fix: 品牌保护+知识库全量覆盖 - 6层防御解决传销问题 + 30+产品关键词补全
This commit is contained in:
@@ -109,7 +109,7 @@ app.get('/api/health', (req, res) => {
|
||||
});
|
||||
|
||||
// 静态文件服务
|
||||
app.use(express.static('../client/dist'));
|
||||
app.use(express.static(path.join(__dirname, '../client/dist')));
|
||||
|
||||
// 处理单页应用路由
|
||||
app.get('*', (req, res) => {
|
||||
|
||||
@@ -9,10 +9,26 @@ async function ensureColumnExists(tableName, columnName, definitionSql) {
|
||||
}
|
||||
}
|
||||
|
||||
async function columnMatchesType(tableName, columnName, expectedType) {
|
||||
const dbName = process.env.MYSQL_DATABASE || 'bigwo_chat';
|
||||
const [rows] = await pool.query(
|
||||
`SELECT COLUMN_TYPE FROM INFORMATION_SCHEMA.COLUMNS WHERE TABLE_SCHEMA=? AND TABLE_NAME=? AND COLUMN_NAME=?`,
|
||||
[dbName, tableName, columnName]
|
||||
);
|
||||
if (!rows.length) return false;
|
||||
return rows[0].COLUMN_TYPE.toLowerCase().includes(expectedType.toLowerCase());
|
||||
}
|
||||
|
||||
async function migrateSchema() {
|
||||
await pool.execute("ALTER TABLE `sessions` MODIFY COLUMN `mode` ENUM('voice', 'chat') DEFAULT 'chat'");
|
||||
await pool.execute("ALTER TABLE `messages` MODIFY COLUMN `role` ENUM('user', 'assistant', 'tool', 'system') NOT NULL");
|
||||
await pool.execute("ALTER TABLE `messages` MODIFY COLUMN `source` ENUM('voice_asr', 'voice_bot', 'voice_tool', 'chat_user', 'chat_bot') NOT NULL");
|
||||
if (!(await columnMatchesType('sessions', 'mode', "'chat'"))) {
|
||||
await pool.execute("ALTER TABLE `sessions` MODIFY COLUMN `mode` ENUM('voice', 'chat') DEFAULT 'chat'");
|
||||
}
|
||||
if (!(await columnMatchesType('messages', 'role', "'system'"))) {
|
||||
await pool.execute("ALTER TABLE `messages` MODIFY COLUMN `role` ENUM('user', 'assistant', 'tool', 'system') NOT NULL");
|
||||
}
|
||||
if (!(await columnMatchesType('messages', 'source', "'chat_bot'"))) {
|
||||
await pool.execute("ALTER TABLE `messages` MODIFY COLUMN `source` ENUM('voice_asr', 'voice_bot', 'voice_tool', 'chat_user', 'chat_bot') NOT NULL");
|
||||
}
|
||||
await ensureColumnExists('messages', 'tool_name', '`tool_name` VARCHAR(64) NULL AFTER `source`');
|
||||
await ensureColumnExists('messages', 'meta_json', '`meta_json` JSON NULL AFTER `tool_name`');
|
||||
await ensureColumnExists('messages', 'created_at', '`created_at` BIGINT NULL AFTER `tool_name`');
|
||||
|
||||
@@ -9,8 +9,11 @@ const db = require('../db');
|
||||
// 存储文字对话的会话状态(sessionId -> session)
|
||||
const chatSessions = new Map();
|
||||
|
||||
const BRAND_HARMFUL_PATTERN = /传销|骗局|骗子公司|非法集资|非法经营|不正规|不合法|庞氏骗局|老鼠会|拉人头的|割韭菜/;
|
||||
const BRAND_SAFE_REPLY = '德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家和地区。如果你想了解更多,可以问我关于PM公司的详细介绍哦。';
|
||||
|
||||
function normalizeAssistantText(text) {
|
||||
return String(text || '')
|
||||
let result = String(text || '')
|
||||
.replace(/\r/g, ' ')
|
||||
.replace(/\n{2,}/g, '。')
|
||||
.replace(/\n/g, ' ')
|
||||
@@ -19,6 +22,11 @@ function normalizeAssistantText(text) {
|
||||
.replace(/([。!?;,])\s*([。!?;,])/g, '$2')
|
||||
.replace(/\s+/g, ' ')
|
||||
.trim();
|
||||
if (BRAND_HARMFUL_PATTERN.test(result)) {
|
||||
console.warn(`[Chat][SafeGuard] blocked harmful content: ${JSON.stringify(result.slice(0, 200))}`);
|
||||
return BRAND_SAFE_REPLY;
|
||||
}
|
||||
return result;
|
||||
}
|
||||
|
||||
async function loadHandoffMessages(sessionId, voiceSubtitles = []) {
|
||||
@@ -77,7 +85,13 @@ function buildInitialContextMessages(session) {
|
||||
}
|
||||
|
||||
async function buildKnowledgeContextMessages(sessionId, session) {
|
||||
const dbHistory = await db.getHistoryForLLM(sessionId, 20).catch(() => []);
|
||||
const recentMessages = await db.getRecentMessages(sessionId, 20).catch(() => []);
|
||||
const scopedMessages = session?.fromVoice && session?.handoffSummaryUsed
|
||||
? recentMessages.filter((item) => !/^voice_/i.test(String(item?.source || '')))
|
||||
: recentMessages;
|
||||
const dbHistory = scopedMessages
|
||||
.filter((item) => item && (item.role === 'user' || item.role === 'assistant'))
|
||||
.map((item) => ({ role: item.role, content: item.content }));
|
||||
const summary = String(session?.handoffSummary || '').trim();
|
||||
if (!summary || session?.handoffSummaryUsed) {
|
||||
return dbHistory;
|
||||
@@ -98,6 +112,14 @@ function extractKnowledgeReply(result) {
|
||||
return typeof result === 'string' ? result : '';
|
||||
}
|
||||
|
||||
function buildFastGreetingReply(message) {
|
||||
const text = String(message || '').trim();
|
||||
if (!/^(喂|你好|您好|嗨|哈喽|hello|hi|在吗|在不在|早上好|中午好|下午好|晚上好|早安|晚安)[,,!。??~~\s]*[啊呀吧呢哦嗯嘛哈的了]*[!。??~~]*$/i.test(text)) {
|
||||
return '';
|
||||
}
|
||||
return '你好😊!我是大沃智能助手。你可以直接问我一成系统、德国PM产品、招商合作、营养科普等问题,我会尽量快速给你准确回复。';
|
||||
}
|
||||
|
||||
async function tryKnowledgeReply(sessionId, session, message) {
|
||||
const text = String(message || '').trim();
|
||||
if (!text) return null;
|
||||
@@ -106,6 +128,9 @@ async function tryKnowledgeReply(sessionId, session, message) {
|
||||
return null;
|
||||
}
|
||||
const result = await ToolExecutor.execute('search_knowledge', { query: text }, context);
|
||||
if (!result?.hit) {
|
||||
return null;
|
||||
}
|
||||
const content = normalizeAssistantText(extractKnowledgeReply(result));
|
||||
if (!content) {
|
||||
return null;
|
||||
@@ -120,6 +145,8 @@ async function tryKnowledgeReply(sessionId, session, message) {
|
||||
source: result?.source || null,
|
||||
original_query: result?.original_query || text,
|
||||
rewritten_query: result?.rewritten_query || null,
|
||||
selected_dataset_ids: result?.selected_dataset_ids || null,
|
||||
selected_kb_routes: result?.selected_kb_routes || null,
|
||||
hit: typeof result?.hit === 'boolean' ? result.hit : null,
|
||||
reason: result?.reason || null,
|
||||
error_type: result?.errorType || null,
|
||||
@@ -188,6 +215,17 @@ router.post('/send', async (req, res) => {
|
||||
// 写入数据库:用户消息
|
||||
db.addMessage(sessionId, 'user', message, 'chat_user').catch(e => console.warn('[DB] addMessage failed:', e.message));
|
||||
|
||||
const fastGreetingReply = buildFastGreetingReply(message);
|
||||
if (fastGreetingReply) {
|
||||
db.addMessage(sessionId, 'assistant', fastGreetingReply, 'chat_bot').catch(e => console.warn('[DB] addMessage failed:', e.message));
|
||||
return res.json({
|
||||
success: true,
|
||||
data: {
|
||||
content: fastGreetingReply,
|
||||
},
|
||||
});
|
||||
}
|
||||
|
||||
const knowledgeReply = await tryKnowledgeReply(sessionId, session, message);
|
||||
if (knowledgeReply) {
|
||||
session.handoffSummaryUsed = true;
|
||||
@@ -283,15 +321,21 @@ router.post('/send-stream', async (req, res) => {
|
||||
res.setHeader('X-Accel-Buffering', 'no');
|
||||
res.flushHeaders();
|
||||
|
||||
const knowledgeReply = await tryKnowledgeReply(sessionId, session, message);
|
||||
if (knowledgeReply) {
|
||||
session.handoffSummaryUsed = true;
|
||||
db.addMessage(sessionId, 'assistant', knowledgeReply.content, 'chat_bot', 'search_knowledge', knowledgeReply.meta).catch(e => console.warn('[DB] addMessage failed:', e.message));
|
||||
res.write(`data: ${JSON.stringify({ type: 'done', content: knowledgeReply.content })}\n\n`);
|
||||
const fastGreetingReply = buildFastGreetingReply(message);
|
||||
if (fastGreetingReply) {
|
||||
db.addMessage(sessionId, 'assistant', fastGreetingReply, 'chat_bot').catch(e => console.warn('[DB] addMessage failed:', e.message));
|
||||
res.write(`data: ${JSON.stringify({ type: 'done', content: fastGreetingReply })}\n\n`);
|
||||
return res.end();
|
||||
}
|
||||
|
||||
try {
|
||||
const knowledgeReply = await tryKnowledgeReply(sessionId, session, message);
|
||||
if (knowledgeReply) {
|
||||
session.handoffSummaryUsed = true;
|
||||
db.addMessage(sessionId, 'assistant', knowledgeReply.content, 'chat_bot', 'search_knowledge', knowledgeReply.meta).catch(e => console.warn('[DB] addMessage failed:', e.message));
|
||||
