https://zerocoding.online/tags/%E5%90%91%E9%87%8F%E6%A3%80%E7%B4%A2/ZeroCodingHugo 0.165.0 & FixIt v1.0.0-alpha.2-mtmrzvotzh-CNhi@zerocoding.online (ZeroCoding)hi@zerocoding.online (ZeroCoding)Thu, 20 Aug 2026 00:00:00 +0800https://zerocoding.online/posts/rag-practice/Thu, 20 Aug 2026 00:00:00 +0800hi@zerocoding.online (ZeroCoding)https://zerocoding.online/posts/rag-practice/AI<p>RAG（检索增强生成）是让大模型「有据可查」的关键手段。落地时，检索质量直接决定了生成效果的上限。</p> <h2 class="heading-element" id="选型三件套"><span>选型三件套</span> <a href="#%e9%80%89%e5%9e%8b%e4%b8%89%e4%bb%b6%e5%a5%97" class="heading-mark"><svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg></a> </h2><ol> <li><strong>Embedding 模型</strong>：中文场景优先选对中文友好的模型，注意向量维度与语种覆盖。</li> <li><strong>向量库</strong>：小规模用 pgvector 即可，规模上来再考虑 Milvus、Qdrant 等专用方案。</li> <li><strong>召回策略</strong>：稠密召回为主，稀疏（BM25）做补充，混合检索通常效果更稳。</li> </ol> <h2 class="heading-element" id="两个常被忽略的优化"><span>两个常被忽略的优化</span> <a href="#%e4%b8%a4%e4%b8%aa%e5%b8%b8%e8%a2%ab%e5%bf%bd%e7%95%a5%e7%9a%84%e4%bc%98%e5%8c%96" class="heading-mark"><svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg></a> </h2><ul> <li><strong>Chunk 切分</strong>：按语义边界切分，保留标题层级，比固定长度效果好很多。</li> <li><strong>重排序</strong>：用 Cross-Encoder 对召回结果二次排序，能明显提升 Top-K 的命中率。</li> </ul>