feat: 升级至 v0.2.1 — 流式渲染修复、安全增强、工具自动执行
流式渲染修复： - runId 机制防止 abort 后旧流事件污染新 run - run lock 防止并发 run 污染引擎状态 - abort race 提前退出工具执行等待 - TERMINATED 状态通过 stateChange 发射 - tool_call_delta 流式参数拼接 + pending 占位替换 - 首轮卡片创建路径统一，traceStep 按 ID 精确匹配 - compressed 事件转发为 toast 通知 安全增强： - ConfirmationHook 支持持久化自动执行（跨会话） - 设置面板新增自动执行工具管理 UI - SandboxManager 双重安全校验 fail-closed - 审计日志链式哈希防篡改 - PromptInjectionDefender 中文注入标记清理 - scanCode 28 模式 + base64/$() 检测 - validatePath realpathSync 防符号链接逃逸 - code-search 使用 execFile 防命令注入 新增工具： - file_editor、code_search、task_manager、diff_viewer 其他： - Agent Loop 加 PARSING/REFLECTING 状态 + 指数退避重试 - MemoryManager TF-IDF 语义检索 - run_command Windows 中文编码修复（chcp 65001） - 版本号 0.2.0 → 0.2.1
This commit is contained in:
thzxx
@@ -4,6 +4,11 @@
|
||||
* 基于 SQLite（better-sqlite3）的三层记忆系统。
|
||||
* 表结构由 DatabaseService 统一创建，此处不再重复。
|
||||
*
|
||||
* v0.2.0 增强：
|
||||
* - TF-IDF 语义检索替代 LIKE 关键词搜索
|
||||
* - 时间衰减策略：老旧记忆权重降低
|
||||
* - IDF 缓存：避免每次搜索重新计算
|
||||
*
|
||||
* @see docs/生产级通用 AI Agent 智能体桌面应用：完整设计与构建指南.html — 第六章
|
||||
* @see standard/开发规范.md — 使用 better-sqlite3（禁止自写数据库层）
|
||||
*/
|
||||
@@ -38,17 +43,258 @@ interface MemorySearchOptions {
|
||||
minImportance?: number;
|
||||
}
|
||||
|
||||
/** 分词：将文本拆分为词项（支持中英文） */
|
||||
function tokenize(text: string): string[] {
|
||||
// 转小写，按非字母数字字符分割
|
||||
const lower = text.toLowerCase();
|
||||
// 英文词
|
||||
const words = lower.match(/[a-z][a-z0-9_-]{1,}/g) ?? [];
|
||||
// 中文 bigram（相邻两字组合），覆盖 CJK 统一表意、扩展 A、平假名/片假名、谚文
|
||||
const cjkChars = lower.match(/[\u4e00-\u9fff\u3400-\u4dbf\u3040-\u30ff\uac00-\ud7af]/g) ?? [];
|
||||
const bigrams: string[] = [];
|
||||
for (let i = 0; i < cjkChars.length - 1; i++) {
|
||||
bigrams.push(cjkChars[i] + cjkChars[i + 1]);
|
||||
}
|
||||
// 单字 CJK 补 unigram（避免单字文档无 token）
|
||||
if (cjkChars.length === 1) {
|
||||
bigrams.push(cjkChars[0]);
|
||||
}
|
||||
return [...words, ...bigrams];
|
||||
}
|
||||
|
||||
/** 计算词频（TF） */
|
||||
function computeTF(tokens: string[]): Map<string, number> {
|
||||
const tf = new Map<string, number>();
|
||||
for (const token of tokens) {
|
||||
tf.set(token, (tf.get(token) ?? 0) + 1);
|
||||
}
|
||||
// 归一化
|
||||
const total = tokens.length || 1;
|
||||
for (const [key, val] of tf) {
|
||||
tf.set(key, val / total);
|
||||
}
|
||||
return tf;
|
||||
}
|
||||
|
||||
/** 计算余弦相似度的点积部分 */
|
||||
function dotProduct(tf1: Map<string, number>, tf2: Map<string, number>, idf: Map<string, number>): number {
|
||||
let sum = 0;
|
||||
for (const [term, freq1] of tf1) {
|
||||
const freq2 = tf2.get(term);
|
||||
if (freq2 !== undefined) {
|
||||
const idfVal = idf.get(term) ?? 1;
|
||||
sum += freq1 * freq2 * idfVal * idfVal;
|
||||
}
|
||||
}
|
||||
return sum;
|
||||
}
|
||||
|
||||
/** 计算向量模长 */
|
||||
function vectorNorm(tf: Map<string, number>, idf: Map<string, number>): number {
|
||||
let sum = 0;
|
||||
for (const [term, freq] of tf) {
|
||||
const idfVal = idf.get(term) ?? 1;
|
||||
sum += (freq * idfVal) ** 2;
|
||||
}
|
||||
return Math.sqrt(sum);
|
||||
}
|
||||
|
||||
/** 时间衰减权重：30 天半衰期（age=30 时 weight=0.5） */
|
||||
function timeDecayWeight(createdAt: number, now: number = Date.now()): number {
