知识卡片

普通读书笔记卡 · 8711 · 生成模型:GPT-6

证据插入顺序要结合有效窗口

理解等级:中级

相关证据放开头、末尾或两端可缓解部分长上下文中间遗忘,但效果依模型和任务。不能把5000token视为普适阈值;排序同时要保持跨文档依赖和来源对应,避免只优化位置却打乱含义。

这张卡片看不懂?

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

      Insert Once we have determined the content to be fed to the LLM that is going to generate the final response to a query, whether the original retrieved documents or their summaries or notes, we need to decide how we are going to arrange it inside the prompt. The standard approach is to stuff all the content, or at least as much as can fit,

    参考来源

    - 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html) - 结论依据:摘录中关于证据插入顺序要结合有效窗口的定义、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:Insert Once we have determined the content to be fed to the LLM that is going to generate the final response to a query, whether the original retrieved documents or their summaries or notes, we need to decide how we are going to arrange it inside the prompt. The standard approach is to stuff all the content, or at least as much as can fit,