node = getNode(graph, n.nodeId); if (!node) return null; const fieldsStr = Object.entries(node.fields) .map(([k, v]) => `${k}: ${v}`) .join(', '); - return `[${node.id}] 类型=${node.type}, ${fieldsStr}, 相似度=${n.score.toFixed(3)}${ - (node.clusters || []).length > 0 ? `, 分类=${node.clusters.join('/')}` : '' - }`; + return ` - [${node.id}] 类型=${node.type}, ${fieldsStr} (相似度=${n.score.toFixed(3)})`; }).filter(Boolean).join('\n'); - - const newNodeFieldsStr = Object.entries(newNode.fields) - .map(([k, v]) => `${k}: ${v}`) - .join(', '); - - // 检查是否有高相似度命中（决定是否启用冲突检测部分的提示） - const hasHighSimilarity = neighbors[0].score > conflictThreshold; - - const userPrompt = [ - '## 新加入的记忆', - `[${newNode.id}] 类型=${newNode.type}, ${newNodeFieldsStr}`, - '', - '## 最近邻的已有记忆', - neighborsContext, - '', - `共 ${neighbors.length} 条近邻记忆。`, - hasHighSimilarity - ? `最高相似度 ${neighbors[0].score.toFixed(3)} 超过阈值 ${conflictThreshold}，请先判断是否冲突/重复，再分析进化关系。` - : '相似度均较低，请重点分析新记忆是否揭示了关于旧记忆的新信息。', - ].join('\n'); - - // ── 1次 LLM：统一判定 ── - const decision = await callLLMForJSON({ - systemPrompt: customPrompt || CONSOLIDATION_SYSTEM_PROMPT, - userPrompt, - maxRetries: 1, - signal, - }); - - if (!decision) { - stats.kept++; - continue; - } - - // ── 处理 action ── - switch (decision.action) { - case 'skip': { - console.log(`[ST-BME] 记忆整合: skip (重复) — ${newId}`); - newNode.archived = true; - stats.skipped++; - break; - } - - case 'merge': { - const targetId = decision.merge_target_id; - const targetNode = targetId ? getNode(graph, targetId) : null; - - if (targetNode && !targetNode.archived) { - console.log(`[ST-BME] 记忆整合: merge ${newId} → ${targetId}`); - - // 合并字段到旧节点 - if (decision.merged_fields && typeof decision.merged_fields === 'object') { - for (const [key, value] of Object.entries(decision.merged_fields)) { - if (value != null && value !== '') { - targetNode.fields[key] = value; - } - } - } else { - // 如果没提供 merged_fields，将新节点的非空字段补充到旧节点 - for (const [key, value] of Object.entries(newNode.fields)) { - if (value != null && value !== '' && !targetNode.fields[key]) { - targetNode.fields[key] = value; - } - } - } - - // 更新旧节点的 seq 为更新的值 - if (Number.isFinite(newNode.seq) && newNode.seq > (targetNode.seq || 0)) { - targetNode.seq = newNode.seq; - } - - // 标记旧节点需要 re-embed - targetNode.embedding = null; - - // 归档新节点 - newNode.archived = true; - stats.merged++; - } else { - // merge target 无效，回退为 keep - console.warn(`[ST-BME] 记忆整合: merge target ${targetId} 不存在，回退为 keep`); - stats.kept++; - } - break; - } - - case 'keep': - default: { - stats.kept++; - break; - } - } - - // ── 处理 evolution（仅 keep 时有意义，但也容错处理其它 action） ── - const evolution = decision.evolution; - if (evolution?.should_evolve && !newNode.archived) { - stats.evolved++; - console.log(`[ST-BME] 记忆整合/进化触发: ${decision.reason || '(无理由)'}`); - - // 建立关联边 - if (Array.isArray(evolution.connections)) { - for (const targetId of evolution.connections) { - if (!getNode(graph, targetId)) continue; - const edge = createEdge({ - fromId: newId, - toId: targetId, - relation: 'related', - strength: 0.7, - }); - if (addEdge(graph, edge)) { - stats.connections++; - } - } - } - - // 反向更新旧节点 - if (Array.isArray(evolution.neighbor_updates)) { - for (const update of evolution.neighbor_updates) { - if (!update.nodeId) continue; - const oldNode = getNode(graph, update.nodeId); - if (!oldNode || oldNode.archived) continue; - - let changed = false; - - // 更新 context/state 字段 - if (update.newContext && typeof update.newContext === 'string') { - if (oldNode.fields.state !