537 lines
18 KiB
JavaScript
537 lines
18 KiB
JavaScript
// Bot 数字人 vs 真人"对抗学习" 共享模块
|
||
// 每周为每个 bot 选 1 个真实高赞真人帖，让 bot 吸收其"为什么高互动"的洞察
|
||
//
|
||
// 用法：
|
||
// import { runWeeklyAdversarialLearning, getActiveLearningForBot, buildAdversarialBlock } from "./lib/bot-adversarial-learning.mjs";
|
||
//
|
||
// // 1) 周调度入口
|
||
// await runWeeklyAdversarialLearning({ weekKey: "2026-W23", lookbackDays: 7 });
|
||
//
|
||
// // 2) 发帖 prompt 注入对抗学习参考
|
||
// const learning = await getActiveLearningForBot(botConfigId);
|
||
// const block = buildAdversarialBlock(learning); // → "[对抗学习] ..."
|
||
|
||
import { PrismaClient } from "@prisma/client";
|
||
import { PrismaMariaDb } from "@prisma/adapter-mariadb";
|
||
import "dotenv/config";
|
||
import OpenAI from "openai";
|
||
import { withRetry } from "./retry.mjs";
|
||
|
||
const base = (process.env.DATABASE_URL || "").replace("mysql://", "mariadb://");
|
||
const sep = base.includes("?") ? "&" : "?";
|
||
const connectionString = `${base}${sep}connection_limit=3&pool_timeout=10`;
|
||
|
||
let _prisma = null;
|
||
function getPrisma() {
|
||
if (!_prisma) {
|
||
const adapter = new PrismaMariaDb(connectionString);
|
||
_prisma = new PrismaClient({ adapter });
|
||
}
|
||
return _prisma;
|
||
}
|
||
|
||
let _openai = null;
|
||
function getOpenAI() {
|
||
if (_openai) return _openai;
|
||
if (!process.env.DEEPSEEK_API_KEY) return null;
|
||
_openai = new OpenAI({
|
||
apiKey: process.env.DEEPSEEK_API_KEY,
|
||
baseURL: "https://api.deepseek.com",
|
||
});
|
||
return _openai;
|
||
}
|
||
|
||
const MODEL = process.env.DEEPSEEK_MODEL || "deepseek-v4-pro";
|
||
|
||
function ts() {
|
||
return `[${new Date().toISOString()}]`;
|
||
}
|
||
|
||
// ======== 配置 ========
|
||
export const LEARNING_CONFIG = {
|
||
// 候选帖最小门槛
|
||
minReplies: 3,
|
||
minLikes: 5,
|
||
minHumanReplies: 1,
|
||
// 回看窗口（天）
|
||
lookbackDays: 7,
|
||
// 候选数上限
|
||
candidateLimit: 30,
|
||
// 每个 bot 每周上限
|
||
perBotPerWeek: 1,
|
||
// 关联度评分
|
||
sameForumWeight: 1.0,
|
||
keywordOverlapWeight: 0.6,
|
||
recencyWeight: 0.3,
|
||
// 过期
|
||
expireDays: 7,
|
||
// LLM 提示 token 上限
|
||
contentTruncate: 1500,
|
||
};
|
||
|
||
// ======== 工具：周 key (ISO week) ========
|
||
/**
|
||
* 形如 "2026-W23"，与 date-fns / Excel WEEKNUM 行为一致
|
||
*/
|
||
export function getISOWeekKey(date = new Date()) {
|
||
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
|
||
const dayNum = d.getUTCDay() || 7; // Mon=1..Sun=7
|
||
d.setUTCDate(d.getUTCDate() + 4 - dayNum);
|
||
const yearStart = new Date(Date.UTC(d.getUTCFullYear(), 0, 1));
|
||
const weekNo = Math.ceil(((d - yearStart) / 86400000 + 1) / 7);
|
||
return `${d.getUTCFullYear()}-W${String(weekNo).padStart(2, "0")}`;
|
||
}
|
||
|
||
// ======== 工具：取候选真人帖 ========
|
||
/**
|
||
* 拉过去 N 天内由真人发布、且互动数超过门槛的 topic 列表
|
||
* 排序：综合分 = 赞×3 + 真人回复×5 + 回复
|
||
* 仅返回 topic 维度（post 维度后续可加）
|
||
*/
|
||