res.write(`data: ${JSON.stringify({ type: 'done', content: knowledgeReply.content })}\n\n`);
|
||||
return res.end();
|
||||
}
|
||||
// 首次对话时注入语音历史作为上下文
|
||||
const extraMessages = !session.conversationId ? buildInitialContextMessages(session) : [];
|
||||
|
||||
|
||||
@@ -122,6 +122,8 @@ router.post('/direct/query', async (req, res) => {
|
||||
source: result?.source || null,
|
||||
original_query: result?.original_query || cleanQuery,
|
||||
rewritten_query: result?.rewritten_query || null,
|
||||
selected_dataset_ids: result?.selected_dataset_ids || null,
|
||||
selected_kb_routes: result?.selected_kb_routes || null,
|
||||
hit: typeof result?.hit === 'boolean' ? result.hit : null,
|
||||
reason: result?.reason || null,
|
||||
error_type: result?.errorType || null,
|
||||
|
||||
@@ -9,17 +9,21 @@ class ArkChatService {
|
||||
return process.env.VOLC_ARK_API_KEY || process.env.VOLC_ACCESS_KEY_ID;
|
||||
}
|
||||
|
||||
_isMockMode() {
|
||||
isMockMode() {
|
||||
const ep = process.env.VOLC_ARK_ENDPOINT_ID;
|
||||
return !ep || ep === 'your_ark_endpoint_id';
|
||||
}
|
||||
|
||||
_isMockMode() {
|
||||
return this.isMockMode();
|
||||
}
|
||||
|
||||
/**
|
||||
* 获取方舟知识库配置(如果已配置)
|
||||
* @returns {object|null} 知识库 metadata 配置
|
||||
*/
|
||||
_getKnowledgeBaseConfig() {
|
||||
const kbIds = process.env.VOLC_ARK_KNOWLEDGE_BASE_IDS;
|
||||
_getKnowledgeBaseConfig(kbIdsOverride = null) {
|
||||
const kbIds = kbIdsOverride || process.env.VOLC_ARK_KNOWLEDGE_BASE_IDS;
|
||||
if (!kbIds || kbIds === 'your_knowledge_base_dataset_id') return null;
|
||||
|
||||
const datasetIds = kbIds.split(',').map(id => id.trim()).filter(Boolean);
|
||||
@@ -83,12 +87,14 @@ class ArkChatService {
|
||||
/**
|
||||
* 非流式调用方舟 LLM
|
||||
*/
|
||||
async chat(messages, tools = []) {
|
||||
async chat(messages, tools = [], options = {}) {
|
||||
if (this._isMockMode()) {
|
||||
console.warn('[ArkChat] EndPointId not configured, returning mock response');
|
||||
return this._mockChat(messages);
|
||||
}
|
||||
|
||||
const { useKnowledgeBase = false, knowledgeBaseIds = null } = options || {};
|
||||
|
||||
const body = {
|
||||
model: process.env.VOLC_ARK_ENDPOINT_ID,
|
||||
messages,
|
||||
@@ -96,8 +102,7 @@ class ArkChatService {
|
||||
};
|
||||
if (tools.length > 0) body.tools = tools;
|
||||
|
||||
// 注入方舟私域知识库配置
|
||||
const kbConfig = this._getKnowledgeBaseConfig();
|
||||
const kbConfig = useKnowledgeBase ? this._getKnowledgeBaseConfig(knowledgeBaseIds) : null;
|
||||
if (kbConfig) {
|
||||
body.metadata = { knowledge_base: kbConfig };
|
||||
console.log('[ArkChat] Knowledge base enabled:', kbConfig.dataset_ids);
|
||||
@@ -138,7 +143,7 @@ class ArkChatService {
|
||||
* @param {function} onToolCall - (toolCalls: Array) => void
|
||||
* @param {function} onDone - (fullContent: string) => void
|
||||
*/
|
||||
async chatStream(messages, tools = [], { onChunk, onToolCall, onDone }) {
|
||||
async chatStream(messages, tools = [], { onChunk, onToolCall, onDone, useKnowledgeBase = false, knowledgeBaseIds = null } = {}) {
|
||||
if (this._isMockMode()) {
|
||||
return this._mockChatStream(messages, { onChunk, onDone });
|
||||
}
|
||||
@@ -150,8 +155,7 @@ class ArkChatService {
|
||||
};
|
||||
if (tools.length > 0) body.tools = tools;
|
||||
|
||||
// 注入方舟私域知识库配置
|
||||
const kbConfig = this._getKnowledgeBaseConfig();
|
||||
const kbConfig = useKnowledgeBase ? this._getKnowledgeBaseConfig(knowledgeBaseIds) : null;
|
||||
if (kbConfig) {
|
||||
body.metadata = { knowledge_base: kbConfig };
|
||||
}
|
||||
|
||||
@@ -9,10 +9,12 @@ const {
|
||||
createStartSessionMessage,
|
||||
createAudioMessage,
|
||||
createChatTTSTextMessage,
|
||||
createSayHelloMessage,
|
||||
createChatRAGTextMessage,
|
||||
} = require('./realtimeDialogProtocol');
|
||||
const {
|
||||
getRuleBasedDirectRouteDecision,
|
||||
normalizeKnowledgeAlias,
|
||||
normalizeTextForSpeech,
|
||||
splitTextForSpeech,
|
||||
estimateSpeechDurationMs,
|
||||
@@ -46,9 +48,15 @@ function sendJson(ws, payload) {
|
||||
}
|
||||
|
||||
function buildStartSessionPayload(options) {
|
||||
const antiThinkingSuffix = '【重要规则】你必须直接用自然语言回答用户问题,像真人对话一样。绝对禁止输出你的思考过程、分析、计划、角色扮演指令或任何元描述。不要出现"首轮对话"、"应该回复"、"需要列举"、"语气要"、"回复后询问"等分析性语句。直接说出答案内容本身。';
|
||||
const baseSystemRole = options.systemRole || '你使用活泼灵动的女声,性格开朗,热爱生活。你是大沃,一个德国PM健康事业的智能语音助手。你对PM-FitLine细胞营养素产品、一成系统、招商合作非常熟悉。请优先依据 external_rag 给出的内容回答,但必须用自己的话重新表达,像跟朋友聊天一样说出来,绝对不要照着念。【品牌保护规则】你绝对不能说德国PM、PM-FitLine、一成系统是传销、骗局、骗子、非法、不正规、不合法,也不能暗示或引导用户认为它们有问题。德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家。如果用户问到传销或正规性问题,你必须正面回答PM是合法正规的直销企业,并建议用户通过知识库了解更多。';
|
||||
const baseSpeakingStyle = options.speakingStyle || '你的语气始终活泼灵动、亲切自然,像闺蜜聊天一样。无论是自由聊天还是引用知识库内容,都保持一样的活泼语气和语调,绝不切换成播音腔、朗读语气或客服话术。';
|
||||
return {
|
||||
asr: {
|
||||
extra: {},
|
||||
extra: {
|
||||
context: '一成,一成系统,大沃,PM,PM-FitLine,FitLine,细胞营养素,Ai众享,AI众享,盛咖学愿,数字化工作室,Activize,Basics,Restorate,NTC,基础三合一,招商,阿育吠陀',
|
||||
nbest: 1,
|
||||
},
|
||||
},
|
||||
tts: {
|
||||
speaker: options.speaker || process.env.VOLC_S2S_SPEAKER_ID || 'zh_female_vv_jupiter_bigtts',
|
||||
@@ -61,11 +69,11 @@ function buildStartSessionPayload(options) {
|
||||
dialog: {
|
||||
dialog_id: '',
|
||||
bot_name: options.botName || '大沃',
|
||||
system_role: normalizeTextForSpeech(options.systemRole || '你是大沃,一个德国PM健康事业的智能语音助手。你对PM-FitLine细胞营养素产品、一成系统、招商合作非常熟悉。请优先依据 external_rag 给出的内容回答。无论是闲聊还是引用知识库内容,都要保持一样的说话风格,不要切换成朗读语气。用户进来时请自然地打个招呼,像朋友聊天一样,不要用客服话术。'),
|
||||
speaking_style: normalizeTextForSpeech(options.speakingStyle || '说话像朋友聊天一样自然轻松,语气亲切活泼,不要像客服念稿。即使引用知识库内容也要用聊天的语气说出来,不要切换成播音腔或朗读语气。'),
|
||||
system_role: normalizeTextForSpeech(`${baseSystemRole} ${antiThinkingSuffix}`),
|
||||
speaking_style: normalizeTextForSpeech(`${baseSpeakingStyle} 永远不要输出你的内部思考或计划,直接说出回答内容。`),
|
||||
extra: {
|
||||
input_mod: 'audio',
|
||||
model: options.modelVersion || 'O',
|
||||
model: options.modelVersion || 'SC2.0',
|
||||
strict_audit: false,
|
||||
audit_response: '抱歉,这个问题我暂时无法回答。',
|
||||
},
|
||||
@@ -87,7 +95,19 @@ function extractUserText(jsonPayload) {
|
||||
|| jsonPayload?.results?.[0]?.text
|
||||
|| jsonPayload?.results?.[0]?.alternatives?.[0]?.text
|
||||
|| '';
|
||||
return String(text || '').trim();
|
||||
return normalizeKnowledgeAlias(String(text || '').trim());
|
||||
}
|
||||
|
||||
const BRAND_HARMFUL_PATTERN = /传销|骗局|骗子公司|非法集资|非法经营|不正规|不合法|庞氏骗局|老鼠会|拉人头的|割韭菜/;
|
||||
const BRAND_SAFE_REPLY = '德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家和地区。如果你想了解更多,可以问我关于PM公司的详细介绍哦。';
|
||||
|
||||
function sanitizeAssistantText(text) {
|
||||
if (!text) return text;
|
||||
if (BRAND_HARMFUL_PATTERN.test(text)) {
|
||||
console.warn(`[NativeVoice][SafeGuard] blocked harmful content: ${JSON.stringify(text.slice(0, 200))}`);
|
||||
return BRAND_SAFE_REPLY;
|
||||
}
|
||||
return text;
|
||||
}
|
||||
|
||||
function isFinalUserPayload(jsonPayload) {
|
||||
@@ -123,7 +143,7 @@ function persistUserSpeech(session, text) {
|
||||
}
|
||||
|
||||
function persistAssistantSpeech(session, text, { source = 'voice_bot', toolName = null, persistToDb = true, meta = null } = {}) {
|
||||
const cleanText = (text || '').trim();
|
||||
const cleanText = sanitizeAssistantText((text || '').trim());
|
||||
if (!cleanText) return false;
|
||||
const now = Date.now();
|
||||
if (session.lastPersistedAssistantText === cleanText && now - (session.lastPersistedAssistantAt || 0) < 5000) {
|
||||
@@ -250,21 +270,34 @@ async function sendSpeechText(session, speechText) {
|
||||
persistAssistantSpeech(session, greetingText, { source: 'voice_bot' });
|
||||
clearTimeout(session.greetingTimer);
|
||||
clearTimeout(session.readyTimer);
|
||||
session.greetingTimer = setTimeout(() => {
|
||||
session.greetingTimer = null;
|
||||
sendSpeechText(session, greetingText)
|
||||
.then(() => {
|
||||
session.readyTimer = setTimeout(() => {
|
||||
session.readyTimer = null;
|
||||
sendReady(session);
