LLM API成本Optimize: 省钱80%的实战Strategy

understand您的账单

多级缓存架构

class SmartCache {
  constructor() {
    // L1: 精确匹配缓存(内存)
    this.exactCache = new LRUCache({ max: 1000, ttl: 3600000 });
    
    // L2: 语义相似缓存(Redis + 向量Data库)
    this.semanticCache = new SemanticCache({
      threshold: 0.95,  // 相似度阈值
      maxResults: 5
    });
    
    // L3: 模板缓存
    this.templateCache = new Map();
  }

  async get(prompt, options = {}) {
    // 1. 检查精确匹配
    const exactKey = this.hashPrompt(prompt);
    const exact = this.exactCache.get(exactKey);
    if (exact) {
      this.metrics.recordHit('exact');
      return exact;
    }
    
    // 2. 检查语义相似
    if (options.allowSemantic) {
      const similar = await this.semanticCache.search(prompt);
      if (similar && similar.score > 0.95) {
        this.metrics.recordHit('semantic');
        return similar.response;
      }
    }
    
    // 3. 检查模板匹配
    const template = this.matchTemplate(prompt);
    if (template) {
      const response = await this.fillTemplate(template, prompt);
      this.metrics.recordHit('template');
      return response;
    }
    
    // 缓存未命中
    this.metrics.recordMiss();
    return null;
  }

  async set(prompt, response, metadata = {}) {
    // 存储到多级缓存
    const key = this.hashPrompt(prompt);
    
    // L1: 精确匹配
    this.exactCache.set(key, response);
    
    // L2: 语义缓存
    if (metadata.cacheable !== false) {
      await this.semanticCache.add(prompt, response, metadata);
    }
    
    // Analyze是否can提取模板
    this.analyzeForTemplate(prompt, response);
  }
}