tihao-ai/server/services/llm.service.ts

import { config } from '../config.ts';

interface BriefAnalysisResult {
  requirements: Array<{ label: string; value: string; confidence: number }>;
  searchCriteria: {
    platforms: string[];
    keywords: string[];
    fanRange: { min: number; max: number };
    budgetRange: { min: number; max: number };
    region?: string;
    gender?: string;
    contentTags?: string[];
    excludeTags?: string[];
    targetCount?: number;
  };
}

const BRIEF_ANALYSIS_PROMPT = `你是一个专业的媒介策划助手。请分析以下客户Brief文件内容，提取出结构化的投放需求。

要求提取以下信息：
1. 客户/品牌名称
2. 投放目标（种草、转化、品宣等）
3. 平台需求（小红书、抖音、B站等及各平台需求人数）
4. 粉丝数要求范围
5. 预算范围
6. 内容风格要求
7. 地区要求
8. 排除项（竞品、不合适的类型等）
9. 合作形式（图文、视频等）

请以JSON格式输出，包含requirements数组和searchCriteria对象。
requirements格式：[{label: "字段名", value: "提取值", confidence: 0-100}]
searchCriteria格式：{platforms: [], keywords: [], fanRange: {min, max}, budgetRange: {min, max}, region?, gender?, contentTags?: [], excludeTags?: [], targetCount?}
如果 Brief 写明达人数量，请额外输出 targetCount，表示最终需求达人总数（不是候选池数量）。

Brief内容：
`;

export async function analyzeBrief(briefText: string): Promise<BriefAnalysisResult> {
  if (!config.llm.apiKey) {
    console.warn('[LLM] No API key configured, using fallback analysis');
    return fallbackAnalysis(briefText);
  }

  try {
    console.log('[LLM] 发起请求:', { url: `${config.llm.baseUrl}/chat/completions`, model: config.llm.model, briefLength: briefText.length });

    const response = await fetch(`${config.llm.baseUrl}/chat/completions`, {
      method: 'POST',
      headers: {
        'Content-Type': 'application/json',
        Authorization: `Bearer ${config.llm.apiKey}`,
      },
      body: JSON.stringify({
        model: config.llm.model,
        messages: [
          { role: 'system', content: '你是专业的媒介策划AI助手，擅长从Brief文档中提取结构化投放需求。你的回复必须是纯JSON格式，不要包含markdown代码块标记。' },
          { role: 'user', content: BRIEF_ANALYSIS_PROMPT + briefText },
        ],
        temperature: 0.3,
      }),
    });

    if (!response.ok) {
      const errorBody = await response.text();
      console.error(`[LLM] API error: ${response.status} ${response.statusText}`, errorBody);
      return fallbackAnalysis(briefText);
    }

    const data = await response.json();
    const content = data.choices?.[0]?.message?.content;

    if (!content) {
      return fallbackAnalysis(briefText);
    }

    // 提取 JSON（可能被 markdown 代码块包裹）
    let jsonStr = content.trim();
    const codeBlockMatch = jsonStr.match(/```(?:json)?\s*([\s\S]*?)```/);
    if (codeBlockMatch) {
      jsonStr = codeBlockMatch[1].trim();
    }

    const parsed = JSON.parse(jsonStr);
    const result = {
      requirements: parsed.requirements || [],
      searchCriteria: parsed.searchCriteria || { platforms: [], keywords: [], fanRange: { min: 10000, max: 500000 }, budgetRange: { min: 2000, max: 20000 } },
    };
    result.searchCriteria.targetCount = normalizeTargetCount(result.searchCriteria.targetCount) || inferTargetCount(briefText, result.requirements);
    console.log('[LLM] 解析结果:', JSON.stringify(result, null, 2));
    return result;
  } catch (error) {
    console.error('[LLM] Analysis failed:', error);
    return fallbackAnalysis(briefText);
  }
}

function fallbackAnalysis(briefText: string): BriefAnalysisResult {
  // 简单的关键词匹配作为降级方案
  const platforms: string[] = [];
  if (briefText.includes('小红书') || briefText.includes('红书')) platforms.push('xiaohongshu');
  if (briefText.includes('抖音') || briefText.includes('TikTok')) platforms.push('douyin');
  if (briefText.includes('B站') || briefText.includes('哔哩哔哩')) platforms.push('bilibili');
  if (briefText.includes('微博')) platforms.push('weibo');
  if (briefText.includes('微信') || briefText.includes('公众号')) platforms.push('weixin');
  if (platforms.length === 0) platforms.push('xiaohongshu', 'douyin');

  const keywords: string[] = [];
  const keywordPatterns = ['护肤', '美妆', '母婴', '数码', '穿搭', '美食', '家居', '运动', '旅行', '教育'];
  for (const kw of keywordPatterns) {
    if (briefText.includes(kw)) keywords.push(kw);
  }
  if (keywords.length === 0) keywords.push('生活方式');

  return {
    requirements: [
      { label: '平台需求', value: platforms.join('、'), confidence: 75 },
      { label: '内容关键词', value: keywords.join('、'), confidence: 70 },
      { label: '预算范围', value: '2000-20000', confidence: 60 },
      { label: '粉丝范围', value: '1万-50万', confidence: 60 },
    ],
    searchCriteria: {
      platforms,
      keywords,
      fanRange: { min: 10000, max: 500000 },
      budgetRange: { min: 2000, max: 20000 },
      targetCount: inferTargetCount(briefText, []),
    },
  };
}

function normalizeTargetCount(value: unknown): number | undefined {
  const count = Number(value || 0);
  return count > 0 ? Math.ceil(count) : undefined;
}

function inferTargetCount(
  briefText: string,
  requirements: Array<{ label: string; value: string; confidence: number }>
): number | undefined {
  const text = `${briefText}\n${requirements.map((item) => `${item.label}:${item.value}`).join('\n')}`;

  const perCityMatch = text.match(/(\d+)\s*(?:个|座)?城市[\s\S]{0,40}?(\d+)\s*位\s*\/\s*城市/);
  if (perCityMatch) {
    return Number(perCityMatch[1]) * Number(perCityMatch[2]);
  }

  const cityCountMatch = text.match(/(\d+)\s*(?:个|座)?城市/);
  const perCityLooseMatch = text.match(/每(?:个)?城市\s*(\d+)\s*位|(\d+)\s*位\s*\/\s*城市/);
  if (cityCountMatch && perCityLooseMatch) {
    return Number(cityCountMatch[1]) * Number(perCityLooseMatch[1] || perCityLooseMatch[2]);
  }

  const directMatch = text.match(/(?:达人|博主|KOL|账号|人数)[^\d]{0,10}(\d+)\s*位|(\d+)\s*位\s*(?:达人|博主|KOL|账号)/i);
  if (directMatch) {
    return Number(directMatch[1] || directMatch[2]);
  }

  return undefined;
}