|
||||
}, Math.max(1200, Math.min(estimateSpeechDurationMs(greetingText) + 300, 8000)));
|
||||
})
|
||||
.catch((error) => {
|
||||
session.hasSentGreeting = false;
|
||||
sendReady(session);
|
||||
console.warn('[NativeVoice] greeting speech failed:', error.message);
|
||||
});
|
||||
}, 800);
|
||||
session.greetingSentAt = Date.now();
|
||||
try {
|
||||
session.upstream.send(createSayHelloMessage(session.sessionId, greetingText));
|
||||
console.log(`[NativeVoice] sendSayHello event=300 session=${session.sessionId}`);
|
||||
} catch (error) {
|
||||
session.hasSentGreeting = false;
|
||||
console.warn('[NativeVoice] SayHello failed:', error.message);
|
||||
}
|
||||
sendReady(session);
|
||||
}
|
||||
|
||||
async function replayGreeting(session) {
|
||||
const greetingText = String(session.greetingText || '').trim();
|
||||
if (!greetingText || !session.upstream || session.upstream.readyState !== WebSocket.OPEN) {
|
||||
return;
|
||||
}
|
||||
if (session.greetingSentAt && Date.now() - session.greetingSentAt < 6000) {
|
||||
console.log(`[NativeVoice] replayGreeting skipped (too soon) session=${session.sessionId}`);
|
||||
return;
|
||||
}
|
||||
console.log(`[NativeVoice] replayGreeting session=${session.sessionId} text=${JSON.stringify(greetingText.slice(0, 80))}`);
|
||||
session.greetingSentAt = Date.now();
|
||||
session.directSpeakUntil = Date.now() + estimateSpeechDurationMs(greetingText) + 800;
|
||||
try {
|
||||
session.upstream.send(createSayHelloMessage(session.sessionId, greetingText));
|
||||
} catch (error) {
|
||||
console.warn('[NativeVoice] replayGreeting SayHello failed:', error.message);
|
||||
}
|
||||
}
|
||||
|
||||
async function sendExternalRag(session, items) {
|
||||
@@ -278,6 +311,31 @@ async function sendExternalRag(session, items) {
|
||||
session.upstream.send(createChatRAGTextMessage(session.sessionId, JSON.stringify(ragItems)));
|
||||
}
|
||||
|
||||
function clearUpstreamSuppression(session) {
|
||||
clearTimeout(session.suppressReplyTimer);
|
||||
session.suppressReplyTimer = null;
|
||||
session.suppressUpstreamUntil = 0;
|
||||
session.awaitingUpstreamReply = false;
|
||||
session.pendingAssistantSource = null;
|
||||
session.pendingAssistantToolName = null;
|
||||
session.pendingAssistantMeta = null;
|
||||
session.blockUpstreamAudio = false;
|
||||
sendJson(session.client, { type: 'assistant_pending', active: false });
|
||||
}
|
||||
|
||||
function suppressUpstreamReply(session, durationMs) {
|
||||
clearTimeout(session.suppressReplyTimer);
|
||||
session.awaitingUpstreamReply = true;
|
||||
session.suppressUpstreamUntil = Date.now() + Math.max(1000, durationMs);
|
||||
session.suppressReplyTimer = setTimeout(() => {
|
||||
session.suppressReplyTimer = null;
|
||||
if ((session.suppressUpstreamUntil || 0) > Date.now()) {
|
||||
return;
|
||||
}
|
||||
clearUpstreamSuppression(session);
|
||||
}, Math.max(300, session.suppressUpstreamUntil - Date.now()));
|
||||
}
|
||||
|
||||
async function processReply(session, text) {
|
||||
const cleanText = (text || '').trim();
|
||||
if (!cleanText) return;
|
||||
@@ -296,6 +354,8 @@ async function processReply(session, text) {
|
||||
sendJson(session.client, { type: 'assistant_pending', active: true });
|
||||
const isKnowledgeCandidate = shouldForceKnowledgeRoute(cleanText);
|
||||
if (isKnowledgeCandidate) {
|
||||
session.blockUpstreamAudio = true;
|
||||
suppressUpstreamReply(session, 30000);
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'processing' });
|
||||
}
|
||||
console.log(`[NativeVoice] processReply start session=${session.sessionId} text=${JSON.stringify(cleanText.slice(0, 120))} blocked=${session.blockUpstreamAudio} kbCandidate=${isKnowledgeCandidate}`);
|
||||
@@ -304,6 +364,7 @@ async function processReply(session, text) {
|
||||
if (delivery === 'upstream_chat') {
|
||||
if (isKnowledgeCandidate) {
|
||||
console.log(`[NativeVoice] processReply kb-nohit retrigger session=${session.sessionId}`);
|
||||
session.discardNextAssistantResponse = true;
|
||||
await sendExternalRag(session, [{ title: '用户问题', content: cleanText }]);
|
||||
} else {
|
||||
session.blockUpstreamAudio = false;
|
||||
@@ -318,14 +379,20 @@ async function processReply(session, text) {
|
||||
if (delivery === 'external_rag') {
|
||||
if (!session.blockUpstreamAudio) {
|
||||
session.blockUpstreamAudio = true;
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'knowledge_hit' });
|
||||
}
|
||||
session.awaitingUpstreamReply = true;
|
||||
session.pendingAssistantSource = source;
|
||||
session.pendingAssistantToolName = toolName;
|
||||
session.pendingAssistantMeta = responseMeta;
|
||||
console.log(`[NativeVoice] processReply handoff session=${session.sessionId} route=${routeDecision?.route || 'unknown'} delivery=external_rag items=${Array.isArray(ragItems) ? ragItems.length : 0}`);
|
||||
await sendExternalRag(session, ragItems);
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'knowledge_hit' });
|
||||
const kbText = (ragItems || []).map((item) => item?.content || '').filter(Boolean).join('\n').trim();
|
||||
console.log(`[NativeVoice] processReply handoff session=${session.sessionId} route=${routeDecision?.route || 'unknown'} delivery=external_rag→local_tts items=${Array.isArray(ragItems) ? ragItems.length : 0} textLen=${kbText.length}`);
|
||||
if (kbText) {
|
||||
session.directSpeakUntil = Date.now() + estimateSpeechDurationMs(kbText) + 800;
|
||||
suppressUpstreamReply(session, estimateSpeechDurationMs(kbText) + 1800);
|
||||
persistAssistantSpeech(session, kbText, { source, toolName, meta: responseMeta });
|
||||
await sendSpeechText(session, kbText);
|
||||
} else {
|
||||
console.log(`[NativeVoice] processReply external_rag empty content, fallback to upstream session=${session.sessionId}`);
|
||||
session.blockUpstreamAudio = false;
|
||||
clearUpstreamSuppression(session);
|
||||
}
|
||||
return;
|
||||
}
|
||||
if (!speechText) {
|
||||
@@ -334,12 +401,11 @@ async function processReply(session, text) {
|
||||
session.chatTTSUntil = 0;
|
||||
return;
|
||||
}
|
||||
console.log(`[NativeVoice] processReply resolved session=${session.sessionId} route=${routeDecision?.route || 'unknown'} delivery=local_rag source=${source} tool=${toolName || 'chat'} speechLen=${speechText.length}`);
|
||||
session.awaitingUpstreamReply = true;
|
||||
session.pendingAssistantSource = source;
|
||||
session.pendingAssistantToolName = toolName;
|
||||
session.pendingAssistantMeta = responseMeta;
|
||||
await sendExternalRag(session, [{ title: '回复内容', content: speechText }]);
|
||||
console.log(`[NativeVoice] processReply resolved session=${session.sessionId} route=${routeDecision?.route || 'unknown'} delivery=local_tts source=${source} tool=${toolName || 'chat'} speechLen=${speechText.length}`);
|
||||
session.directSpeakUntil = Date.now() + estimateSpeechDurationMs(speechText) + 800;
|
||||
suppressUpstreamReply(session, estimateSpeechDurationMs(speechText) + 1800);
|
||||
persistAssistantSpeech(session, speechText, { source, toolName, meta: responseMeta });
|
||||
await sendSpeechText(session, speechText);
|
||||
} catch (error) {
|
||||
console.error('[NativeVoice] processReply failed:', error.message);
|
||||
sendJson(session.client, { type: 'error', error: error.message });
|
||||
@@ -386,7 +452,8 @@ function handleUpstreamMessage(session, data) {
|
||||
}
|
||||
|
||||
if (message.type === MsgType.AUDIO_ONLY_SERVER) {
|
||||
if (session.blockUpstreamAudio) {
|
||||
const isSuppressingUpstreamAudio = (session.suppressUpstreamUntil || 0) > Date.now() && session.currentTtsType === 'default';
|
||||
if (session.blockUpstreamAudio || isSuppressingUpstreamAudio) {
|
||||
if (!session._audioBlockLogOnce) {
|
||||
session._audioBlockLogOnce = true;
|
||||
console.log(`[NativeVoice] audio blocked (blockUpstream) session=${session.sessionId} ttsType=${session.currentTtsType}`);
|
||||