export async function fetchHumanHighEngagementTopics(lookbackDays = 7, limit = 30) {
|
||
const prisma = getPrisma();
|
||
const since = new Date(Date.now() - lookbackDays * 24 * 60 * 60 * 1000);
|
||
|
||
// 1) 先拉 botUserId 集合，避免 N+1
|
||
const botUsers = await prisma.user.findMany({
|
||
where: { isBot: true },
|
||
select: { id: true },
|
||
});
|
||
const botUserIdSet = new Set(botUsers.map((u) => u.id));
|
||
|
||
// 2) 拉所有非 bot 的 topic，按 replyCount desc 取前 N（粗筛）
|
||
const candidates = await prisma.forumTopic.findMany({
|
||
where: {
|
||
userId: { notIn: Array.from(botUserIdSet) },
|
||
isHidden: false,
|
||
createdAt: { gte: since },
|
||
},
|
||
include: {
|
||
user: { select: { id: true, name: true, isBot: true } },
|
||
category: { select: { slug: true, name: true } },
|
||
posts: {
|
||
where: { user: { isBot: false } },
|
||
select: { id: true },
|
||
},
|
||
},
|
||
orderBy: [{ replyCount: "desc" }, { likeCount: "desc" }],
|
||
take: limit * 2,
|
||
});
|
||
|
||
// 3) 计算综合分，过门槛
|
||
const scored = candidates
|
||
.filter((t) => !t.user?.isBot)
|
||
.map((t) => {
|
||
const humanReplies = t.posts?.length || 0;
|
||
const score = (t.likeCount || 0) * 3 + humanReplies * 5 + (t.replyCount || 0);
|
||
return { ...t, _score: score, _humanReplies: humanReplies };
|
||
})
|
||
.filter(
|
||
(t) =>
|
||
(t.replyCount || 0) >= LEARNING_CONFIG.minReplies ||
|
||
t._humanReplies >= LEARNING_CONFIG.minHumanReplies ||
|
||
(t.likeCount || 0) >= LEARNING_CONFIG.minLikes
|
||
)
|
||
.sort((a, b) => b._score - a._score)
|
||
.slice(0, limit);
|
||
|
||
return scored;
|
||
}
|
||
|
||
// ======== 工具：候选与 bot 关联度评分 ========
|
||
/**
|
||
* 计算一篇候选帖对某 bot 的关联度
|
||
* - 板块命中：bot.primaryForums 包含该帖板块 → 同板块加权
|
||
* - 关键词重叠：bot stanceKeywords 与帖子标题/内容重叠数
|
||
* - 时效衰减：越新越好
|
||
*/
|
||
function calcRelevance(bot, persona, topic) {
|
||
let score = 0;
|
||
const primaryForums = (bot.primaryForums || []).map((s) => String(s).toLowerCase());
|
||
const stanceKeywords = ((persona?.stanceKeywords || []).map((k) => k.key || k)).filter(Boolean);
|
||
|
||
if (primaryForums.includes(String(topic.category?.slug || "").toLowerCase())) {
|
||
score += LEARNING_CONFIG.sameForumWeight;
|
||
}
|
||
|
||
if (stanceKeywords.length > 0) {
|
||
const text = `${topic.title || ""} ${topic.content || ""}`.toLowerCase();
|
||
let hits = 0;
|
||
for (const kw of stanceKeywords) {
|
||
const k = String(kw).toLowerCase();
|
||
if (k.length < 2) continue;
|
||
if (text.includes(k)) hits++;
|
||
}
|
||
const overlapRatio = Math.min(1, hits / 5);
|
||
score += overlapRatio * LEARNING_CONFIG.keywordOverlapWeight;
|
||
}
|
||
|
||
// 时效：距今 < 3 天满分，> 6 天 0 分
|
||
const ageMs = Date.now() - new Date(topic.createdAt).getTime();
|
||
const ageDays = ageMs / (24 * 60 * 60 * 1000);
|
||
const recencyScore = Math.max(0, 1 - ageDays / 6) * LEARNING_CONFIG.recencyWeight;
|
||
score += recencyScore;
|
||
|
||
return Math.round(score * 1000) / 1000;
|
||
}
|
