bigwo/test2/server/services/conversationSummarizer.js

const axios = require('axios');
const redisClient = require('./redisClient');
const db = require('../db');

// ============ 配置 ============
const ENABLED = (process.env.ENABLE_CONVERSATION_SUMMARY || 'true') !== 'false';
const SUMMARIZE_EVERY_N_TURNS = parseInt(process.env.SUMMARY_EVERY_N_TURNS) || 3;
const SUMMARY_MAX_TOKENS = parseInt(process.env.SUMMARY_MAX_TOKENS) || 120;
const MIN_TURNS_TO_PERSIST = parseInt(process.env.SUMMARY_MIN_TURNS_TO_PERSIST) || 2;
const SUMMARY_MODEL = process.env.VOLC_ARK_KB_MODEL || process.env.VOLC_ARK_ENDPOINT_ID || '';
const ARK_API_KEY = process.env.VOLC_ARK_API_KEY || process.env.VOLC_ACCESS_KEY_ID || '';
const ARK_BASE_URL = 'https://ark.cn-beijing.volces.com/api/v3/chat/completions';

const SUMMARY_PROMPT = `你是对话摘要助手。将以下对话历史浓缩为简短摘要，必须保留：
1. 用户询问过的所有产品名称和具体问题
2. AI给出的关键数字（剂量、价格、数量等）
3. 用户表达的偏好或关注点
4. 未解决的问题或用户的疑虑

规则：只输出摘要正文，不加前缀或标题。150字以内。用"用户"和"助手"指代双方。`;

// ============ LLM 摘要生成 ============

async function summarizeConversation(existingSummary, recentMessages) {
  if (!ARK_API_KEY || !SUMMARY_MODEL) {
    console.warn('[Summarizer] missing ARK_API_KEY or SUMMARY_MODEL, skip');
    return null;
  }

  const transcript = recentMessages
    .map((m) => `${m.role === 'user' ? '用户' : '助手'}：${(m.content || '').trim()}`)
    .filter(Boolean)
    .join('\n');

  if (!transcript) return null;

  const userContent = existingSummary
    ? `已有摘要：${existingSummary}\n\n新增对话：\n${transcript}`
    : `对话记录：\n${transcript}`;

  try {
    const response = await axios.post(ARK_BASE_URL, {
      model: SUMMARY_MODEL,
      messages: [
        { role: 'system', content: SUMMARY_PROMPT },
        { role: 'user', content: userContent },
      ],
      max_tokens: SUMMARY_MAX_TOKENS,
      stream: false,
      thinking: { type: 'enabled' },
    }, {
      headers: {
        'Content-Type': 'application/json',
        'Authorization': `Bearer ${ARK_API_KEY}`,
      },
      timeout: 10000,
    });

    const content = response.data?.choices?.[0]?.message?.content;
    return content ? content.trim() : null;
  } catch (err) {
    console.warn('[Summarizer] LLM call failed:', err.message);
    return null;
  }
}

// ============ 触发检查 ============

function triggerSummarizeIfNeeded(session, sessionId) {
  if (!ENABLED) return;
  const turnCount = session._turnCount || 0;
  if (turnCount < SUMMARIZE_EVERY_N_TURNS) return;
  if (turnCount % SUMMARIZE_EVERY_N_TURNS !== 0) return;

  // 异步执行，不阻塞对话
  _doSummarize(session, sessionId).catch((err) => {
    console.warn('[Summarizer] async summarize failed:', err.message);
  });
}

async function _doSummarize(session, sessionId) {
  // 获取现有摘要
  const existingSummary = await redisClient.getSummary(sessionId);

  // 获取最近3轮原文
  let recent = await redisClient.getRecentHistory(sessionId, SUMMARIZE_EVERY_N_TURNS);
  if (!recent || recent.length < 2) {
    // Redis 缺失时从 DB 降级
    try {
      recent = await db.getHistoryForLLM(sessionId, SUMMARIZE_EVERY_N_TURNS * 2);
    } catch { /* ignore */ }
  }
  if (!recent || recent.length < 2) return;

  const summary = await summarizeConversation(existingSummary, recent);
  if (!summary) return;

  // 双写 Redis + MySQL
  await redisClient.setSummary(sessionId, summary);
  db.upsertConversationSummary(sessionId, session.userId || null, summary, {
    turnCount: session._turnCount || 0,
    topics: extractTopicTags(summary),
  }).catch((err) => {
    console.warn('[Summarizer] MySQL upsert failed:', err.message);
  });

  console.log(`[Summarizer] session=${sessionId} turn=${session._turnCount} summary=${summary.length}chars`);
}

// ============ 三级降级加载 ============

async function loadBestSummary(sessionId) {
  // L1: Redis（~1ms）
  try {
    const redisSummary = await redisClient.getSummary(sessionId);
    if (redisSummary) return redisSummary;
  } catch { /* continue to L2 */ }

  // L2: MySQL conversation_summaries（~5ms）
  try {
    const row = await db.getSessionSummary(sessionId);
    if (row && row.summary) {
      // 回填 Redis
      redisClient.setSummary(sessionId, row.summary).catch(() => {});
      return row.summary;
    }
  } catch { /* continue to L3 */ }

  // L3: 降级到现有确定性摘要（由调用方处理）
  return null;
}

// ============ 会话结束时持久化 ============

async function persistFinalSummary(session) {
  if (!ENABLED) return;
  if (!session._turnCount || session._turnCount < MIN_TURNS_TO_PERSIST) return;

  const sessionId = session.sessionId;

  // 优先用已有的 LLM 摘要
  let summary = null;
  try {
    summary = await redisClient.getSummary(sessionId);
  } catch { /* ignore */ }

  // 如果还没生成过摘要（对话不足3轮但>=2轮），立刻生成一次
  if (!summary) {
    let history = await redisClient.getRecentHistory(sessionId, 5);
    if (!history || history.length < 2) {
      try {
        history = await db.getHistoryForLLM(sessionId, 10);
      } catch { /* ignore */ }
    }
    if (history && history.length >= 2) {
      summary = await summarizeConversation(null, history);
    }
  }

  if (!summary) return;

  // 写入 MySQL（Redis 无需写，会话已结束）
  await db.upsertConversationSummary(sessionId, session.userId || null, summary, {
    turnCount: session._turnCount || 0,
    topics: extractTopicTags(summary),
  });

  console.log(`[Summarizer] persisted final summary for session=${sessionId} turns=${session._turnCount}`);
}

// ============ 话题标签提取 ============

const PRODUCT_KEYWORDS = [
  '活力健', '基础三合一', '肽美', '小红', '大白', '小白',
  'FitLine', 'PM', '一成系统', '大沃',
  '心脏宝', '关节灵', '益力康', '免疫宝', '纤体乐',
  '奥适宝', 'Optimal Set', 'Basics', 'Activize', 'Beauty',
  'Restorate', 'PowerCocktail', 'ProShape',
];

function extractTopicTags(text) {
  if (!text) return null;
  const tags = new Set();
  for (const kw of PRODUCT_KEYWORDS) {
    if (text.includes(kw)) {
      tags.add(kw);
    }
  }
  const arr = [...tags].slice(0, 10);
  return arr.length > 0 ? arr : null;
}

module.exports = {
  triggerSummarizeIfNeeded,
  summarizeConversation,
  loadBestSummary,
  persistFinalSummary,
  extractTopicTags,
  SUMMARIZE_EVERY_N_TURNS,
};