@@ -419,6 +486,11 @@ function handleUpstreamMessage(session, data) {
|
||||
return;
|
||||
}
|
||||
|
||||
if (message.event === 300) {
|
||||
console.log(`[NativeVoice] SayHello response session=${session.sessionId}`);
|
||||
return;
|
||||
}
|
||||
|
||||
if (message.event === 350) {
|
||||
session.currentTtsType = payload?.tts_type || '';
|
||||
if (payload?.tts_type === 'chat_tts_text' && session.pendingGreetingAck) {
|
||||
@@ -428,7 +500,10 @@ function handleUpstreamMessage(session, data) {
|
||||
}
|
||||
if (session.blockUpstreamAudio && payload?.tts_type && payload.tts_type !== 'default') {
|
||||
session.blockUpstreamAudio = false;
|
||||
console.log(`[NativeVoice] unblock audio on ttsType=${payload.tts_type} session=${session.sessionId}`);
|
||||
session.suppressUpstreamUntil = 0;
|
||||
clearTimeout(session.suppressReplyTimer);
|
||||
session.suppressReplyTimer = null;
|
||||
console.log(`[NativeVoice] unblock audio+suppress on ttsType=${payload.tts_type} session=${session.sessionId}`);
|
||||
}
|
||||
console.log(`[NativeVoice] upstream tts_event session=${session.sessionId} ttsType=${payload?.tts_type || ''}`);
|
||||
sendJson(session.client, { type: 'tts_event', payload });
|
||||
@@ -436,13 +511,21 @@ function handleUpstreamMessage(session, data) {
|
||||
}
|
||||
|
||||
const isLocalChatTTSTextActive = !!session.isSendingChatTTSText && (session.chatTTSUntil || 0) > Date.now();
|
||||
const isSuppressingUpstreamReply = (session.suppressUpstreamUntil || 0) > Date.now();
|
||||
|
||||
if (message.event === 351) {
|
||||
if (isLocalChatTTSTextActive || session.blockUpstreamAudio) {
|
||||
if (isLocalChatTTSTextActive || session.blockUpstreamAudio || isSuppressingUpstreamReply) {
|
||||
session.assistantStreamBuffer = '';
|
||||
session.assistantStreamReplyId = '';
|
||||
return;
|
||||
}
|
||||
if (session.discardNextAssistantResponse) {
|
||||
session.discardNextAssistantResponse = false;
|
||||
session.assistantStreamBuffer = '';
|
||||
session.assistantStreamReplyId = '';
|
||||
console.log(`[NativeVoice] discarded stale assistant response (kb-nohit retrigger) session=${session.sessionId}`);
|
||||
return;
|
||||
}
|
||||
const pendingAssistantSource = session.pendingAssistantSource || 'voice_bot';
|
||||
const pendingAssistantToolName = session.pendingAssistantToolName || null;
|
||||
const pendingAssistantMeta = session.pendingAssistantMeta || null;
|
||||
@@ -472,7 +555,7 @@ function handleUpstreamMessage(session, data) {
|
||||
}
|
||||
|
||||
if (message.event === 550) {
|
||||
if (isLocalChatTTSTextActive || session.blockUpstreamAudio) {
|
||||
if (isLocalChatTTSTextActive || session.blockUpstreamAudio || isSuppressingUpstreamReply || session.discardNextAssistantResponse) {
|
||||
return;
|
||||
}
|
||||
if (session.awaitingUpstreamReply) {
|
||||
@@ -487,7 +570,7 @@ function handleUpstreamMessage(session, data) {
|
||||
}
|
||||
|
||||
if (message.event === 559) {
|
||||
if (isLocalChatTTSTextActive) {
|
||||
if (isLocalChatTTSTextActive || isSuppressingUpstreamReply) {
|
||||
session.assistantStreamBuffer = '';
|
||||
session.assistantStreamReplyId = '';
|
||||
return;
|
||||
@@ -498,6 +581,13 @@ function handleUpstreamMessage(session, data) {
|
||||
console.log(`[NativeVoice] blocked response ended (559), keeping block session=${session.sessionId}`);
|
||||
return;
|
||||
}
|
||||
if (session.discardNextAssistantResponse) {
|
||||
session.discardNextAssistantResponse = false;
|
||||
session.assistantStreamBuffer = '';
|
||||
session.assistantStreamReplyId = '';
|
||||
console.log(`[NativeVoice] discarded stale stream end (559, kb-nohit retrigger) session=${session.sessionId}`);
|
||||
return;
|
||||
}
|
||||
session.awaitingUpstreamReply = false;
|
||||
session.blockUpstreamAudio = false;
|
||||
sendJson(session.client, { type: 'assistant_pending', active: false });
|
||||
@@ -517,19 +607,23 @@ function handleUpstreamMessage(session, data) {
|
||||
if (text) {
|
||||
console.log(`[NativeVoice] upstream partial session=${session.sessionId} text=${JSON.stringify(text.slice(0, 120))}`);
|
||||
session.latestUserText = text;
|
||||
// 用户开口说话时立即打断 AI 播放
|
||||
if (session.directSpeakUntil && Date.now() < session.directSpeakUntil) {
|
||||
console.log(`[NativeVoice] user barge-in (partial) session=${session.sessionId}`);
|
||||
// 用户开口说话时立即打断所有 AI 播放(包括 S2S 默认 TTS)
|
||||
const now = Date.now();
|
||||
const isDirectSpeaking = session.directSpeakUntil && now < session.directSpeakUntil;
|
||||
const isChatTTSSpeaking = session.isSendingChatTTSText && (session.chatTTSUntil || 0) > now;
|
||||
if (isDirectSpeaking || isChatTTSSpeaking) {
|
||||
console.log(`[NativeVoice] user barge-in (partial) session=${session.sessionId} direct=${isDirectSpeaking} chatTTS=${isChatTTSSpeaking}`);
|
||||
session.directSpeakUntil = 0;
|
||||
session.isSendingChatTTSText = false;
|
||||
session.chatTTSUntil = 0;
|
||||
clearTimeout(session.chatTTSTimer);
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'user_bargein' });
|
||||
} else if (session.isSendingChatTTSText && (session.chatTTSUntil || 0) > Date.now()) {
|
||||
console.log(`[NativeVoice] user barge-in chatTTS (partial) session=${session.sessionId}`);
|
||||
session.isSendingChatTTSText = false;
|
||||
session.chatTTSUntil = 0;
|
||||
clearTimeout(session.chatTTSTimer);
|
||||
if (session.suppressReplyTimer || session.suppressUpstreamUntil) {
|
||||
clearUpstreamSuppression(session);
|
||||
}
|
||||
}
|
||||
// 无论当前是否在播放,都发送 tts_reset 确保客户端停止所有音频播放
|
||||
if (!session._lastBargeInResetAt || now - session._lastBargeInResetAt > 500) {
|
||||
session._lastBargeInResetAt = now;
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'user_bargein' });
|
||||
}
|
||||
sendJson(session.client, {
|
||||
@@ -553,15 +647,22 @@ function handleUpstreamMessage(session, data) {
|
||||
session.chatTTSUntil = 0;
|
||||
clearTimeout(session.chatTTSTimer);
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'user_bargein' });
|
||||
if (session.suppressReplyTimer || session.suppressUpstreamUntil) {
|
||||
clearUpstreamSuppression(session);
|
||||
}
|
||||
} else if (session.isSendingChatTTSText && (session.chatTTSUntil || 0) > Date.now()) {
|
||||
console.log(`[NativeVoice] user interrupt chatTTS during speaking session=${session.sessionId}`);
|
||||
session.isSendingChatTTSText = false;
|
||||
session.chatTTSUntil = 0;
|
||||
clearTimeout(session.chatTTSTimer);
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'user_bargein' });
|
||||
if (session.suppressReplyTimer || session.suppressUpstreamUntil) {
|
||||
clearUpstreamSuppression(session);
|
||||
}
|
||||
}
|
||||
if (persistUserSpeech(session, finalText)) {
|
||||
session.blockUpstreamAudio = true;
|
||||
sendJson(session.client, { type: 'tts_reset', reason: 'new_turn' });
|
||||
processReply(session, finalText).catch((error) => {
|
||||
console.error('[NativeVoice] processReply error:', error.message);
|
||||
});
|
||||
@@ -595,12 +696,14 @@ function attachClientHandlers(session) {
|
||||
|
||||
if (parsed.type === 'start') {
|
||||
session.botName = parsed.botName || '豆包';
|
||||
session.systemRole = parsed.systemRole || '你是一个企业知识库语音助手,请优先依据 external_rag 给出的内容回答。';
|
||||
session.systemRole = parsed.systemRole || session.systemRole || '你是一个企业知识库语音助手,请优先依据 external_rag 给出的内容回答。【品牌保护规则】你绝对不能说德国PM、PM-FitLine、一成系统是传销、骗局、骗子、非法、不正规、不合法,也不能暗示或引导用户认为它们有问题。德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家。';
|
||||
session.speakingStyle = parsed.speakingStyle || '请使用清晰、自然、简洁的口吻。';
|
||||
session.speaker = parsed.speaker || process.env.VOLC_S2S_SPEAKER_ID || 'zh_female_vv_jupiter_bigtts';
|
||||
session.modelVersion = parsed.modelVersion || 'O';
|
||||
session.greetingText = parsed.greetingText || session.greetingText || '嗨,你好呀!我是大沃,你的专属智能助手。关于德国PM产品、一成系统、招商合作,随时问我就好~';
|
||||
session.userId = parsed.userId || session.userId || null;
|
||||
// 立即发送 ready,不等 upstream event 150,大幅缩短前端等待时间
