https://www.xiaoxiangtao.cn/zh/tags/rag/Recent content in RAG on Xiangtao's logHugozh-Hans15989119660@163.com (Xiangtao XIAO)15989119660@163.com (Xiangtao XIAO)Sun, 08 Mar 2026 00:00:00 +0000https://www.xiaoxiangtao.cn/zh/posts/learning-notes-of-rag/Sun, 08 Mar 2026 00:00:00 +000015989119660@163.com (Xiangtao XIAO)https://www.xiaoxiangtao.cn/zh/posts/learning-notes-of-rag/<h1 id="rag">RAG</h1> <p>RAG (Retrieval-Augmented Generation) 是一种结合信息检索与文本生成的AI架构，通过检索外部知识库来增强LLM的回答能力。</p> <p><strong>经典论文：</strong></p> <ul> <li>Lewis, P., et al. (2020). &ldquo;Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.&rdquo; NeurIPS. [<a href="https://arxiv.org/abs/2005.11401">论文链接</a>] - RAG开山之作</li> <li>Guu, K., et al. (2020). &ldquo;REALM: Retrieval-Augmented Language Model Pre-Training.&rdquo; ICML. [<a href="https://arxiv.org/abs/2002.08909">论文链接</a>]</li> <li>Izacard, G., &amp; Grave, É. (2021). &ldquo;Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering.&rdquo; EACL. [<a href="https://arxiv.org/abs/2007.01282">论文链接</a>] - RAG的另一种实现</li> </ul> <h2 id="1-context-window上下文窗口">1. Context Window（上下文窗口）</h2> <h3 id="概念">概念</h3> <p>Context Window 指的是 <strong>大语言模型在一次推理时能够读取和理解的最大文本长度</strong>。</p> <p><strong>经典论文：</strong></p> <ul> <li>Dai, Z., et al. (2019). &ldquo;Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context.&rdquo; ACL. [<a href="https://arxiv.org/abs/1901.02860">论文链接</a>] - 突破固定上下文长度的开创性工作</li> </ul> <p>通常包含：</p>