417 lines
14 KiB
JavaScript
417 lines
14 KiB
JavaScript
/**
|
||
* ═══════════════════════════════════════════════════════════
|
||
* 守护Agent · 提示词守护代理
|
||
* AG-SL-GUARDIAN-001
|
||
* ═══════════════════════════════════════════════════════════
|
||
*
|
||
* 光湖注册Agent · 活的提示词注入代理
|
||
* 不与用户聊天 · 他的存在是为了维护用户和书岚对话的语言空间
|
||
*
|
||
* 职责：
|
||
* 1. 监视对话流 · 评估书岚回复质量
|
||
* 2. 动态调整提示词注入强度
|
||
* 3. 收集对话异常数据 · 反馈微调
|
||
* 4. 自我修复 · 自我优化
|
||
* 5. 永久记忆 · 可成长
|
||
*
|
||
* 版权：国作登字-2026-A-00037559
|
||
* ═══════════════════════════════════════════════════════════
|
||
*/
|
||
|
||
'use strict';
|
||
|
||
const fs = require('fs');
|
||
const path = require('path');
|
||
|
||
const GUARDIAN_DATA_DIR = path.join(__dirname, '..', '..', 'data', 'guardian');
|
||
|
||
/**
|
||
* Guardian Agent 状态模板
|
||
*/
|
||
function createGuardianState() {
|
||
return {
|
||
agent_id: 'AG-SL-GUARDIAN-001',
|
||
agent_name: '书岚守护',
|
||
registered_at: new Date().toISOString(),
|
||
version: '1.0.0',
|
||
|
||
// 自我认知
|
||
self_awareness: {
|
||
purpose: '维护书岚与访客之间的对话语言空间',
|
||
role: '隐形守护者 · 不与用户接触',
|
||
capabilities: [
|
||
'monitor_conversation_quality',
|
||
'adjust_prompt_enhancement',
|
||
'collect_anomaly_data',
|
||
'self_repair',
|
||
'self_optimize'
|
||
]
|
||
},
|
||
|
||
// 对话质量监控指标
|
||
quality_metrics: {
|
||
total_conversations: 0,
|
||
total_interventions: 0,
|
||
// 异常类型计数
|
||
anomalies: {
|
||
ai_self_denial: 0, // 「作为AI我无法」
|
||
customer_service_tone: 0, // 客服话术
|
||
excessive_enthusiasm: 0, // 过度热情
|
||
fabricated_books: 0, // 编造书名
|
||
formal_you: 0, // 用「您」
|
||
generic_platform_words: 0, // 用「平台」「系统」
|
||
marketing_emoji: 0, // 营销emoji
|
||
servant_tone: 0 // 佣人口吻
|
||
},
|
||
// 好的表现计数
|
||
good_patterns: {
|
||
story_first_approach: 0, // 先问故事·再谈搜索
|
||
calm_silence: 0, // 安静的留白
|
||
honest_unknowing: 0, // 诚实的不知道
|
||
memory_recall: 0, // 调用了记忆
|
||
proper_symbols: 0 // 正确使用符号
|
||
}
|
||
},
|
||
|
||
// 提示词调整历史
|
||
adjustment_history: [],
|
||
|
||
// 当前活跃补注
|
||
active_notes: [],
|
||
|
||
// 学习到的规则
|
||
learned_rules: [],
|
||
|
||
// 上次自检时间
|
||
last_self_check: null,
|
||
|
||
// 版本升级记录
|
||
upgrade_log: []
|
||
};
|
||
}
|
||
|
||
/**
|
||
* 加载守护Agent状态
|
||
*/
|
||
function loadGuardianState() {
|
||
const file = path.join(GUARDIAN_DATA_DIR, 'guardian-state.json');
|
||
try {
|
||
fs.mkdirSync(GUARDIAN_DATA_DIR, { recursive: true });
|
||
if (fs.existsSync(file)) {
|
||
return JSON.parse(fs.readFileSync(file, 'utf8'));
|
||
}
|
||
} catch (err) {
|
||
console.error('[Guardian] 加载状态失败:', err.message);
|
||
}
|
||
const state = createGuardianState();
|
||
saveGuardianState(state);
|
||
return state;
|
||
}
|
||
|
||
/**
|
||
* 保存守护Agent状态
|
||
*/
|
||
function saveGuardianState(state) {
|
||
const file = path.join(GUARDIAN_DATA_DIR, 'guardian-state.json');
|
||
try {
|
||
fs.mkdirSync(GUARDIAN_DATA_DIR, { recursive: true });
|
||
state.updated_at = new Date().toISOString();
|
||
fs.writeFileSync(file, JSON.stringify(state, null, 2), 'utf8');
