知识卡片
级联先尝试便宜模型再升级
小模型结果满足验收阈值就接受,否则依次升级更大模型。适合多数请求能在前级完成的分布;困难请求会累计多轮成本与延迟,不能只比较最终大模型价格。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
LLM Cascades While using the state-of-the-art LLM for processing all our inputs is an option, realistically this might be cost-prohibitive or latency sensitive. To optimize costs while keeping performance standards high, we could leverage multiple LLMs, organized in a cascade architecture. Let’s illustrate LLM cascades. Consider you have an application using three LLMs: one small, one medium, and one large, as illustrated in Figure 13-1.
参考来源
- 位置:《Designing Large Language Model Applications》第13章《设计模式与系统架构》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch13.html)
- 结论依据:摘录中关于级联先尝试便宜模型再升级的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:LLM Cascades While using the state-of-the-art LLM for processing all our inputs is an option, realistically this might be cost-prohibitive or latency sensitive. To optimize costs while keeping performance standards high, we could leverage multiple LLMs, organized in a cascade architecture. Let’s illustrate LLM cascades. Consider you have an application using three LLMs: one small, one medium, and one large, as illustrated in Figure 13-1.