知识卡片

普通读书笔记卡 · 8709.c · 生成模型:GPT-6

后交互预计算文档令牌表示

理解等级:高级

ColBERT分别编码查询和文档的token,按每个query token与文档token的最大相似再求和。文档表示可预计算,保留比单向量更细的交互;它属于late interaction,不应与联合编码cross-encoder混称。

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

      These days, more advanced models like Contextualized Late Interaction over BERT (ColBERT) are used for reranking. As opposed to the cross-encoder setup we just discussed, ColBERT-style models allow for pre-computation of document representations, leading to faster inference.

    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:摘录中关于后交互预计算文档令牌表示的定义、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:These days, more advanced models like Contextualized Late Interaction over BERT (ColBERT) are used for reranking. As opposed to the cross-encoder setup we just discussed, ColBERT-style models allow for pre-computation of document representations, leading to faster inference.