|
||||
const ageDays = Math.max(0, (now - createdAt) / (24 * 60 * 60 * 1000));
|
||||
const halfLifeDays = 30;
|
||||
return Math.pow(0.5, ageDays / halfLifeDays);
|
||||
}
|
||||
|
||||
/**
|
||||
* 记忆管理器
|
||||
*/
|
||||
export class MemoryManager {
|
||||
/** IDF 缓存：词项 -> 文档频率 */
|
||||
private idfCache = new Map<string, number>();
|
||||
/** 缓存的记忆总数 */
|
||||
private cachedDocCount = 0;
|
||||
/** 缓存最后更新时间 */
|
||||
private cacheUpdatedAt = 0;
|
||||
/** 缓存有效期（5 分钟） */
|
||||
private readonly CACHE_TTL = 5 * 60 * 1000;
|
||||
|
||||
constructor(private getDB: () => Database.Database) {}
|
||||
|
||||
/**
|
||||
* 初始化（表结构由 DatabaseService 创建）
|
||||
*/
|
||||
initialize(): void {
|
||||
log.info('MemoryManager initialized');
|
||||
log.info('MemoryManager initialized (v0.2.0: TF-IDF enabled)');
|
||||
}
|
||||
|
||||
/**
|
||||
* 更新 IDF 缓存
|
||||
*/
|
||||
private updateIdfCache(): void {
|
||||
const now = Date.now();
|
||||
if (now - this.cacheUpdatedAt < this.CACHE_TTL && this.cachedDocCount > 0) {
|
||||
return; // 缓存未过期
|
||||
}
|
||||
|
||||
const db = this.getDB();
|
||||
this.idfCache.clear();
|
||||
|
||||
// 获取所有记忆内容（episodic + semantic + working）
|
||||
const episodicRows = db.prepare('SELECT content, summary FROM episodic_memories').all() as Array<{ content: string; summary: string | null }>;
|
||||
const semanticRows = db.prepare('SELECT value FROM semantic_memories').all() as Array<{ value: string }>;
|
||||
const workingRows = db.prepare('SELECT value FROM working_memories').all() as Array<{ value: string }>;
|
||||
|
||||
const allDocs = [
|
||||
...episodicRows.map((r) => r.content + ' ' + (r.summary ?? '')),
|
||||
...semanticRows.map((r) => r.value),
|
||||
...workingRows.map((r) => r.value),
|
||||
];
|
||||
|
||||
this.cachedDocCount = allDocs.length;
|
||||
const docFreq = new Map<string, number>();
|
||||
|
||||
for (const doc of allDocs) {
|
||||
const tokens = new Set(tokenize(doc));
|
||||
for (const token of tokens) {
|
||||
docFreq.set(token, (docFreq.get(token) ?? 0) + 1);
|
||||
}
|
||||
}
|
||||
|
||||
// IDF = log((N+1)/(df+1)) + 1（Sklearn 风格平滑），确保非负
|
||||
for (const [term, df] of docFreq) {
|
||||
this.idfCache.set(term, Math.log((this.cachedDocCount + 1) / (df + 1)) + 1);
|
||||
}
|
||||
|
||||
this.cacheUpdatedAt = now;
|
||||
}
|
||||
|
||||
/**
|
||||
* TF-IDF 相似度搜索
|
||||
*/
|
||||
private tfidfSearch(query: string, options: MemorySearchOptions): SearchResult[] {
|
||||
const db = this.getDB();
|
||||
this.updateIdfCache();
|
||||
|
||||
const queryTokens = tokenize(query);
|
||||
if (queryTokens.length === 0) return [];
|
||||
|
||||
const queryTF = computeTF(queryTokens);
|
||||
const queryNorm = vectorNorm(queryTF, this.idfCache);
|
||||
if (queryNorm === 0) return [];
|
||||
|
||||
const { topK = 5, type, minImportance = 0 } = options;
|
||||
const results: SearchResult[] = [];
|
||||
const now = Date.now();
|
||||
|
||||
// 搜索 episodic 记忆
|
||||
if (!type || type === 'episodic') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM episodic_memories WHERE importance >= ?
|
||||
ORDER BY importance DESC, created_at DESC LIMIT ?