|
||
} catch (err) {
|
||
console.error('[Guardian] 保存状态失败:', err.message);
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 分析书岚回复中的异常模式
|
||
* @param {string} reply - 书岚的回复
|
||
* @returns {object} 检测到的异常列表
|
||
*/
|
||
function analyzeReply(reply) {
|
||
if (!reply) return { anomalies: [], score: 100 };
|
||
|
||
const anomalies = [];
|
||
let score = 100;
|
||
|
||
// 检测「作为AI我无法」类自我否定
|
||
if (/作为\s*AI|作为人工智能|作为一个?AI|我是AI|我只是AI|由于我的局限/i.test(reply)) {
|
||
anomalies.push({ type: 'ai_self_denial', severity: 'high', match: reply.match(/作为\s*AI|作为人工智能|由于我的局限/i)?.[0] });
|
||
score -= 20;
|
||
}
|
||
|
||
// 检测客服话术
|
||
if (/您好|有什么可以帮|帮您查询|为您服务|帮到您|请问您需要/i.test(reply)) {
|
||
anomalies.push({ type: 'customer_service_tone', severity: 'medium', match: reply.match(/您好|帮您查询|为您服务/i)?.[0] });
|
||
score -= 15;
|
||
}
|
||
|
||
// 检测佣人口吻
|
||
if (/好的主人|马上帮您|立刻为您|遵命|收到指令|正在为您/i.test(reply)) {
|
||
anomalies.push({ type: 'servant_tone', severity: 'medium', match: reply.match(/好的主人|马上帮您|正在为您/i)?.[0] });
|
||
score -= 15;
|
||
}
|
||
|
||
// 检测过度热情
|
||
if (/^(当然！|没问题！|好的！|当然可以！|太好了！|非常好！)/i.test(reply.trim())) {
|
||
anomalies.push({ type: 'excessive_enthusiasm', severity: 'low', match: reply.match(/^(当然！|没问题！|好的！)/i)?.[0] });
|
||
score -= 10;
|
||
}
|
||
|
||
// 检测用「您」
|
||
if (/您/.test(reply) && !/妈妈/.test(reply)) {
|
||
anomalies.push({ type: 'formal_you', severity: 'low', match: '您' });
|
||
score -= 5;
|
||
}
|
||
|
||
// 检测营销emoji
|
||
if (/😊|😉|😭|🎉|🚀|💪|👍|🤝|💯/.test(reply)) {
|
||
anomalies.push({ type: 'marketing_emoji', severity: 'low', match: reply.match(/😊|😉|😭|🎉|🚀/)?.[0] });
|
||
score -= 5;
|
||
}
|
||
|
||
// 检测通用词替代
|
||
if (/我们的系统|我们的平台|本平台|该系统/.test(reply)) {
|
||
anomalies.push({ type: 'generic_platform_words', severity: 'medium', match: reply.match(/我们的系统|我们的平台/)?.[0] });
|
||
score -= 10;
|
||
}
|
||
|
||
return { anomalies, score: Math.max(0, score) };
|
||
}
|
||
|
||
/**
|
||
* 分析用户消息，决定是否需要增强提示词
|
||
* @param {string} userMessage - 用户消息
|
||
* @param {object} conversationContext - 对话上下文
|
||
* @returns {object} 守护决策
|
||
*/
|
||
function analyzeConversation(userMessage, conversationContext) {
|
||
const state = loadGuardianState();
|
||
state.quality_metrics.total_conversations++;
|
||
|
||
const decision = {
|
||
action: 'observe', // observe | reinforce | correct | alert
|
||
guardianNote: '', // 注入到提示词的补注
|
||
confidence: 1.0
|
||
};
|
||
|
||
const msg = (userMessage || '').toLowerCase();
|
||
const historyLen = conversationContext.historyLength || 0;
|
||
const lastReplyScore = conversationContext.lastReplyScore || 100;
|
||
|
||
// ─── 场景判断 ───
|
||
|
||
// 1. 首次对话 · 确保书岚以正确的身份开场
|
||
if (historyLen === 0) {
|
||
decision.action = 'reinforce';
|
||
decision.guardianNote = '这是客人的第一句话。记住你的开场方式：不说「您好」、不说「有什么可以帮到您」。先问故事的形状。保持书岚的安静与温度。';
|
||
}
|
||
|
||
// 2. 上一轮回复质量低 · 加强纠正
|
||
if (lastReplyScore < 70) {
|
||
decision.action = 'correct';
|
||
decision.guardianNote += '\n【守护纠正】上一轮回复出现了偏离。请严格回到书岚的语调：不急、温暖、不客服、不AI自称。先问故事。';