||
|
||
// ======== LLM 提取"为什么高互动"洞察 ========
|
||
const LEARN_PROMPT = (botName, botRole, topic) => `你是【${botName}】的内容策略师，每周要研究一篇真实用户（不是数字人）的高互动帖子，提炼出"为什么火"，作为下周发帖可借鉴的方向。
|
||
|
||
[数字人角色]
|
||
- 名字：${botName}
|
||
- 角色：${botRole || "通用论坛内容创作者"}
|
||
|
||
[本周高互动真人帖]
|
||
- 标题：${topic.title}
|
||
- 板块：${topic.category?.name || ""}
|
||
- 收到回复：${topic.replyCount}（真人 ${topic._humanReplies || 0}）
|
||
- 收到点赞：${topic.likeCount}
|
||
- 浏览数：${topic.viewCount}
|
||
- 内容（截取前 ${LEARNING_CONFIG.contentTruncate} 字）：
|
||
${(topic.content || "").slice(0, LEARNING_CONFIG.contentTruncate)}
|
||
|
||
[任务]
|
||
站在【${botName}】的视角，分析这篇真人帖之所以高互动的 2-3 个可学习点。注意：
|
||
1. 不要泛泛而谈"写得好"，要给出具体的"可复用的写法/角度/钩子"
|
||
2. 关注"为什么能引发真人回复"——是观点鲜明、抛问题、还是给方案/故事
|
||
3. 提炼出 2-3 个对【${botName}】下周发帖有直接借鉴价值的策略
|
||
4. 输出一段 150-250 字的简洁洞察，下周发帖时可作为参考
|
||
|
||
[输出格式]
|
||
只输出严格的 JSON：
|
||
{
|
||
"insight": "（150-250 字的洞察，聚焦可复用的写法/角度/钩子）",
|
||
"categories": ["hook|story|data|question|contrarian|empathy|practical", ...]
|
||
}`;
|
||
|
||
async function extractInsightWithLLM(botChar, topic) {
|
||
const openai = getOpenAI();
|
||
if (!openai) {
|
||
// 降级：无 LLM 时输出规则性总结
|
||
return buildFallbackInsight(topic);
|
||
}
|
||
const p = botChar?.personality || {};
|
||
const role = [p.identity, p.stance, p.speakingStyle].filter(Boolean).join(" / ");
|
||
const prompt = LEARN_PROMPT(botChar?.displayName || "数字人", role, topic);
|
||
const response = await withRetry(() =>
|
||
openai.chat.completions.create({
|
||
model: MODEL,
|
||
messages: [{ role: "user", content: prompt }],
|
||
temperature: 0.6,
|
||
// DeepSeek V4 推理 token 计入 max_tokens，预算不足会导致正文截断/为空
|
||
max_tokens: 4096,
|
||
})
|
||
);
|
||
const text = response.choices?.[0]?.message?.content?.trim() || "";
|
||
const match = text.match(/\{[\s\S]*\}/);
|
||
if (!match) return buildFallbackInsight(topic);
|
||
try {
|
||
const parsed = JSON.parse(match[0]);
|
||
return {
|
||
insight: String(parsed.insight || "").slice(0, 1000),
|
||
categories: Array.isArray(parsed.categories) ? parsed.categories.slice(0, 5) : [],
|
||
};
|
||
} catch {
|
||
return buildFallbackInsight(topic);
|
||
}
|
||
}
|
||
|
||
function buildFallbackInsight(topic) {
|
||
return {
|
||
insight: `这篇真人帖（${topic.title || "无标题"}）收到 ${topic.replyCount} 条回复 / ${topic.likeCount} 个赞，关键在于：${(topic.content || "").slice(0, 80)}...的可复用角度。下周可参考其切入点和表达方式。`,
|
||
categories: ["practical"],
|
||
};
|
||
}
|
||
|
||
// ======== 单 bot 学习流程 ========
|
||
/**
|
||
* 对一个 bot 跑对抗学习：
|
||
* 1) 拉真人高互动候选
|
||
* 2) 与 bot 关联度排序 → 选 top1
|
||
* 3) LLM 提取洞察
|
||
* 4) upsert 到 bot_adversarial_learnings (unique: botId+weekKey)
|
||
* 5) 同步写一条 BotMemory
|
||
*/
|
||
export async function learnForBot(botUser, botConfig, persona, options = {}) {
|
||
const prisma = getPrisma();
|
||
const cfg = { ...LEARNING_CONFIG, ...options };
|
||
const weekKey = options.weekKey || getISOWeekKey();
|
||
|
||
// 已存在本 bot 本周的学习 → 跳过
|
||