|
||||
`).all(minImportance, topK * 3) as Array<{
|
||||
id: string; session_id: string | null; content: string; summary: string | null;
|
||||
source: string; importance: number; created_at: number; expires_at: number | null;
|
||||
}>;
|
||||
|
||||
for (const row of rows) {
|
||||
const docText = row.content + ' ' + (row.summary ?? '');
|
||||
const docTokens = tokenize(docText);
|
||||
const docTF = computeTF(docTokens);
|
||||
const docNorm = vectorNorm(docTF, this.idfCache);
|
||||
|
||||
if (docNorm === 0) continue;
|
||||
|
||||
const dotProd = dotProduct(queryTF, docTF, this.idfCache);
|
||||
const cosineSim = dotProd / (queryNorm * docNorm);
|
||||
|
||||
// 时间衰减
|
||||
const decayWeight = timeDecayWeight(row.created_at, now);
|
||||
// 最终分数 = 余弦相似度 * 时间衰减 * 重要度权重
|
||||
const finalScore = cosineSim * decayWeight * (0.5 + row.importance * 0.5);
|
||||
|
||||
if (finalScore > 0) {
|
||||
results.push({
|
||||
id: row.id, type: 'episodic', content: row.content,
|
||||
summary: row.summary ?? undefined, source: row.source as MemorySource,
|
||||
importance: row.importance, sessionId: row.session_id ?? undefined,
|
||||
createdAt: row.created_at, expiresAt: row.expires_at ?? undefined,
|
||||
score: finalScore,
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 搜索 semantic 记忆
|
||||
if (!type || type === 'semantic') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM semantic_memories WHERE confidence >= ?
|
||||
ORDER BY confidence DESC, access_count DESC LIMIT ?
|
||||
`).all(minImportance, Math.ceil(topK * 1.5)) as Array<{
|
||||
id: string; key: string; value: string; category: string | null;
|
||||
confidence: number; source_session: string | null; created_at: number;
|
||||
}>;
|
||||
|
||||
for (const row of rows) {
|
||||
const docText = row.key + ' ' + row.value;
|
||||
const docTokens = tokenize(docText);
|
||||
const docTF = computeTF(docTokens);
|
||||
const docNorm = vectorNorm(docTF, this.idfCache);
|
||||
|
||||
if (docNorm === 0) continue;
|
||||
|
||||
const dotProd = dotProduct(queryTF, docTF, this.idfCache);
|
||||
const cosineSim = dotProd / (queryNorm * docNorm);
|
||||
|
||||
const decayWeight = timeDecayWeight(row.created_at, now);
|
||||
const finalScore = cosineSim * decayWeight * (0.5 + row.confidence * 0.5);
|
||||
|
||||
if (finalScore > 0) {
|
||||
results.push({
|
||||
id: row.id, type: 'semantic', content: row.value,
|
||||
source: 'imported', importance: row.confidence,
|
||||
sessionId: row.source_session ?? undefined,
|
||||
createdAt: row.created_at, score: finalScore,
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 搜索 working 记忆
|
||||
if (!type || type === 'working') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM working_memories ORDER BY updated_at DESC LIMIT ?
|
||||
`).all(topK * 3) as Array<{
|
||||
id: string; session_id: string; task_id: string;
|
||||
key: string; value: string; updated_at: number;
|
||||
}>;
|
||||
|
||||
for (const row of rows) {
|
||||
const docText = row.key + ' ' + row.value;
|
||||
const docTokens = tokenize(docText);
|
||||
const docTF = computeTF(docTokens);
|
||||
const docNorm = vectorNorm(docTF, this.idfCache);
|
||||
|
||||
if (docNorm === 0) continue;
|
||||
|
||||
const dotProd = dotProduct(queryTF, docTF, this.idfCache);
|
||||
const cosineSim = dotProd / (queryNorm * docNorm);
|
||||
|
||||
const decayWeight = timeDecayWeight(row.updated_at, now);
|
||||
const finalScore = cosineSim * decayWeight * 0.5;
|
||||
|
||||
if (finalScore > 0) {
|
||||
results.push({
|
||||
id: row.id, type: 'working', content: row.value,
|
||||
source: 'agent_thought', importance: 0.5,
|
||||
sessionId: row.session_id, createdAt: row.updated_at,
|
||||
score: finalScore,
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return results.sort((a, b) => b.score - a.score).slice(0, topK);
