知识卡片
知识冲突必须回到来源判断
模型参数记忆与取回文本冲突,可能来自过时、错误或任务设定不同。应比较来源、时间及适用条件,无法解决时呈现不确定;模型更自信或偏好上下文不能直接决定哪方真实。
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
- · 检索增强生成
原书出处 · 1 处
- · 查看原书依据
How Do LLMs Deal with Contradictory Information? Sometimes the knowledge captured in the LLM’s internal representations can be contradictory to the content retrieved during the RAG. This can happen due to a multitude of reasons: outdated or incorrect content in the LLM’s training datasets, errors in the user-provided context, or retrieval of incorrect or irrelevant documents during RAG. In these cases, we want the LLM
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:摘录中关于知识冲突必须回到来源判断的定义、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:How Do LLMs Deal with Contradictory Information? Sometimes the knowledge captured in the LLM’s internal representations can be contradictory to the content retrieved during the RAG. This can happen due to a multitude of reasons: outdated or incorrect content in the LLM’s training datasets, errors in the user-provided context, or retrieval of incorrect or irrelevant documents during RAG. In these cases, we want the LLM