知识卡片
检索增强不能保证无幻觉
提供相关证据可降低依赖参数记忆的需求,却仍有检索缺失、噪声、长上下文利用不当和无依据补写。应分别诊断证据是否找到、是否充分以及生成是否忠实,不能仅因接入知识库就省去这些检查。
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
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
RAG is not a panacea for the hallucination problem for the following reasons: Feeding ground-truth factual data in the prompt reduces hallucinations but does not eliminate them completely, especially when the context is large. Using RAG shifts the bottleneck toward the retrieval process. If the retrieval process is not able to return the relevant data needed, the model may do worse at the task than
参考来源
- 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html)
- 结论依据:摘录中关于检索增强不能保证无幻觉的定义、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:RAG is not a panacea for the hallucination problem for the following reasons: Feeding ground-truth factual data in the prompt reduces hallucinations but does not eliminate them completely, especially when the context is large. Using RAG shifts the bottleneck toward the retrieval process. If the retrieval process is not able to return the relevant data needed, the model may do worse at the task than