📄 中文摘要
许多 RAG 教程遵循相同的流程：将文档分割成块，生成嵌入，存储在向量数据库中，并在用户提问时检索前 K 个结果。然而，当将其部署到真实用户时，情况却大相径庭。经过几个月的构建和迭代，发现这些教程中的几乎所有假设在大规模应用时都不成立。实际有效的做法与教程中所述存在显著差异，本文将揭示这些关键点和遗漏之处。
生产中的 RAG 与教程大相径庭
出处: RAG in production is nothing like the tutorials
发布: 2026年2月25日
📄 English Summary
RAG in production is nothing like the tutorials
Many RAG tutorials follow a similar script: take documents, split them into chunks, generate embeddings, store them in a vector database, and retrieve the top K results when a user asks a question. However, deploying this to real users reveals significant issues. After months of building and iterating on a production RAG pipeline, it became clear that nearly every assumption from those tutorials fails at scale. The article highlights what actually works and the critical omissions from the tutorials.