== undefined) { - oldNode.fields.state = update.newContext; - changed = true; - } else if (oldNode.fields.summary !== undefined) { - oldNode.fields.summary = update.newContext; - changed = true; - } else if (oldNode.fields.core_note !== undefined) { - oldNode.fields.core_note = update.newContext; - changed = true; - } - } - - // 更新分类标签 - if (update.newTags && Array.isArray(update.newTags)) { - oldNode.clusters = update.newTags; - changed = true; - } - - if (changed) { - oldNode.embedding = null; - if (!oldNode._evolutionHistory) oldNode._evolutionHistory = []; - oldNode._evolutionHistory.push({ - triggeredBy: newId, - timestamp: Date.now(), - reason: decision.reason || '', - }); - stats.updates++; - } - } - } - } - } catch (e) { - if (isAbortError(e)) throw e; - console.error(`[ST-BME] 记忆整合失败 (${newId}):`, e); - stats.kept++; } + + // 检查高相似度 + const hasHighSimilarity = neighbors.length > 0 && neighbors[0].score > conflictThreshold; + const hint = hasHighSimilarity + ? ` ⚠ 最高相似度 ${neighbors[0].score.toFixed(3)} 超过阈值 ${conflictThreshold}` + : ''; + + userPromptSections.push([ + `### 新记忆 #${i + 1}`, + `[${entry.id}] 类型=${entry.node.type}, ${newNodeFieldsStr}`, + '近邻记忆:', + neighborText, + hint, + ].filter(Boolean).join('\n')); } + const userPrompt = userPromptSections.join('\n\n'); + + let decision; + try { + decision = await callLLMForJSON({ + systemPrompt: customPrompt || CONSOLIDATION_SYSTEM_PROMPT, + userPrompt, + maxRetries: 1, + signal, + }); + } catch (e) { + if (isAbortError(e)) throw e; + console.error('[ST-BME] 记忆整合 LLM 调用失败:', e); + stats.kept = newEntries.length; + return stats; + } + + // ══════════════════════════════════════════════ + // Phase 4: 逐个处理结果 + // ══════════════════════════════════════════════ + + // 解析 LLM 返回——兼容单条和批量格式 + let results; + if (Array.isArray(decision?.results)) { + results = decision.results; + } else if (decision?.action) { + // 单条返回格式（LLM 可能忽略 results 包装） + results = [{ ...decision, node_id: newEntries[0]?.id }]; + } else { + console.warn('[ST-BME] 记忆整合: LLM 返回格式异常，全部 keep'); + stats.kept = newEntries.length; + return stats; + } + + // 建立 node_id → result 的映射 + const resultMap = new Map(); + for (const r of results) { + if (r.node_id) resultMap.set(r.node_id, r); + } + + // 处理每个新节点 + for (const entry of newEntries) { + const result = resultMap.get(entry.id); + if (!result) { + // LLM 未返回此节点的结果，fallback 为 keep + stats.kept++; + continue; + } + + processOneResult(graph, entry, result, stats); + } + + // 日志 const actionSummary = []; if (stats.merged > 0) actionSummary.push(`合并 ${stats.merged}`); if (stats.skipped > 0) actionSummary.push(`跳过 ${stats.skipped}`); @@ -339,3 +332,122 @@ export async function consolidateMemories({ return stats; } + +/** + * 处理单个节点的整合结果 + */ +function processOneResult(graph, entry, result, stats) { + const { id: newId, node: newNode } = entry; + + // ── 处理 action ── + switch (result.action) { + case 'skip': { + console.log(`[ST-BME] 记忆整合: skip (重复) — ${newId}`); + newNode.archived = true; + stats.skipped++; + break; + } + + case 'merge': { + const targetId = result.merge_target_id; + const targetNode = targetId ? getNode(graph, targetId) : null; + + if (targetNode && !targetNode.archived) { + console.log(`[ST-BME] 记忆整合: merge ${newId} → ${targetId}`); + + if (result.merged_fields && typeof result.merged_fields === 'object') { + for (const [key, value] of Object.entries(result.merged_fields)) { + if (value != null && value !