https://index.zshipu.com/ai002/tags/PDF%E8%A7%A3%E6%9E%90/ Recent content in PDF解析 on 知识铺的博客 Hugo -- gohugo.io zh-CN Fri, 04 Sep 2026 00:00:00 +0000 https://index.zshipu.com/ai002/post/20260904/RAG%E7%AE%A1%E9%81%93%E9%87%8C%E7%9C%9F%E6%AD%A3%E5%86%B3%E5%AE%9A%E4%B8%8A%E9%99%90%E7%9A%84%E7%8E%AF%E8%8A%82%E4%B8%8D%E6%98%AF%E6%A8%A1%E5%9E%8B%E8%80%8C%E6%98%AF%E6%91%84%E5%8F%96/ Fri, 04 Sep 2026 00:00:00 +0000 https://index.zshipu.com/ai002/post/20260904/RAG%E7%AE%A1%E9%81%93%E9%87%8C%E7%9C%9F%E6%AD%A3%E5%86%B3%E5%AE%9A%E4%B8%8A%E9%99%90%E7%9A%84%E7%8E%AF%E8%8A%82%E4%B8%8D%E6%98%AF%E6%A8%A1%E5%9E%8B%E8%80%8C%E6%98%AF%E6%91%84%E5%8F%96/ 末端调优无法弥补摄取阶段的缺陷 多数团队调试RAG效果不佳时选择更换模型、调整prompt或提高top k，但答案质量几乎不变，因为问题早在查询到达前就已产生。检索系统只能返回已索引的内容，一切下游结果都继承自摄取步骤，没有重排序能修复被打乱的chunk。 这个现象在实际项目中反复出现