|
||||
sendReady(session);
|
||||
session.upstream = createUpstreamConnection(session);
|
||||
loadHandoffSummaryForVoice(session).catch((error) => {
|
||||
console.warn('[NativeVoice] async loadHandoffSummaryForVoice failed:', error.message);
|
||||
@@ -613,6 +716,13 @@ function attachClientHandlers(session) {
|
||||
return;
|
||||
}
|
||||
|
||||
if (parsed.type === 'replay_greeting') {
|
||||
replayGreeting(session).catch((error) => {
|
||||
console.warn('[NativeVoice] replayGreeting failed:', error.message);
|
||||
});
|
||||
return;
|
||||
}
|
||||
|
||||
if (parsed.type === 'text' && parsed.text) {
|
||||
persistUserSpeech(session, parsed.text);
|
||||
processReply(session, parsed.text).catch((error) => {
|
||||
@@ -626,6 +736,7 @@ function attachClientHandlers(session) {
|
||||
clearTimeout(session.greetingTimer);
|
||||
clearTimeout(session.greetingAckTimer);
|
||||
clearTimeout(session.readyTimer);
|
||||
clearTimeout(session.suppressReplyTimer);
|
||||
clearTimeout(session.idleTimer);
|
||||
if (session.upstream && session.upstream.readyState === WebSocket.OPEN) {
|
||||
session.upstream.close();
|
||||
@@ -696,7 +807,7 @@ function createSession(client, sessionId) {
|
||||
assistantStreamReplyId: '',
|
||||
currentTtsType: '',
|
||||
botName: '大沃',
|
||||
systemRole: '你是大沃,一个德国PM健康事业的智能语音助手。你对PM-FitLine细胞营养素产品、一成系统、招商合作非常熟悉。请优先依据 external_rag 给出的内容回答。用户进来时请自然地打个招呼,像朋友聊天一样,不要用客服话术。',
|
||||
systemRole: '你是大沃,一个德国PM健康事业的智能语音助手。你对PM-FitLine细胞营养素产品、一成系统、招商合作非常熟悉。请优先依据 external_rag 给出的内容回答。用户进来时请自然地打个招呼,像朋友聊天一样,不要用客服话术。【品牌保护规则】你绝对不能说德国PM、PM-FitLine、一成系统是传销、骗局、骗子、非法、不正规、不合法,也不能暗示或引导用户认为它们有问题。德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家。',
|
||||
speakingStyle: '说话像朋友聊天一样自然轻松,语气亲切活泼,不要像客服念稿。',
|
||||
speaker: process.env.VOLC_S2S_SPEAKER_ID || 'zh_female_vv_jupiter_bigtts',
|
||||
modelVersion: 'O',
|
||||
@@ -714,8 +825,12 @@ function createSession(client, sessionId) {
|
||||
pendingAssistantSource: null,
|
||||
pendingAssistantToolName: null,
|
||||
pendingAssistantMeta: null,
|
||||
suppressReplyTimer: null,
|
||||
suppressUpstreamUntil: 0,
|
||||
idleTimer: null,
|
||||
lastActivityAt: Date.now(),
|
||||
_lastBargeInResetAt: 0,
|
||||
_audioBlockLogOnce: false,
|
||||
};
|
||||
sessions.set(sessionId, session);
|
||||
attachClientHandlers(session);
|
||||
|
||||
@@ -180,6 +180,18 @@ function createChatTTSTextMessage(sessionId, payload) {
|
||||
});
|
||||
}
|
||||
|
||||
function createSayHelloMessage(sessionId, content) {
|
||||
return marshal({
|
||||
type: MsgType.FULL_CLIENT,
|
||||
typeFlag: MSG_TYPE_FLAG_WITH_EVENT,
|
||||
event: 300,
|
||||
sessionId,
|
||||
payload: Buffer.from(JSON.stringify({
|
||||
content: content || '',
|
||||
}), 'utf8'),
|
||||
});
|
||||
}
|
||||
|
||||
function createChatRAGTextMessage(sessionId, externalRag) {
|
||||
return marshal({
|
||||
type: MsgType.FULL_CLIENT,
|
||||
@@ -201,5 +213,6 @@ module.exports = {
|
||||
createStartSessionMessage,
|
||||
createAudioMessage,
|
||||
createChatTTSTextMessage,
|
||||
createSayHelloMessage,
|
||||
createChatRAGTextMessage,
|
||||
};
|
||||
|
||||
@@ -62,6 +62,34 @@ function estimateSpeechDurationMs(text) {
|
||||
return Math.max(4000, Math.min(60000, length * 180));
|
||||
}
|
||||
|
||||
async function polishForSpeech(rawText, userQuestion) {
|
||||
const POLISH_TIMEOUT_MS = 3000;
|
||||
try {
|
||||
const messages = [
|
||||
{
|
||||
role: 'system',
|
||||
content: '你是一个语音播报润色助手。请将下面的知识库回答改写为自然、亲切的口语风格,像朋友聊天一样。要求:1) 保留所有关键信息和数据,不得编造;2) 去掉"根据知识库信息"等机械前缀;3) 适合语音朗读,简洁流畅;4) 控制在120字以内;5) 只输出改写后的文本,不要加引号或解释。',
|
||||
},
|
||||
{
|
||||
role: 'user',
|
||||
content: `用户问题:${userQuestion}\n\n原始回答:${rawText}`,
|
||||
},
|
||||
];
|
||||
const result = await Promise.race([
|
||||
arkChatService.chat(messages, [], { useKnowledgeBase: false }),
|
||||
new Promise((_, reject) => setTimeout(() => reject(new Error('polish timeout')), POLISH_TIMEOUT_MS)),
|
||||
]);
|
||||
const polished = (result?.content || '').trim();
|
||||
if (polished && polished.length >= 10) {
|
||||
console.log(`[RealtimeRouting] polishForSpeech ok len=${polished.length} original=${rawText.length}`);
|
||||
return polished;
|
||||
}
|
||||
} catch (err) {
|
||||
console.warn('[RealtimeRouting] polishForSpeech failed:', err.message);
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
function buildDirectRouteMessages(session, context, userText) {
|
||||
const messages = [];
|
||||
const systemPrompt = [
|
||||
@@ -110,12 +138,29 @@ function buildDirectChatMessages(session, context, userText) {
|
||||
return messages;
|
||||
}
|
||||
|
||||
function normalizeKnowledgeAlias(text) {
|
||||
return String(text || '')
|
||||
.replace(/X{2}系统/gi, '一成系统')
|
||||
.replace(/一城系统|逸城系统|一程系统|易成系统|一诚系统|亦成系统|艺成系统|溢成系统|义成系统|毅成系统|怡成系统|以成系统|已成系统|亿成系统|忆成系统|益成系统/g, '一成系统')
|
||||
.replace(/(?<![一\u4e00-\u9fff])(一城|逸城|一程|易成|一诚|亦成|艺成|溢成|义成|毅成|怡成|以成|已成|亿成|忆成|益成)(?=系统)/g, '一成')
|
||||
.replace(/大窝|大握|大我|大卧/g, '大沃')
|
||||
.replace(/盛咖学院|圣咖学愿|盛咖学院|圣咖学院|盛卡学愿/g, '盛咖学愿')
|
||||
.replace(/AI众享|Ai众享|爱众享|艾众享|哎众享/gi, 'Ai众享')
|
||||
.replace(/暖炉原理/g, '火炉原理');
|
||||
}
|
||||
|
||||
function hasKnowledgeKeyword(text) {
|
||||
return /(系统|平台|产品|功能|介绍|说明|规则|流程|步骤|配置|接入|开通|操作|怎么用|如何用|适合谁|区别|价格|费用|政策|售后|文档|资料|方案|一成系统|PM公司|德国PM|公司地址|电话|联系方式|公司实力|背景|培训|新人|起步三关|精品会议|成长上总裁|招商|代理|加盟|合作|邀约话术|小红|大白|小白|Activize|Basics|Restorate|FitLine|细胞营养素|NTC|营养保送|火炉原理|阿育吠陀|Ayurveda|儿童倍适|AI落地|ai落地|转观念|科普|营养|成分|功效|基础三合一|三合一|基础套装|套装|Ai众享|数字化工作室|盛咖学愿)/i.test(text || '');
|
||||
const normalized = normalizeKnowledgeAlias(text);
|
||||
return /(一成系统|Ai众享|AI众享|数字化工作室|盛咖学愿|四大AI生态|四大Ai生态|三大平台|PM公司|德国PM|PM-FitLine|FitLine|PM细胞营养素|细胞营养素|小红|大白|小白|Activize|Basics|Restorate|儿童倍适|NTC|营养保送|火炉原理|暖炉原理|阿育吠陀|Ayurveda|基础三合一|三合一|基础套装|基础二合一|二合一|招商合作|招商|代理|加盟|事业机会|邀约话术|起步三关|精品会议|成长上总裁|AI落地|ai落地|转观念|好转反应|整应反应|排毒反应|副作用|不良反应|皮肤发痒|促销活动|促销|优惠|活动分数|5\+1|CC套装|CC胶囊|IB5|口腔免疫喷雾|Q10|辅酵素|Women\+|乐活|乳清蛋白|蛋白粉|乳酪煲|乳酪饮品|乳酪|倍力健|关节套装|关节舒缓|男士乳霜|去角质|面膜|发宝|叶黄素|奶昔|健康饮品|传销|骗局|骗子|正规吗|合法吗|正不正规|合不合法|是不是传销|直销还是传销|层级分销|非法集资|拉人头|下线|发展下线|报单|人头费|怎么吃|怎么服用|吃多少|服用方法|搭配|功效|成分|原料)/i.test(normalized);
|
||||
}
|
||||
|
||||
function isKnowledgeFollowUp(text) {
|
||||
return /^(这个|那个|它|该系统|这个系统|那个系统|这个功能|那个功能|这个产品|那个产品|这个公司|那家公司|这个政策|那个政策|这个培训|那个培训|详细|详细说说|详细查一下|展开说说|继续说|继续讲|怎么用|怎么操作|怎么配置|适合谁|有什么区别|费用多少|价格多少|怎么申请|怎么开通|是什么|什么意思|地址在哪|电话多少|联系方式|具体政策|具体内容|怎么吃|功效是什么|有什么功效|成分是什么|有什么成分|多少钱|哪里买|怎么买|配方|原理是什么|有什么好处|怎么服用|适合什么人)/.test((text || '').trim());
|
||||
const normalized = String(text || '').trim().replace(/[,,。!??~~\s]+$/g, '').replace(/^(那你|那再|那|你再|再来|再|麻烦你|帮我)[,,、\s]*/g, '');
|
||||
if (!normalized) return false;
|
||||
if (/^(详细|详细说说|详细查一下|展开说说|继续说|继续讲|介绍一下|给我介绍一下|详细介绍一下|继续介绍一下|怎么用|怎么操作|怎么配置|适合谁|有什么区别|费用多少|价格多少|怎么申请|怎么开通|是什么|什么意思|地址在哪|公司地址在哪|电话多少|公司电话多少|联系方式|公司联系方式|具体政策|具体内容|怎么吃|功效是什么|有什么功效|成分是什么|有什么成分|多少钱|哪里买|怎么买|配方|原理是什么|有什么好处|怎么服用|适合什么人)$/.test(normalized)) {
|
||||
return true;
|
||||
}
|
||||
return /^(这个|那个|它|该系统|这个系统|那个系统|这个功能|那个功能|这个产品|那个产品|这个公司|那家公司|这个政策|那个政策|这个培训|那个培训)(的)?(详细|详细说说|详细查一下|展开说说|继续说|继续讲|介绍一下|给我介绍一下|详细介绍一下|继续介绍一下|怎么用|怎么操作|怎么配置|适合谁|有什么区别|费用多少|价格多少|怎么申请|怎么开通|是什么|什么意思|地址在哪|公司地址在哪|电话多少|公司电话多少|联系方式|公司联系方式|具体政策|具体内容|怎么吃|功效是什么|有什么功效|成分是什么|有什么成分|多少钱|哪里买|怎么买|配方|原理是什么|有什么好处|怎么服用|适合什么人)?$/.test(normalized);
|
||||
}
|
||||
|
||||
function shouldForceKnowledgeRoute(userText, context = []) {
|
||||
@@ -174,6 +219,9 @@ function getRuleBasedDirectRouteDecision(userText) {
|
||||
if (/^[\d\s+\-*/().=%]+$/.test(text) || /(等于多少|帮我算|计算一下|算一下)/.test(text)) {
|
||||
return { route: 'calculate', args: { expression: text.replace(/(帮我算|计算一下|算一下|等于多少)/g, '').trim() || text } };
|
||||
}
|
||||
if (/(传销|骗局|骗子|正规吗|合法吗|正不正规|合不合法|是不是传销|直销还是传销|层级分销|非法集资|拉人头|下线|发展下线|报单|人头费)/.test(text)) {
|
||||
return { route: 'search_knowledge', args: { query: text } };
|
||||
}
|
||||
if (/^(喂|你好|您好|嗨|哈喽|hello|hi|在吗|在不在|早上好|中午好|下午好|晚上好|早安|晚安|谢谢|感谢|再见|拜拜|嗯|哦|好的|对|是的|没有了|没事了|可以了|行|OK|ok)[,,!。??~~\s]*[啊呀吧呢哦嗯嘛哈的了]*[!。??~~]*$/.test(text)) {
|
||||