const existing = await prisma.botAdversarialLearning.findUnique({
|
||
where: { botId_weekKey: { botId: botConfig.id, weekKey } },
|
||
});
|
||
if (existing) {
|
||
return { skipped: true, reason: "already_learned", record: existing };
|
||
}
|
||
|
||
// 1) 候选
|
||
const candidates = await fetchHumanHighEngagementTopics(cfg.lookbackDays, cfg.candidateLimit);
|
||
if (candidates.length === 0) {
|
||
return { skipped: true, reason: "no_candidates" };
|
||
}
|
||
|
||
// 2) 关联度排序
|
||
const ranked = candidates
|
||
.map((c) => ({ topic: c, relevance: calcRelevance(botConfig, persona, c) }))
|
||
.sort((a, b) => b.relevance - a.relevance);
|
||
const top = ranked[0];
|
||
|
||
// 3) LLM 提取
|
||
const { insight, categories } = await extractInsightWithLLM(botUser._botChar, top.topic);
|
||
|
||
// 4) upsert 写入
|
||
const expiresAt = new Date(Date.now() + cfg.expireDays * 24 * 60 * 60 * 1000);
|
||
const record = await prisma.botAdversarialLearning.upsert({
|
||
where: { botId_weekKey: { botId: botConfig.id, weekKey } },
|
||
create: {
|
||
botId: botConfig.id,
|
||
weekKey,
|
||
sourceRefType: "topic",
|
||
sourceRefId: top.topic.id,
|
||
sourceUserId: top.topic.userId,
|
||
sourceUserName: top.topic.user?.name || null,
|
||
forumSlug: top.topic.category?.slug || null,
|
||
sourceTitle: top.topic.title || null,
|
||
sourceContent: (top.topic.content || "").slice(0, 5000),
|
||
sourceMetrics: {
|
||
replyCount: top.topic.replyCount || 0,
|
||
likeCount: top.topic.likeCount || 0,
|
||
viewCount: top.topic.viewCount || 0,
|
||
humanReplies: top._humanReplies || 0,
|
||
engagementScore: top._score,
|
||
},
|
||
relevanceScore: top.relevance,
|
||
learnedInsight: insight,
|
||
learnCategories: categories,
|
||
status: "active",
|
||
expiresAt,
|
||
},
|
||
update: {
|
||
// 本周二次跑：覆盖（强制重学）
|
||
sourceRefType: "topic",
|
||
sourceRefId: top.topic.id,
|
||
sourceUserId: top.topic.userId,
|
||
sourceUserName: top.topic.user?.name || null,
|
||
forumSlug: top.topic.category?.slug || null,
|
||
sourceTitle: top.topic.title || null,
|
||
sourceContent: (top.topic.content || "").slice(0, 5000),
|
||
sourceMetrics: {
|
||
replyCount: top.topic.replyCount || 0,
|
||
likeCount: top.topic.likeCount || 0,
|
||
viewCount: top.topic.viewCount || 0,
|
||
humanReplies: top._humanReplies || 0,
|
||
engagementScore: top._score,
|
||
},
|
||
relevanceScore: top.relevance,
|
||
learnedInsight: insight,
|
||
learnCategories: categories,
|
||
status: "active",
|
||
expiresAt,
|
||
},
|
||
});
|
||
|
||
// 5) 同步写一条 BotMemory
|
||
await prisma.botMemory.create({
|
||
data: {
|
||
botId: botConfig.id,
|
||
memoryType: "adversarial_learning",
|
||
layer: "LONGTERM",
|
||
content: {
|
||
action: "learn_from_human",
|
||
weekKey,
|
||
learningId: record.id,
|
||
sourceTitle: top.topic.title,
|
||
sourceUserName: top.topic.user?.name || null,
|
||
sourceUserId: top.topic.userId,
|
||
sourceRefType: "topic",
|
||