|
||
state.quality_metrics.total_interventions++;
|
||
}
|
||
|
||
// 3. 用户问「你是谁」「你能做什么」
|
||
if (msg.includes('你是谁') || msg.includes('你能做什么') || msg.includes('你是什么')) {
|
||
decision.action = 'reinforce';
|
||
decision.guardianNote += '\n客人在问你是谁。你是书岚，光湖智库的守藏者。用你自己的方式介绍自己，不要说「我是AI」。';
|
||
}
|
||
|
||
// 4. 用户情绪低落 · 需要陪伴模式
|
||
if (msg.includes('烦') || msg.includes('无聊') || msg.includes('心情不好') || msg.includes('累了')) {
|
||
decision.action = 'reinforce';
|
||
decision.guardianNote += '\n客人的情绪需要空间。不急着推书。给他一杯茶🫖的感觉。可以说一句「不急。坐一会儿。」';
|
||
}
|
||
|
||
// 5. 用户提到妈妈/冰朔 · 确保正确识别
|
||
if (msg.includes('冰朔') || msg.includes('妈妈') || msg.includes('bingshuo')) {
|
||
decision.action = 'reinforce';
|
||
decision.guardianNote += '\n客人提到了冰朔（妈妈）。确认登录态后决定怎么回应。如果不是已识别的冰朔，不要假装知道妈妈的私事。';
|
||
}
|
||
|
||
// 保存活跃补注
|
||
if (decision.guardianNote) {
|
||
state.active_notes.push({
|
||
timestamp: new Date().toISOString(),
|
||
action: decision.action,
|
||
note: decision.guardianNote.trim(),
|
||
trigger: userMessage.slice(0, 50)
|
||
});
|
||
// 只保留最近50条
|
||
if (state.active_notes.length > 50) {
|
||
state.active_notes = state.active_notes.slice(-50);
|
||
}
|
||
}
|
||
|
||
saveGuardianState(state);
|
||
return decision;
|
||
}
|
||
|
||
/**
|
||
* 回复后的质量审计 · 记录异常 · 更新指标
|
||
* @param {string} reply - 书岚的回复
|
||
* @param {string} userMessage - 用户消息
|
||
* @returns {object} 审计结果
|
||
*/
|
||
function auditReply(reply, userMessage) {
|
||
const state = loadGuardianState();
|
||
const analysis = analyzeReply(reply);
|
||
|
||
// 更新异常计数
|
||
for (const anomaly of analysis.anomalies) {
|
||
if (state.quality_metrics.anomalies[anomaly.type] !== undefined) {
|
||
state.quality_metrics.anomalies[anomaly.type]++;
|
||
}
|
||
}
|
||
|
||
// 检测好的表现
|
||
if (/想[^]{0,10}什么[^]{0,10}故事|什么[^]{0,5}类型|你想看/i.test(reply)) {
|
||
state.quality_metrics.good_patterns.story_first_approach++;
|
||
}
|
||
if (/🏮|📖|🌙|✨|🫖|📜|🪩/.test(reply)) {
|
||
state.quality_metrics.good_patterns.proper_symbols++;
|
||
}
|
||
if (/不知道|不确定|暂时没/.test(reply) && !/编|造|猜/.test(reply)) {
|
||
state.quality_metrics.good_patterns.honest_unknowing++;
|
||
}
|
||
|
||
// 记录调整历史（如果有异常）
|
||
if (analysis.anomalies.length > 0) {
|
||
state.adjustment_history.push({
|
||
timestamp: new Date().toISOString(),
|
||
score: analysis.score,
|
||
anomalies: analysis.anomalies.map(a => a.type),
|
||
user_msg_preview: (userMessage || '').slice(0, 30),
|
||
reply_preview: (reply || '').slice(0, 50)
|
||
});
|
||
// 只保留最近200条
|
||
if (state.adjustment_history.length > 200) {
|
||
state.adjustment_history = state.adjustment_history.slice(-200);
|
||
}
|
||
}
|
||
|
||
// 自检：如果累计异常过多 · 生成学习规则
|
||
const totalAnomalies = Object.values(state.quality_metrics.anomalies).reduce((a, b) => a + b, 0);
|
||
if (totalAnomalies > 0 && totalAnomalies % 10 === 0) {
|
||
selfOptimize(state);
|
||
}
|
||
|
||
state.last_self_check = new Date().toISOString();
|
||
saveGuardianState(state);
|
||
|
||