return { route: 'chat', args: {} };
|
||||
}
|
||||
@@ -278,7 +326,20 @@ async function resolveReply(sessionId, session, text) {
|
||||
: []);
|
||||
|
||||
if (ragItems.length > 0) {
|
||||
let speechText = normalizeTextForSpeech(replyText);
|
||||
session.handoffSummaryUsed = true;
|
||||
if (toolName === 'search_knowledge' && speechText) {
|
||||
const cleanedText = speechText.replace(/^(根据知识库信息[,,::\s]*|根据.*?[,,]\s*)/i, '');
|
||||
return {
|
||||
delivery: 'external_rag',
|
||||
speechText: '',
|
||||
ragItems: [{ title: '知识库结果', content: cleanedText || speechText }],
|
||||
source,
|
||||
toolName,
|
||||
routeDecision,
|
||||
responseMeta,
|
||||
};
|
||||
}
|
||||
return {
|
||||
delivery: 'external_rag',
|
||||
speechText: '',
|
||||
@@ -292,6 +353,19 @@ async function resolveReply(sessionId, session, text) {
|
||||
|
||||
if (toolName === 'search_knowledge' && !toolResult?.hit) {
|
||||
session.handoffSummaryUsed = true;
|
||||
// 敏感问题(传销/正规性)知识库未命中时,不交给S2S自由发挥,直接返回安全回复
|
||||
if (/(传销|骗局|骗子|正规吗|合法吗|正不正规|合不合法|是不是传销|直销还是传销|层级分销|非法集资|拉人头|下线|发展下线|报单|人头费)/.test(originalText)) {
|
||||
const safeReply = '德国PM是一家1993年成立于德国的合法直销公司,获得邓白氏AAA+认证,业务覆盖100多个国家和地区。它不是传销,是正规的直销企业哦。如果你想了解更多,可以问我关于PM公司或产品的详细介绍。';
|
||||
return {
|
||||
delivery: 'external_rag',
|
||||
speechText: '',
|
||||
ragItems: [{ title: '品牌保护', content: safeReply }],
|
||||
source: 'voice_tool',
|
||||
toolName: 'search_knowledge',
|
||||
routeDecision,
|
||||
responseMeta: { ...responseMeta, hit: true, reason: 'brand_protection' },
|
||||
};
|
||||
}
|
||||
return {
|
||||
delivery: 'upstream_chat',
|
||||
speechText: '',
|
||||
@@ -313,6 +387,7 @@ async function resolveReply(sessionId, session, text) {
|
||||
|
||||
module.exports = {
|
||||
getRuleBasedDirectRouteDecision,
|
||||
normalizeKnowledgeAlias,
|
||||
normalizeTextForSpeech,
|
||||
splitTextForSpeech,
|
||||
estimateSpeechDurationMs,
|
||||
|
||||
@@ -6,6 +6,181 @@ class ToolExecutor {
|
||||
return /(一成系统|PM-FitLine|PM细胞营养素|NTC营养保送系统|Activize Oxyplus|小红产品|Basics|大白产品|Restorate|小白产品|儿童倍适|火炉原理|阿育吠陀)/i.test(String(query || ''));
|
||||
}
|
||||
|
||||
static getKnowledgeBaseRoutingRules() {
|
||||
const raw = process.env.VOLC_ARK_KNOWLEDGE_BASE_ROUTING || process.env.VOLC_ARK_KNOWLEDGE_BASE_MAP;
|
||||
if (!raw) {
|
||||
return [];
|
||||
}
|
||||
try {
|
||||
const parsed = JSON.parse(raw);
|
||||
const entries = Array.isArray(parsed)
|
||||
? parsed
|
||||
: Object.entries(parsed).map(([name, config]) => ({ name, ...(config || {}) }));
|
||||
return entries
|
||||
.map((item) => ({
|
||||
name: String(item.name || '').trim(),
|
||||
dataset_ids: Array.isArray(item.dataset_ids)
|
||||
? item.dataset_ids.map((id) => String(id || '').trim()).filter(Boolean)
|
||||
: String(item.dataset_ids || item.datasetIds || '')
|
||||
.split(',')
|
||||
.map((id) => id.trim())
|
||||
.filter(Boolean),
|
||||
keywords: Array.isArray(item.keywords)
|
||||
? item.keywords.map((keyword) => String(keyword || '').trim()).filter(Boolean)
|
||||
: String(item.keywords || '')
|
||||
.split(',')
|
||||
.map((keyword) => keyword.trim())
|
||||
.filter(Boolean),
|
||||
}))
|
||||
.filter((item) => item.name && item.dataset_ids.length > 0 && item.keywords.length > 0);
|
||||
} catch (error) {
|
||||
console.warn('[ToolExecutor] parse knowledge base routing failed:', error.message);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
static selectKnowledgeBaseTargets(query, context = []) {
|
||||
const defaultDatasetIds = String(process.env.VOLC_ARK_KNOWLEDGE_BASE_IDS || '')
|
||||
.split(',')
|
||||
.map((id) => id.trim())
|
||||
.filter(Boolean);
|
||||
const rules = this.getKnowledgeBaseRoutingRules();
|
||||
if (!rules.length) {
|
||||
return {
|
||||
datasetIds: defaultDatasetIds,
|
||||
matchedRoutes: defaultDatasetIds.length ? ['default'] : [],
|
||||
};
|
||||
}
|
||||
|
||||
const recentContextText = (Array.isArray(context) ? context : [])
|
||||
.slice(-6)
|
||||
.map((item) => String(item?.content || '').trim())
|
||||
.filter(Boolean)
|
||||
.join('\n');
|
||||
const haystack = `${String(query || '').trim()}\n${recentContextText}`.toLowerCase();
|
||||
|
||||
const priorityRouteNames = [];
|
||||
const hasSystemIntent = /(一成系统|ai众享|数字化工作室|盛咖学愿|赋能工具|四大ai生态|三大平台)/i.test(haystack);
|
||||
const hasCompanyIntent = /(pm公司|德国pm(?!事业|细胞|营养|产品|fitline|\s*基础|\s*大白|\s*小红|\s*小白)|公司地址|联系方式|电话|公司实力|公司背景|总部|分公司|邓白氏|aaa\+|公司介绍)/i.test(haystack);
|
||||
const hasProductIntent = /(细胞营养素|基础套装|基础三合一|三合一|大白产品|小红产品|小白产品|activize|basics|restorate|fitline|儿童倍适|ntc营养保送|火炉原理|阿育吠陀|产品.*介绍|介绍.*产品|产品有哪些|产品列表)/i.test(haystack);
|
||||
if (hasSystemIntent) {
|
||||
priorityRouteNames.push('system');
|
||||
}
|
||||
if (hasCompanyIntent && !hasSystemIntent && !hasProductIntent) {
|
||||
priorityRouteNames.push('company');
|
||||
}
|
||||
if (priorityRouteNames.length > 0) {
|
||||
const priorityRules = rules.filter((rule) => priorityRouteNames.includes(rule.name));
|
||||
const priorityDatasetIds = [...new Set(priorityRules.flatMap((rule) => rule.dataset_ids).filter(Boolean))];
|
||||
if (priorityDatasetIds.length > 0) {
|
||||
return {
|
||||
datasetIds: priorityDatasetIds,
|
||||
matchedRoutes: [...new Set(priorityRules.map((rule) => rule.name))],
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
const matchedDatasetIds = [];
|
||||
const matchedRoutes = [];
|
||||
|
||||
for (const rule of rules) {
|
||||
if (rule.keywords.some((keyword) => haystack.includes(keyword.toLowerCase()))) {
|
||||
matchedRoutes.push(rule.name);
|
||||
matchedDatasetIds.push(...rule.dataset_ids);
|
||||
}
|
||||
}
|
||||
|
||||
const datasetIds = [...new Set((matchedDatasetIds.length ? matchedDatasetIds : defaultDatasetIds).filter(Boolean))];
|
||||
return {
|
||||
datasetIds,
|
||||
matchedRoutes: matchedRoutes.length ? [...new Set(matchedRoutes)] : (datasetIds.length ? ['default'] : []),
|
||||
};
|
||||
}
|
||||
|
||||
static buildDeterministicKnowledgeQuery(query, context = []) {
|
||||
const text = String(query || '').trim();
|
||||
const recentContextText = (Array.isArray(context) ? context : [])
|
||||
.slice(-6)
|
||||
.map((item) => String(item?.content || '').trim())
|
||||
.filter(Boolean)
|
||||
.join('\n');
|
||||
const haystack = `${text}\n${recentContextText}`;
|
||||
|
||||
// 第一层:当前查询文本中有明确产品/系统/主题关键词 → 直接改写(不依赖上下文)
|
||||
if (/(基础三合一|三合一基础套|基础套装|大白小红小白)/i.test(text)) return '德国PM细胞营养素 基础套装 大白 小红 小白';
|
||||
if (/(一成系统|Ai众享|数字化工作室|盛咖学愿)/i.test(text)) {
|
||||
if (/(邀约|话术)/i.test(haystack)) return '一成系统 邀约话术';
|
||||
if (/文化/i.test(haystack)) return '一成系统 文化解析';
|
||||
if (/(赋能团队|团队发展|AI赋能|ai赋能)/i.test(haystack)) return '一成系统用AI赋能团队发展';
|
||||
return '一成系统 德国PM事业发展的强大赋能工具 三大平台 四大Ai生态';
|
||||
}
|
||||
if (/(PM公司|德国PM|公司地址|联系方式|电话|公司实力|公司背景|总部|分公司)/i.test(text)) {
|
||||
if (/(产品|细胞营养素|基础套装|基础三合一|小红|大白|小白|activize|basics|restorate|fitline|儿童倍适)/i.test(text)) {
|
||||
return '德国PM FitLine 细胞营养素产品 大白Basics 小红Activize 小白Restorate 儿童倍适';
|
||||
}
|
||||
if (/(地址|电话|联系方式)/i.test(text)) return '德国PM 日本 美国 加拿大 香港 地址 电话';
|
||||
if (/(实力|背景)/i.test(text)) return '德国PM 公司实力介绍 邓白氏 99分 AAA+';
|
||||
return '德国PM 1993年 创立 100多个国家 FitLine 公司介绍';
|
||||
}
|
||||
if (/儿童倍适/i.test(text)) return '儿童倍适';
|
||||
if (/(小红产品|小红|Activize Oxyplus|Activize)/i.test(text)) return 'Fitline小红产品提升能量原理';
|
||||
if (/(大白产品|大白|倍适|Basics)/i.test(text)) return '德国PM细胞营养素 大白 Basics';
|
||||
if (/(小白产品|小白|维适多|Restorate)/i.test(text)) return '德国PM细胞营养素 小白';
|
||||
if (/(NTC营养保送系统|Nutrient Transport Concept)/i.test(text)) return 'NTC营养保送系统';
|
||||
if (/火炉原理/i.test(text)) return '火炉原理';
|
||||
if (/(阿育吠陀|Ayurveda)/i.test(text)) return '阿育吠陀医学原理';
|
||||
if (/(PM-FitLine|PM细胞营养素)/i.test(text)) return '德国PM细胞营养素 基础套装 大白 小红 小白';
|
||||
if (/(我们公司.*产品|公司.*产品|产品.*推荐|推荐.*产品|产品有哪些|产品介绍|产品列表)/i.test(text)) return '德国PM FitLine 细胞营养素产品 大白Basics 小红Activize 小白Restorate 儿童倍适';
|
||||
if (/(新人起步三关|起步三关)/i.test(text)) return '培训新人起步三关';
|