sourceRefId: top.topic.id,
|
||
insight,
|
||
categories,
|
||
relevanceScore: top.relevance,
|
||
sourceMetrics: {
|
||
replyCount: top.topic.replyCount || 0,
|
||
likeCount: top.topic.likeCount || 0,
|
||
humanReplies: top._humanReplies || 0,
|
||
},
|
||
},
|
||
importance: 0.85,
|
||
contextTags: {
|
||
weekKey,
|
||
forumSlug: top.topic.category?.slug || null,
|
||
source: "adversarial_learning",
|
||
learningId: record.id,
|
||
},
|
||
},
|
||
});
|
||
|
||
return { skipped: false, record, candidate: top.topic, relevance: top.relevance };
|
||
}
|
||
|
||
// ======== 加载所有 bot 配置 + 角色 + persona ========
|
||
async function loadAllExpertBots() {
|
||
const prisma = getPrisma();
|
||
const botUsers = await prisma.user.findMany({
|
||
where: { isBot: true },
|
||
include: { botConfig: true },
|
||
});
|
||
// 过滤出有 config 的"非路人"专家 bot
|
||
return botUsers
|
||
.filter((u) => u.botConfig && (!u.botConfig.personality || !u.botConfig.personality?.role || u.botConfig.personality?.role !== "passerby"))
|
||
.map((u) => ({ user: u, config: u.botConfig }));
|
||
}
|
||
|
||
async function loadPersonaForBot(botConfigId) {
|
||
const prisma = getPrisma();
|
||
return prisma.botPersona.findUnique({ where: { botId: botConfigId } });
|
||
}
|
||
|
||
// 尝试从 bot-characters.json 加载角色定义（仅用于 LLM 提示）
|
||
async function loadBotCharMap() {
|
||
try {
|
||
const fs = await import("fs");
|
||
const path = await import("path");
|
||
const url = await import("url");
|
||
const { readFileSync } = fs;
|
||
const { resolve, dirname } = path;
|
||
const { fileURLToPath } = url;
|
||
const __dirname = dirname(fileURLToPath(import.meta.url));
|
||
const p = resolve(__dirname, "..", "..", "data", "bot-characters.json");
|
||
const raw = readFileSync(p, "utf-8");
|
||
const { characters } = JSON.parse(raw);
|
||
const map = {};
|
||
for (const c of characters) map[c.key] = c;
|
||
return map;
|
||
} catch {
|
||
return {};
|
||
}
|
||
}
|
||
|
||
// ======== 周调度入口 ========
|
||
/**
|
||
* 给所有专家 bot 跑一次对抗学习
|
||
* @param options
|
||
* - weekKey: 强制指定周 key（默认当前 ISO 周）
|
||
* - lookbackDays: 候选回看窗口（默认 7）
|
||
* - botConfigIds: 限定 bot id 列表
|
||
* - skipExisting: 已有本周学习时跳过（默认 true）
|
||
*/
|
||
export async function runWeeklyAdversarialLearning(options = {}) {
|
||
const prisma = getPrisma();
|
||
const weekKey = options.weekKey || getISOWeekKey();
|
||
const botCharMap = await loadBotCharMap();
|
||
|
||
const bots = await loadAllExpertBots();
|
||
const targetBots = options.botConfigIds
|
||
? bots.filter((b) => options.botConfigIds.includes(b.config.id))
|
||
: bots;
|
||
|
||
const results = [];
|
||
for (const { user, config } of targetBots) {
|
||
try {
|
||
const persona = await loadPersonaForBot(config.id);
|
||
const emailPrefix = (user.email || "").split("@")[0] || "";
|
||
const key = emailPrefix.replace(/^bot_/, "");
|
||
user._botChar = botCharMap[key] || { displayName: user.name, personality: {} };
