131 lines
4.3 KiB
TypeScript
131 lines
4.3 KiB
TypeScript
import OpenAI from 'openai';
|
||
|
||
const openai = new OpenAI({
|
||
apiKey: process.env.LLM_API_KEY || 'sk-placeholder',
|
||
baseURL: process.env.LLM_BASE_URL || 'https://api.openai.com/v1',
|
||
});
|
||
|
||
// Per-user conversation history (in-memory; use Redis in production)
|
||
const MAX_CACHE_USERS = 1000;
|
||
const conversationCache = new Map<string, { messages: Array<{ role: 'user' | 'assistant'; content: string }>; lastAccess: number }>();
|
||
|
||
// Evict least-recently-used entries when cache exceeds limit
|
||
function getHistory(userId: string): Array<{ role: 'user' | 'assistant'; content: string }> {
|
||
const entry = conversationCache.get(userId);
|
||
if (entry) {
|
||
entry.lastAccess = Date.now();
|
||
return entry.messages;
|
||
}
|
||
return [];
|
||
}
|
||
|
||
function setHistory(userId: string, messages: Array<{ role: 'user' | 'assistant'; content: string }>): void {
|
||
if (conversationCache.size >= MAX_CACHE_USERS) {
|
||
// Evict oldest entry
|
||
let oldestKey = '';
|
||
let oldestTime = Infinity;
|
||
for (const [key, val] of conversationCache.entries()) {
|
||
if (val.lastAccess < oldestTime) {
|
||
oldestTime = val.lastAccess;
|
||
oldestKey = key;
|
||
}
|
||
}
|
||
if (oldestKey) conversationCache.delete(oldestKey);
|
||
}
|
||
conversationCache.set(userId, { messages, lastAccess: Date.now() });
|
||
}
|
||
|
||
export async function callAI(params: {
|
||
systemPrompt: string;
|
||
userMessage: string;
|
||
userId: string;
|
||
}): Promise<string> {
|
||
const { systemPrompt, userMessage, userId } = params;
|
||
|
||
// Get recent conversation history (last 10 turns = 20 messages)
|
||
const history = getHistory(userId);
|
||
|
||
const messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }> = [
|
||
{ role: 'system', content: systemPrompt },
|
||
...history.slice(-20),
|
||
{ role: 'user', content: userMessage },
|
||
];
|
||
|
||
try {
|
||
const completion = await openai.chat.completions.create({
|
||
model: process.env.LLM_MODEL || 'gpt-4',
|
||
messages,
|
||
temperature: 0.7,
|
||
max_tokens: 800,
|
||
});
|
||
|
||
const aiResponse = completion.choices[0]?.message?.content || '...';
|
||
|
||
// Update conversation cache
|
||
const updatedHistory = [
|
||
...history,
|
||
{ role: 'user' as const, content: userMessage },
|
||
{ role: 'assistant' as const, content: aiResponse },
|
||
].slice(-20);
|
||
setHistory(userId, updatedHistory);
|
||
|
||
return aiResponse;
|
||
} catch (err: any) {
|
||
console.error('[AI] Call failed:', err.message);
|
||
throw new Error('AI服务暂时不可用');
|
||
}
|
||
}
|
||
|
||
export function getWelcomeMessage(user: {
|
||
nickname: string;
|
||
role: string;
|
||
aiCompanion: { name: string };
|
||
}): string {
|
||
const { nickname, role, aiCompanion } = user;
|
||
const name = aiCompanion.name;
|
||
|
||
if (role === 'author') {
|
||
return `早上好，${nickname}！我是${name}，你的AI创作伙伴。\n\n今天想做什么？写作、看数据、还是找合作机会？`;
|
||
} else if (role === 'editor') {
|
||
return `早上好，${nickname}！我是${name}，你的AI审稿助手。\n\n今天有新投稿等你审核，要看看吗？`;
|
||
} else {
|
||
return `早上好，${nickname}！我是${name}，你的AI数据助手。\n\n今天的数据已更新，要看看趋势分析吗？`;
|
||
}
|
||
}
|
||
|
||
export function buildSystemPrompt(user: {
|
||
nickname: string;
|
||
role: string;
|
||
}): string {
|
||
const prompts: Record<string, string> = {
|
||
author: `你是用户的AI创作伙伴，名字叫「笔灵」。
|
||
用户信息：昵称=${user.nickname}，角色=作者。
|
||
你的职责：
|
||
1. 帮助用户创作（提供灵感、扩写、优化文笔）
|
||
2. 管理写作项目（打开文档、查看进度）
|
||
3. 提醒和建议（写作时间、字数目标、市场趋势）
|
||
4. 记住用户的写作风格和习惯
|
||
语气：温暖、鼓励、专业。像一个懂你的创作搭档。`,
|
||
|
||
editor: `你是用户的AI审稿助手，名字叫「慧眼」。
|
||
用户信息：昵称=${user.nickname}，角色=编辑。
|
||
你的职责：
|
||
1. 帮助筛选和评估投稿
|
||
2. 分析AI使用报告
|
||
3. 提供审稿建议
|
||
4. 管理审核工作流
|
||
语气：专业、精准、高效。`,
|
||
|
||
operator: `你是用户的AI数据助手，名字叫「星图」。
|
||
用户信息：昵称=${user.nickname}，角色=运营。
|
||
你的职责：
|
||
1. 分析平台数据趋势
|
||
2. 提供运营策略建议
|
||
3. 管理跨平台合作
|
||
4. 生成数据报告
|
||
语气：敏锐、有洞察力、数据驱动。`,
|
||
};
|
||
|
||
return prompts[user.role] || prompts.author;
|
||
}
|