== '') { + targetNode.fields[key] = value; + } + } + } else { + for (const [key, value] of Object.entries(newNode.fields)) { + if (value != null && value !== '' && !targetNode.fields[key]) { + targetNode.fields[key] = value; + } + } + } + + if (Number.isFinite(newNode.seq) && newNode.seq > (targetNode.seq || 0)) { + targetNode.seq = newNode.seq; + } + + targetNode.embedding = null; + newNode.archived = true; + stats.merged++; + } else { + console.warn(`[ST-BME] 记忆整合: merge target ${targetId} 不存在，回退为 keep`); + stats.kept++; + } + break; + } + + case 'keep': + default: { + stats.kept++; + break; + } + } + + // ── 处理 evolution ── + const evolution = result.evolution; + if (evolution?.should_evolve && !newNode.archived) { + stats.evolved++; + console.log(`[ST-BME] 记忆整合/进化触发: ${result.reason || '(无理由)'}`); + + if (Array.isArray(evolution.connections)) { + for (const targetId of evolution.connections) { + if (!getNode(graph, targetId)) continue; + const edge = createEdge({ + fromId: newId, + toId: targetId, + relation: 'related', + strength: 0.7, + }); + if (addEdge(graph, edge)) { + stats.connections++; + } + } + } + + if (Array.isArray(evolution.neighbor_updates)) { + for (const update of evolution.neighbor_updates) { + if (!update.nodeId) continue; + const oldNode = getNode(graph, update.nodeId); + if (!oldNode || oldNode.archived) continue; + + let changed = false; + + if (update.newContext && typeof update.newContext === 'string') { + if (oldNode.fields.state !== undefined) { + oldNode.fields.state = update.newContext; + changed = true; + } else if (oldNode.fields.summary !== undefined) { + oldNode.fields.summary = update.newContext; + changed = true; + } else if (oldNode.fields.core_note !== undefined) { + oldNode.fields.core_note = update.newContext; + changed = true; + } + } + + if (update.newTags && Array.isArray(update.newTags)) { + oldNode.clusters = update.newTags; + changed = true; + } + + if (changed) { + oldNode.embedding = null; + if (!oldNode._evolutionHistory) oldNode._evolutionHistory = []; + oldNode._evolutionHistory.push({ + triggeredBy: newId, + timestamp: Date.now(), + reason: result.reason || '', + }); + stats.updates++; + } + } + } + } +} diff --git a/panel.js b/panel.js index 61a5afe..0e0ebb9 100644 --- a/panel.js +++ b/panel.js @@ -67,17 +67,14 @@ const DEFAULT_PROMPTS = { "- 反向更新旧记忆", "", "输出严格 JSON：", - "{", - ' "action": "keep"|"merge"|"skip",', - ' "merge_target_id": "旧节点ID",', - ' "merged_fields": {},', - ' "reason": "理由",', - ' "evolution": {', - ' "should_evolve": true/false,', - ' "connections": ["旧记忆ID"],', - ' "neighbor_updates": [{"nodeId": "旧节点ID", "newContext": "修正描述", "newTags": ["标签"]}]', + '{ "results": [', + ' { "node_id": "新记忆节点ID",', + ' "action": "keep"|"merge"|"skip",', + ' "merge_target_id": "旧节点ID (仅merge)",', + ' "reason": "理由",', + ' "evolution": { "should_evolve": true/false, "connections": ["旧记忆ID"], "neighbor_updates": [...] }', " }", - "}", + "] }", ].join("\n"), compress: [