📄 中文摘要
发布了一本118页的手册，旨在分享在构建RAG系统过程中获得的经验和模式。手册涵盖了多个关键问题，包括向量搜索返回“足够接近”的结果而非精确匹配，提供了同时进行语义和关键词搜索的混合检索方法；文档切分时出现奇怪分割的情况，介绍了语义切分、基于抽象语法树的代码感知切分以及保持上下文完整的父子结构；评估检索质量的框架，能够在没有手动标记测试数据的情况下进行评估。这些内容为正在构建RAG的开发者提供了实用的解决方案和指导。
从100多个生产RAG部署中获得的经验教训（免费118页手册）
📄 English Summary
What we learned from 100+ production RAG deployments (free 118-page handbook)
A 118-page handbook has been published to share insights gained from building RAG systems. It addresses several critical issues faced during development, such as vector search returning 'close enough' results instead of exact matches, introducing hybrid retrieval methods that run semantic and keyword searches in parallel. It also tackles the problem of odd document chunking by discussing semantic chunking, code-aware chunking using Abstract Syntax Trees (ASTs), and parent-child structures that maintain context. Furthermore, it presents evaluation frameworks that assess retrieval quality without the need for manually labeled test data. This resource offers practical solutions and guidance for developers currently building RAG systems.