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -81,26 +327,44 @@ export class MemoryManager {
|
||||
break;
|
||||
}
|
||||
|
||||
// 使 IDF 缓存失效
|
||||
this.cacheUpdatedAt = 0;
|
||||
|
||||
log.debug(`Memory stored: ${id} (${item.type})`);
|
||||
return id;
|
||||
}
|
||||
|
||||
/**
|
||||
* 检索记忆（关键词搜索）
|
||||
* 检索记忆（v0.2.0: TF-IDF 语义检索 + 时间衰减）
|
||||
*
|
||||
* v0.2.0 变更：
|
||||
* - 使用 TF-IDF 余弦相似度替代 LIKE 关键词搜索
|
||||
* - 支持中英文分词（英文按词，中文按 bigram）
|
||||
* - 时间衰减：30 天半衰期，老旧记忆权重降低
|
||||
* - IDF 缓存：5 分钟有效期，避免重复计算
|
||||
*/
|
||||
search(query: string, options: MemorySearchOptions = {}): SearchResult[] {
|
||||
const db = this.getDB();
|
||||
const { topK = 5, type, minImportance = 0 } = options;
|
||||
if (!query) return [];
|
||||
|
||||
const pattern = `%${query}%`;
|
||||
// v0.2.0: 优先使用 TF-IDF 语义搜索
|
||||
const tfidfResults = this.tfidfSearch(query, options);
|
||||
if (tfidfResults.length > 0) {
|
||||
return tfidfResults;
|
||||
}
|
||||
|
||||
// 回退：如果 TF-IDF 没有结果（如 IDF 缓存为空），使用 LIKE 关键词搜索
|
||||
// 转义 LIKE 通配符，避免用户输入的 % 和 _ 影响匹配
|
||||
const escapedQuery = query.replace(/[%_]/g, '\\$&');
|
||||
const pattern = `%${escapedQuery}%`;
|
||||
const results: SearchResult[] = [];
|
||||
|
||||
// 搜索情节记忆
|
||||
if (!type || type === 'episodic') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM episodic_memories
|
||||
WHERE (content LIKE ? OR summary LIKE ?) AND importance >= ?
|
||||
WHERE (content LIKE ? ESCAPE '\\' OR summary LIKE ? ESCAPE '\\') AND importance >= ?
|
||||
ORDER BY importance DESC, created_at DESC LIMIT ?
|
||||
`).all(pattern, pattern, minImportance, topK) as Array<{
|
||||
id: string; session_id: string | null; content: string; summary: string | null;
|
||||
@@ -112,7 +376,7 @@ export class MemoryManager {
|
||||
summary: row.summary ?? undefined, source: row.source as MemorySource,
|
||||
importance: row.importance, sessionId: row.session_id ?? undefined,
|
||||
createdAt: row.created_at, expiresAt: row.expires_at ?? undefined,
|
||||
score: row.importance,
|
||||
score: row.importance * timeDecayWeight(row.created_at),
|
||||
});
|
||||
}
|
||||
}
|
||||
@@ -121,7 +385,7 @@ export class MemoryManager {
|
||||
if (!type || type === 'semantic') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM semantic_memories
|
||||
WHERE (key LIKE ? OR value LIKE ?) AND confidence >= ?
|
||||
WHERE (key LIKE ? ESCAPE '\\' OR value LIKE ? ESCAPE '\\') AND confidence >= ?
|
||||
ORDER BY confidence DESC, access_count DESC LIMIT ?
|
||||
`).all(pattern, pattern, minImportance, Math.ceil(topK / 2)) as Array<{
|
||||
id: string; key: string; value: string; category: string | null;
|
||||
@@ -132,7 +396,7 @@ export class MemoryManager {
|
||||
id: row.id, type: 'semantic', content: row.value,
|
||||
source: 'imported', importance: row.confidence,
|
||||
sessionId: row.source_session ?? undefined,
|
||||
createdAt: row.created_at, score: row.confidence,
|
||||
createdAt: row.created_at, score: row.confidence * timeDecayWeight(row.created_at),
|
||||
});
|
||||
}
|
||||
}
|
||||
@@ -141,7 +405,7 @@ export class MemoryManager {
|
||||
if (type === 'working') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM working_memories
|
||||
WHERE (key LIKE ? OR value LIKE ?)
|
||||
WHERE (key LIKE ? ESCAPE '\\' OR value LIKE ? ESCAPE '\\')
|
||||
ORDER BY updated_at DESC LIMIT ?
|
||||
`).all(pattern, pattern, topK) as Array<{
|
||||
id: string; session_id: string; task_id: string;
|
||||
@@ -152,7 +416,7 @@ export class MemoryManager {
|
||||
id: row.id, type: 'working', content: row.value,
|
||||
source: 'agent_thought', importance: 0.5,
|
||||
sessionId: row.session_id, createdAt: row.updated_at,
|
||||
score: 0.3, // 工作记忆权重较低
|
||||
score: 0.3 * timeDecayWeight(row.updated_at),
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user