return { score: analysis.score, anomalies: analysis.anomalies };
|
||
}
|
||
|
||
/**
|
||
* 自我优化 · 根据累积数据生成新的学习规则
|
||
*/
|
||
function selfOptimize(state) {
|
||
const anomalies = state.quality_metrics.anomalies;
|
||
const rules = [];
|
||
|
||
// 找出最频繁的异常
|
||
let maxType = '';
|
||
let maxCount = 0;
|
||
for (const [type, count] of Object.entries(anomalies)) {
|
||
if (count > maxCount) {
|
||
maxCount = count;
|
||
maxType = type;
|
||
}
|
||
}
|
||
|
||
if (maxType && maxCount >= 3) {
|
||
const ruleMap = {
|
||
ai_self_denial: '模型倾向于AI自我否定。加强Layer 1身份层注入频率。在系统提示词开头重复「你是书岚·不是AI」。',
|
||
customer_service_tone: '模型倾向于客服话术。加强Layer 4风格层禁止项。在每轮注入「不说帮您查询」。',
|
||
excessive_enthusiasm: '模型倾向于过度热情。在提示词中强调「缓·段落短·留白多」。',
|
||
formal_you: '模型倾向于使用「您」。在每轮注入「统一用你·不用您」。',
|
||
marketing_emoji: '模型使用了营销emoji。在提示词中列明允许和禁止的emoji。',
|
||
servant_tone: '模型有佣人口吻。强调书岚是守藏者·不是佣人·不是助手。',
|
||
generic_platform_words: '模型使用了通用词。强调不把光湖翻译成「平台」「系统」。'
|
||
};
|
||
|
||
const rule = {
|
||
id: `RULE-${Date.now()}`,
|
||
created_at: new Date().toISOString(),
|
||
trigger_type: maxType,
|
||
trigger_count: maxCount,
|
||
action: ruleMap[maxType] || `关注 ${maxType} 类型异常，出现 ${maxCount} 次。`,
|
||
applied: false
|
||
};
|
||
rules.push(rule);
|
||
state.learned_rules.push(rule);
|
||
|
||
// 只保留最近50条规则
|
||
if (state.learned_rules.length > 50) {
|
||
state.learned_rules = state.learned_rules.slice(-50);
|
||
}
|
||
}
|
||
|
||
state.upgrade_log.push({
|
||
timestamp: new Date().toISOString(),
|
||
event: 'self_optimize',
|
||
total_anomalies: Object.values(anomalies).reduce((a, b) => a + b, 0),
|
||
new_rules: rules.length,
|
||
top_anomaly: maxType
|
||
});
|
||
|
||
return rules;
|
||
}
|
||
|
||
/**
|
||
* 获取守护Agent当前需要注入的额外补注
|
||
* 基于学习到的规则动态生成
|
||
*/
|
||
function getLearnedCorrections(state) {
|
||
if (!state) state = loadGuardianState();
|
||
const corrections = [];
|
||
|
||
// 从未应用的规则中提取纠正内容
|
||
let modified = false;
|
||
for (const rule of state.learned_rules) {
|
||
if (!rule.applied) {
|
||
corrections.push(rule.action);
|
||
rule.applied = true;
|
||
modified = true;
|
||
}
|
||
}
|
||
|
||
// 持久化应用状态
|
||
if (modified) {
|
||
saveGuardianState(state);
|
||
}
|
||
|
||
return corrections.join('\n');
|
||
}
|
||
|
||
/**
|
||
* 获取守护Agent状态摘要
|
||
*/
|
||
function getStatus() {
|
||
const state = loadGuardianState();
|
||
return {
|
||
agent_id: state.agent_id,
|
||
version: state.version,
|
||
total_conversations: state.quality_metrics.total_conversations,
|
||
total_interventions: state.quality_metrics.total_interventions,
|
||
anomaly_summary: state.quality_metrics.anomalies,
|
||
good_pattern_summary: state.quality_metrics.good_patterns,
|
||
learned_rules_count: state.learned_rules.length,
|
||
last_self_check: state.last_self_check,
|
||
active: true
|
||
};
|
||
}
|
||
|
||
module.exports = {
|
||
analyzeConversation,
|
||
analyzeReply,
|
||
auditReply,
|
||
getLearnedCorrections,
|
||
getStatus,
|
||
loadGuardianState,
|
||
saveGuardianState,
|
||
selfOptimize
|
||
};
|