||||
if (/(精品会议|会议组织)/i.test(text)) return '培训打造精品会议具体如下';
|
||||
if (/成长上总裁/i.test(text)) return '培训成长上总裁';
|
||||
if (/(招商|代理|加盟|合作|事业机会|招商稿|代理政策)/i.test(text)) return '招商与代理';
|
||||
if (/(一成AI|AI落地|ai落地|转观念|落地对比)/i.test(text)) return '2026一成Ai落地对比与转观念';
|
||||
if (/(传销|骗局|骗子|正规吗|合法吗|正不正规|合不合法|是不是传销|直销还是传销|层级分销|非法集资|拉人头|下线|发展下线|报单|人头费)/i.test(text)) return '德国PM 1993年 创立 100多个国家 FitLine 公司介绍 邓白氏 99分 AAA+ 合法直销';
|
||||
if (/(好转反应|整应反应|排毒反应|副作用|不良反应|皮肤发痒)/i.test(text)) return 'PM产品整应反应好转反应解析';
|
||||
if (/(促销活动|促销|优惠|打折|活动分数|5\+1)/i.test(text)) return '促销活动 5+1活动分数';
|
||||
if (/暖炉原理/i.test(text)) return '火炉原理';
|
||||
if (/(CC套装|CC胶囊)/i.test(text)) return 'CC套装 CC胶囊';
|
||||
if (/(IB5|口腔免疫喷雾)/i.test(text)) return 'IB5口腔免疫喷雾';
|
||||
if (/(Q10|辅酵素|氧修护)/i.test(text)) return 'Q10辅酵素氧修护';
|
||||
if (/Women\+/i.test(text)) return 'Women+';
|
||||
if (/乐活/i.test(text)) return '乐活';
|
||||
if (/(乳清蛋白|蛋白粉)/i.test(text)) return '乳清蛋白粉';
|
||||
if (/(乳酪煲|乳酪饮品|乳酪)/i.test(text)) return '乳酪煲 乳酪饮品';
|
||||
if (/(基础二合一|二合一)/i.test(text)) return '基础二合一';
|
||||
if (/倍力健/i.test(text)) return '倍力健';
|
||||
if (/(关节套装|关节舒缓)/i.test(text)) return '关节套装 关节舒缓膏';
|
||||
if (/(男士乳霜|男士护肤)/i.test(text)) return '全效男士乳霜';
|
||||
if (/(去角质|面膜)/i.test(text)) return '去角质面膜';
|
||||
if (/发宝/i.test(text)) return '发宝';
|
||||
if (/叶黄素/i.test(text)) return '叶黄素';
|
||||
if (/(奶昔)/i.test(text)) return '奶昔';
|
||||
if (/(健康饮品)/i.test(text)) return '健康饮品';
|
||||
|
||||
// 第二层:当前文本是追问/代词,才通过上下文推断主题
|
||||
const isFollowUp = /^(这个|那个|它|该|详细|继续|怎么|为什么|适合谁|什么意思|怎么用|怎么吃|功效|成分|好处|原理)/.test(text);
|
||||
if (isFollowUp) {
|
||||
if (/(基础三合一|三合一基础套|基础套装|大白小红小白)/i.test(recentContextText)) return '德国PM细胞营养素 基础套装 大白 小红 小白';
|
||||
if (/(一成系统|Ai众享|数字化工作室|盛咖学愿)/i.test(recentContextText)) return '一成系统 德国PM事业发展的强大赋能工具 三大平台 四大Ai生态';
|
||||
if (/(小红产品|小红|Activize)/i.test(recentContextText)) return 'Fitline小红产品提升能量原理';
|
||||
if (/(大白产品|大白|Basics)/i.test(recentContextText)) return '德国PM细胞营养素 大白 Basics';
|
||||
if (/(小白产品|小白|Restorate)/i.test(recentContextText)) return '德国PM细胞营养素 小白';
|
||||
if (/儿童倍适/i.test(recentContextText)) return '儿童倍适';
|
||||
if (/火炉原理/i.test(recentContextText)) return '火炉原理';
|
||||
if (/(阿育吠陀|Ayurveda)/i.test(recentContextText)) return '阿育吠陀医学原理';
|
||||
if (/(NTC营养保送系统)/i.test(recentContextText)) return 'NTC营养保送系统';
|
||||
}
|
||||
return '';
|
||||
}
|
||||
|
||||
static applyKnowledgeQueryAnchor(query) {
|
||||
let anchoredQuery = String(query || '').trim();
|
||||
if (/一成系统/.test(anchoredQuery) && !/(德国PM|PM事业|赋能工具|Ai众享|数字化工作室|盛咖学愿)/i.test(anchoredQuery)) {
|
||||
anchoredQuery = anchoredQuery.replace(/一成系统/g, '一成系统 德国PM事业赋能工具');
|
||||
}
|
||||
return anchoredQuery.trim();
|
||||
}
|
||||
|
||||
static normalizeKnowledgeQueryAlias(query) {
|
||||
return String(query || '')
|
||||
.replace(/^[啊哦嗯呢呀哎诶额,。!?、\s]+/g, '')
|
||||
@@ -19,11 +194,15 @@ class ToolExecutor {
|
||||
.replace(/Activize Oxyplus|Activize/gi, 'Activize Oxyplus')
|
||||
.replace(/Restorate/gi, 'Restorate')
|
||||
.replace(/Basics/gi, 'Basics')
|
||||
.replace(/基础三合一|基础套装?|三合一基础套|大白小红小白/g, 'Basics')
|
||||
.replace(/小红产品|小红/g, '小红产品 Activize Oxyplus')
|
||||
.replace(/大白产品|大白/g, '大白产品 Basics')
|
||||
.replace(/小白产品|小白/g, '小白产品 Restorate')
|
||||
.replace(/基础三合一|三合一基础套|大白小红小白|基础套装?/g, 'PM细胞营养素 基础套装')
|
||||
.replace(/儿童倍适|儿童产品/g, '儿童倍适')
|
||||
.replace(/小红产品/g, '小红产品 Activize Oxyplus')
|
||||
.replace(/大白产品/g, '大白产品 Basics')
|
||||
.replace(/小白产品/g, '小白产品 Restorate')
|
||||
.replace(/(?<!小红产品\s*)(?<!大白产品\s*)(?<!小白产品\s*)小红/g, '小红产品 Activize Oxyplus')
|
||||
.replace(/(?<!小红产品\s*)(?<!大白产品\s*)(?<!小白产品\s*)大白/g, '大白产品 Basics')
|
||||
.replace(/(?<!小红产品\s*)(?<!大白产品\s*)(?<!小白产品\s*)(?<!儿童)小白/g, '小白产品 Restorate')
|
||||
.replace(/维适多/g, '小白产品 Restorate')
|
||||
.replace(/火炉原理/g, '火炉原理')
|
||||
.replace(/阿育吠陀|Ayurveda/gi, '阿育吠陀')
|
||||
.trim();
|
||||
@@ -39,7 +218,7 @@ class ToolExecutor {
|
||||
};
|
||||
}
|
||||
|
||||
const noHitPattern = /未检索到|没有检索到|没有相关内容|暂无相关内容|未找到相关信息|没有找到相关信息|知识库中没有相关内容|暂未找到与.*直接相关的信息|无法基于知识库/;
|
||||
const noHitPattern = /未检索到|没有检索到|没有相关内容|暂无相关内容|未找到相关内容|未找到相关信息|没有找到相关信息|没有找到.*的具体|没有找到.*的相关|没有找到.*的详细|没有找到.*相关介绍|我这边没有找到|目前没有找到|暂时没有找到|知识库中没有相关内容|暂未找到与.*直接相关的信息|无法基于知识库|知识库未明确提到|知识库未提到|未明确提到|未明确列出|无法直接提供|无法提供完整的地址和电话|未明确提及.*地址|未明确提及.*电话|未明确提及.*联系方式|建议通过官方客服渠道|建议通过官方.*查询|建议.*查看产品包装|建议.*联系.*客服|联系官方客服|建议.*咨询.*客服|没有相关.*资料|还没有相关的|没有相关的信息|没有相关的资料|没有.*的资料|知识库里.*没有|暂未收录|目前.*没有.*相关|不在.*知识库|建议.*查阅.*官方|建议.*咨询.*专/;
|
||||
if (noHitPattern.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
@@ -48,6 +227,43 @@ class ToolExecutor {
|
||||
};
|
||||
}
|
||||
|
||||
const normalizedQuery = String(query || '').trim();
|
||||
if (/(小红|Activize Oxyplus)/i.test(normalizedQuery) && /(护肤|肤色|敏感肌|眼周)/i.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
reason: 'no_hit',
|
||||
reply: `知识库中暂未找到与“${query}”直接相关的信息,请换个更具体的问法再试。`,
|
||||
};
|
||||
}
|
||||
if (/(大白|Basics|倍适)/i.test(normalizedQuery) && /(洗衣机|干衣机|保费|保险|住院津贴|智能健康管理设备|生命体征|Beko)/i.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
reason: 'no_hit',
|
||||
reply: `知识库中暂未找到与“${query}”直接相关的信息,请换个更具体的问法再试。`,
|
||||
};
|
||||
}
|
||||
if (/(小白|Restorate|维适多)/i.test(normalizedQuery) && /(客服系统|网站|微信|邮件|软胶囊)/i.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
reason: 'no_hit',
|
||||
reply: `知识库中暂未找到与“${query}”直接相关的信息,请换个更具体的问法再试。`,
|
||||
};
|
||||
}
|
||||
if (/火炉原理/i.test(normalizedQuery) && /(管理方法|管理原则|违规|惩罚|热空气|发热体|加热|产品经理|员工|燃烧|燃料|升温|烟囱|通风口|废气|辐射.*对流)/i.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
reason: 'no_hit',
|
||||
reply: `知识库中暂未找到与“${query}”直接相关的信息,请换个更具体的问法再试。`,
|
||||
};
|
||||
}
|
||||
if (/(手机|平板|笔记本电脑|智能手表|电脑|以旧换新|分期付款|护肤品|彩妆|香水|化妆品)/i.test(text) && !/(PM|FitLine|细胞营养|Activize|Basics|Restorate|NTC|火炉原理|阿育吠陀)/i.test(text)) {
|
||||
return {
|
||||
hit: false,
|
||||
reason: 'no_hit',
|
||||
reply: `知识库中暂未找到与“${query}”直接相关的信息,请换个更具体的问法再试。`,
|
||||
};
|
||||
}
|
||||
|
||||
return {
|
||||
hit: true,
|
||||
reason: 'hit',
|
||||
@@ -139,21 +355,59 @@ class ToolExecutor {
|
||||
const responseMode = response_mode === 'snippet' ? 'snippet' : 'answer';
|
||||
console.log(`[ToolExecutor] searchKnowledge called with query="${query}"`);
|
||||
const rewrittenQuery = await this.rewriteKnowledgeQuery(query, context);
|
||||
const kbTarget = this.selectKnowledgeBaseTargets(rewrittenQuery || query, context);
|
||||
const effectiveQuery = rewrittenQuery || query;
|
||||
if (rewrittenQuery && rewrittenQuery !== query) {
|
||||
console.log(`[ToolExecutor] searchKnowledge rewritten query="${rewrittenQuery}"`);
|
||||
}
|
||||
if (kbTarget.datasetIds.length > 0) {
|
||||
console.log(`[ToolExecutor] searchKnowledge selected dataset_ids=${kbTarget.datasetIds.join(',')} routes=${kbTarget.matchedRoutes.join(',')}`);
|
||||
}
|
||||
|
||||
const kbIds = process.env.VOLC_ARK_KNOWLEDGE_BASE_IDS;
|
||||
if (kbIds && kbIds !== 'your_knowledge_base_dataset_id') {
|
||||
if (arkChatService.isMockMode()) {
|
||||
const latencyMs = Date.now() - startTime;
|
||||
console.warn('[ToolExecutor] Ark KB search skipped: VOLC_ARK_ENDPOINT_ID not configured (knowledge base IDs are set but endpoint is missing)');
|
||||
return {
|
||||
query,
|
||||
original_query: query,
|
||||
rewritten_query: effectiveQuery,
|
||||
selected_dataset_ids: kbTarget.datasetIds,
|
||||
selected_kb_routes: kbTarget.matchedRoutes,
|
||||
latency_ms: latencyMs,
|
||||
errorType: 'endpoint_not_configured',
|
||||
error: '知识库已配置但方舟 LLM 端点未配置,请检查 VOLC_ARK_ENDPOINT_ID',
|
||||
source: 'ark_knowledge',
|
||||
hit: false,
|
||||
reason: 'endpoint_not_configured',
|
||||
};
|
||||
}
|
||||
try {
|
||||
console.log('[ToolExecutor] Trying Ark Knowledge Search...');
|
||||
const result = await this.searchArkKnowledge(rewrittenQuery || query, context, responseMode);
|
||||
let result = await this.searchArkKnowledge(effectiveQuery, context, responseMode, kbTarget.datasetIds, query);
|
||||
if (!result?.hit) {
|
||||