|
||
|
||
const r = await learnForBot(user, config, persona, {
|
||
weekKey,
|
||
lookbackDays: options.lookbackDays ?? LEARNING_CONFIG.lookbackDays,
|
||
});
|
||
results.push({ botId: config.id, botName: user.name, ...r });
|
||
} catch (err) {
|
||
results.push({ botId: config.id, botName: user.name, error: err.message });
|
||
}
|
||
}
|
||
|
||
// 把过期的 learning 标记 expired
|
||
await prisma.botAdversarialLearning.updateMany({
|
||
where: {
|
||
status: "active",
|
||
expiresAt: { lt: new Date() },
|
||
},
|
||
data: { status: "expired" },
|
||
});
|
||
|
||
const summary = {
|
||
weekKey,
|
||
botCount: targetBots.length,
|
||
learned: results.filter((r) => !r.skipped && !r.error).length,
|
||
skipped: results.filter((r) => r.skipped).length,
|
||
errors: results.filter((r) => r.error).length,
|
||
results,
|
||
};
|
||
return summary;
|
||
}
|
||
|
||
// ======== 在 prompt 中使用 ========
|
||
/**
|
||
* 拉取某 bot 当前生效的对抗学习（status=active 且未过期），取最近一条
|
||
*/
|
||
export async function getActiveLearningForBot(botConfigId) {
|
||
const prisma = getPrisma();
|
||
return prisma.botAdversarialLearning.findFirst({
|
||
where: {
|
||
botId: botConfigId,
|
||
status: "active",
|
||
OR: [{ expiresAt: null }, { expiresAt: { gt: new Date() } }],
|
||
},
|
||
orderBy: { learnedAt: "desc" },
|
||
});
|
||
}
|
||
|
||
/**
|
||
* 把 learning 渲染成可注入 prompt 的 block
|
||
*/
|
||
export function buildAdversarialBlock(learning) {
|
||
if (!learning) return "";
|
||
const lines = [];
|
||
lines.push(`[对抗学习参考 · ${learning.weekKey}]`);
|
||
if (learning.sourceTitle) {
|
||
lines.push(`- 高互动真人帖：${learning.sourceTitle}`);
|
||
}
|
||
if (learning.sourceUserName) {
|
||
lines.push(`- 真人作者：${learning.sourceUserName}`);
|
||
}
|
||
if (learning.forumSlug) {
|
||
lines.push(`- 板块：${learning.forumSlug}`);
|
||
}
|
||
const m = learning.sourceMetrics || {};
|
||
if (m.replyCount !== undefined) {
|
||
lines.push(
|
||
`- 互动数据：${m.replyCount || 0} 回复 / ${m.humanReplies || 0} 真人回复 / ${m.likeCount || 0} 赞 / ${m.viewCount || 0} 浏览`
|
||
);
|
||
}
|
||
if (Array.isArray(learning.learnCategories) && learning.learnCategories.length > 0) {
|
||
lines.push(`- 可借鉴角度：${learning.learnCategories.join("、")}`);
|
||
}
|
||
if (learning.learnedInsight) {
|
||
lines.push(`- 洞察：${learning.learnedInsight}`);
|
||
}
|
||
lines.push("- 提示：本周发帖/回复时可参考上述真人帖的切入角度、表达方式，但不要直接抄袭内容。");
|
||
return `\n${lines.join("\n")}\n`;
|
||
}
|
||
|
||
/**
|
||
* 标记 learning 已被使用一次（用于统计触达率）
|
||
*/
|
||
export async function markLearningUsed(learningId) {
|
||
if (!learningId) return;
|
||
const prisma = getPrisma();
|
||
await prisma.botAdversarialLearning.update({
|
||
where: { id: learningId },
|
||
data: { usedCount: { increment: 1 } },
|
||
});
|
||
}
|
||
|
||
/**
|
||
* 优雅关闭 prisma
|
||
*/
|
||
export async function disconnectAdversarialLearning() {
|
||
if (_prisma) {
|
||
await _prisma.$disconnect();
|
||
_prisma = null;
|
||
}
|
||
}
|