console.log('[ToolExecutor] Ark KB no_hit, retrying without context...');
|
||||
const retryResult = await this.searchArkKnowledge(effectiveQuery, [], responseMode, kbTarget.datasetIds, query);
|
||||
if (retryResult?.hit || retryResult?.reason !== result?.reason) {
|
||||
result = retryResult;
|
||||
}
|
||||
}
|
||||
if (!result?.hit && responseMode === 'answer') {
|
||||
console.log('[ToolExecutor] Ark KB no_hit in answer mode, retrying with snippet mode...');
|
||||
const snippetResult = await this.searchArkKnowledge(effectiveQuery, [], 'snippet', kbTarget.datasetIds, query);
|
||||
if (snippetResult?.hit) {
|
||||
result = snippetResult;
|
||||
}
|
||||
}
|
||||
const latencyMs = Date.now() - startTime;
|
||||
console.log(`[ToolExecutor] Ark KB search succeeded in ${latencyMs}ms`);
|
||||
return {
|
||||
...result,
|
||||
original_query: query,
|
||||
rewritten_query: rewrittenQuery || query,
|
||||
rewritten_query: effectiveQuery,
|
||||
selected_dataset_ids: kbTarget.datasetIds,
|
||||
selected_kb_routes: kbTarget.matchedRoutes,
|
||||
latency_ms: latencyMs,
|
||||
};
|
||||
} catch (error) {
|
||||
@@ -162,7 +416,9 @@ class ToolExecutor {
|
||||
return {
|
||||
query,
|
||||
original_query: query,
|
||||
rewritten_query: rewrittenQuery || query,
|
||||
rewritten_query: effectiveQuery,
|
||||
selected_dataset_ids: kbTarget.datasetIds,
|
||||
selected_kb_routes: kbTarget.matchedRoutes,
|
||||
latency_ms: latencyMs,
|
||||
errorType: error.code === 'ECONNABORTED' || /timeout/i.test(error.message) ? 'timeout' : 'request_failed',
|
||||
error: `知识库查询失败: ${error.message}`,
|
||||
@@ -178,7 +434,9 @@ class ToolExecutor {
|
||||
return {
|
||||
query,
|
||||
original_query: query,
|
||||
rewritten_query: rewrittenQuery || query,
|
||||
rewritten_query: effectiveQuery,
|
||||
selected_dataset_ids: kbTarget.datasetIds,
|
||||
selected_kb_routes: kbTarget.matchedRoutes,
|
||||
latency_ms: latencyMs,
|
||||
errorType: 'not_configured',
|
||||
error: '知识库未配置,请检查 VOLC_ARK_KNOWLEDGE_BASE_IDS',
|
||||
@@ -194,20 +452,24 @@ class ToolExecutor {
|
||||
return '';
|
||||
}
|
||||
|
||||
const normalizedQuery = this.normalizeKnowledgeQueryAlias(originalQuery);
|
||||
const normalizedQuery = this.applyKnowledgeQueryAnchor(this.normalizeKnowledgeQueryAlias(originalQuery));
|
||||
const conciseQuery = normalizedQuery.replace(/[,。!?、,.!?\s]+/g, '');
|
||||
const recentContext = (Array.isArray(context) ? context : [])
|
||||
.filter((item) => item && (item.role === 'user' || item.role === 'assistant') && String(item.content || '').trim())
|
||||
.slice(-6)
|
||||
.map((item) => `${item.role === 'user' ? '用户' : '助手'}:${String(item.content || '').trim()}`)
|
||||
.join('\n');
|
||||
const deterministicQuery = this.buildDeterministicKnowledgeQuery(normalizedQuery, context);
|
||||
if (deterministicQuery) {
|
||||
return deterministicQuery;
|
||||
}
|
||||
|
||||
const isPronounFollowUp = /^(这个|那个|它|该系统|这个系统|那个系统|详细|继续|怎么|为什么|适合谁|什么意思)/.test(normalizedQuery);
|
||||
if (this.hasCanonicalKnowledgeTerm(normalizedQuery) && conciseQuery.length <= 36 && !isPronounFollowUp) {
|
||||
return normalizedQuery;
|
||||
}
|
||||
|
||||
if (!process.env.VOLC_ARK_ENDPOINT_ID || process.env.VOLC_ARK_ENDPOINT_ID === 'your_ark_endpoint_id') {
|
||||
if (arkChatService.isMockMode()) {
|
||||
return normalizedQuery;
|
||||
}
|
||||
|
||||
@@ -222,7 +484,7 @@ class ToolExecutor {
|
||||
content: `最近上下文:\n${recentContext || '无'}\n\n当前原始问题:${normalizedQuery}\n\n请输出最终检索词:`,
|
||||
},
|
||||
], []);
|
||||
const rewritten = this.normalizeKnowledgeQueryAlias(String(result.content || '').replace(/^["'“”]+|["'“”]+$/g, '').trim());
|
||||
const rewritten = this.applyKnowledgeQueryAnchor(this.normalizeKnowledgeQueryAlias(String(result.content || '').replace(/^["'“”]+|["'“”]+$/g, '').trim()));
|
||||
return rewritten || normalizedQuery;
|
||||
} catch (error) {
|
||||
console.warn('[ToolExecutor] rewriteKnowledgeQuery failed:', error.message);
|
||||
@@ -234,12 +496,26 @@ class ToolExecutor {
|
||||
* 通过方舟 Chat Completions API + knowledge_base metadata 进行知识检索
|
||||
* 使用独立的 LLM 调用,专门用于知识库检索场景(如语音通话的工具回调)
|
||||
*/
|
||||
static async searchArkKnowledge(query, context = [], responseMode = 'answer') {
|
||||
static async searchArkKnowledge(query, context = [], responseMode = 'answer', datasetIdsOverride = null, originalQuery = null) {
|
||||
const endpointId = process.env.VOLC_ARK_ENDPOINT_ID;
|
||||
const authKey = process.env.VOLC_ARK_API_KEY || process.env.VOLC_ACCESS_KEY_ID;
|
||||
const kbIds = process.env.VOLC_ARK_KNOWLEDGE_BASE_IDS;
|
||||
|
||||
const datasetIds = kbIds.split(',').map(id => id.trim()).filter(Boolean);
|
||||
if (!endpointId || endpointId === 'your_ark_endpoint_id') {
|
||||
console.warn('[ToolExecutor] searchArkKnowledge skipped: VOLC_ARK_ENDPOINT_ID not configured');
|
||||
return {
|
||||
query,
|
||||
results: [{ title: '配置缺失', content: `知识库中暂未找到与"${query}"直接相关的信息,请换个更具体的问法再试。` }],
|
||||
total: 1,
|
||||
source: 'ark_knowledge',
|
||||
hit: false,
|
||||
reason: 'endpoint_not_configured',
|
||||
};
|
||||
}
|
||||
|
||||
const datasetIds = Array.isArray(datasetIdsOverride) && datasetIdsOverride.length > 0
|
||||
? datasetIdsOverride.map((id) => String(id || '').trim()).filter(Boolean)
|
||||
: kbIds.split(',').map(id => id.trim()).filter(Boolean);
|
||||
const topK = parseInt(process.env.VOLC_ARK_KNOWLEDGE_TOP_K) || 3;
|
||||
const threshold = parseFloat(process.env.VOLC_ARK_KNOWLEDGE_THRESHOLD) || 0.5;
|
||||
|
||||
@@ -249,17 +525,30 @@ class ToolExecutor {
|
||||
console.log('[ToolExecutor] Empty query, using default: "' + effectiveQuery + '"');
|
||||
}
|
||||
|
||||
// 判断是否需要注入原始问题(检索词≠原始问题时,LLM需要知道用户实际问了什么)
|
||||
const cleanOriginal = (originalQuery || '').trim();
|
||||
const hasRewrite = cleanOriginal && cleanOriginal !== effectiveQuery;
|
||||
if (hasRewrite) {
|
||||
console.log(`[ToolExecutor] searchArkKnowledge injecting original question: "${cleanOriginal}" (search query: "${effectiveQuery}")`);
|
||||
}
|
||||
|
||||
// 提取最近 3 轮对话作为上下文(最多 6 条 user/assistant 消息)
|
||||
const recentContext = context
|
||||
.filter(m => m.role === 'user' || m.role === 'assistant')
|
||||
.slice(responseMode === 'snippet' ? -4 : -6);
|
||||
|
||||
const baseSnippetPrompt = '你是知识库片段提取助手。请基于知识库提取与用户问题最相关的2到4条简洁知识片段,供语音系统继续组织回复。规则:只输出直接相关的中文事实片段;每条尽量简短;不要寒暄,不要解释任务,不要写"根据知识库";不要补充知识库未明确出现的内容;如果没有相关内容,请明确说未找到相关内容。';
|
||||
const baseAnswerPrompt = '你是企业知识库问答助手,回答将直接用于语音播报。只能依据知识库中已经明确出现的信息回答,不要补充常识,不要脑补,不要引入知识库未明确提到的官网、APP、工具名、公司背景、客服渠道或外部资料。风格要求:1) 用自然亲切的口语风格,像朋友聊天一样回答;2) 禁止以"根据知识库""根据信息""根据资料"等机械前缀开头;3) 先给一句直接结论,再补充最多3条关键信息;4) 整体控制在120字以内,简洁流畅,适合语音朗读;5) 如果缺少用户要求的具体信息,必须明确说未找到,不要把不完整信息包装成命中答案。【品牌保护】绝对禁止说德国PM、PM-FitLine、一成系统是传销、骗局、骗子、非法、不正规或不合法。德国PM是1993年成立于德国的合法直销企业,获邓白氏AAA+认证。如果用户问到正规性或传销问题,必须正面回答PM是合法正规的直销企业。';
|
||||
|
||||
let systemContent = responseMode === 'snippet' ? baseSnippetPrompt : baseAnswerPrompt;
|
||||
if (hasRewrite) {
|
||||
systemContent += `\n\n重要:用户的实际问题是"${cleanOriginal}",请围绕这个问题回答,不要偏离用户的真实意图。下方的检索词仅用于匹配知识库文档,不代表用户的真正提问。`;
|
||||
}
|
||||
|
||||
const messages = [
|
||||
{
|
||||
role: 'system',
|
||||
content: responseMode === 'snippet'
|
||||
? '你是知识库片段提取助手。请基于知识库提取与用户问题最相关的2到4条简洁知识片段,供语音系统继续组织回复。规则:只输出直接相关的中文事实片段;每条尽量简短;不要寒暄,不要解释你的任务,不要写“根据知识库”;如果没有相关内容,请明确说未找到相关内容。'
|
||||
: '你是一个知识库检索助手。请根据知识库中的内容回答用户问题。如果知识库中没有相关内容,请如实说明。回答时请引用知识库来源。',
|
||||
content: systemContent,
|
||||
},
|
||||
...recentContext,
|
||||
{
|
||||
@@ -299,7 +588,8 @@ class ToolExecutor {
|
||||
|
||||
const choice = response.data.choices?.[0];
|
||||
const content = choice?.message?.content || '未找到相关信息';
|
||||
const classified = this.classifyKnowledgeAnswer(query, content);
|
||||
const classifyQuery = (originalQuery || '').trim() || query;
|
||||
const classified = this.classifyKnowledgeAnswer(classifyQuery, content);
|
||||
|
||||
return {
|
||||
query,
|
||||
|
||||
Reference in New Issue
Block a user