[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-config":3,"primary-menu":6,"cat-rag":85,"catposts-rag":99,"post-detail-post-rag-embedding-027aa771":208},{"config":4,"socialLinks":5},null,[],{"metadata":7,"spec":10,"menuItems":20},{"name":8,"creationTimestamp":9},"primary","2026-09-12T06:13:55.024154Z",{"displayName":11,"menuItems":12},"主菜单",[13,14,15,16,17,18,19],"88c3f10b-321c-4092-86a8-70db00251b74","menu-item-izrtmxer","c4c814d1-0c2c-456b-8c96-4864965fee94","menu-item-column","menu-item-tools","menu-item-zsxq","menu-item-ocrxlczu",[21,30,40,49,58,67,76],{"displayName":22,"metadata":23,"spec":25,"status":28,"children":29},"首页",{"name":13,"creationTimestamp":24},"2026-09-12T06:13:54.422830Z",{"displayName":22,"href":26,"priority":27,"menuName":8},"\u002F",0,{"displayName":22,"href":26},[],{"displayName":31,"metadata":32,"spec":34,"status":38,"children":39},"Java面试八股",{"name":14,"creationTimestamp":33},"2026-09-13T09:42:02.854891Z",{"displayName":31,"href":35,"target":36,"priority":37,"menuName":8},"\u002Fjava-interview","_self",1,{"displayName":31,"href":35},[],{"displayName":41,"metadata":42,"spec":44,"status":47,"children":48},"文章",{"name":15,"creationTimestamp":43},"2026-09-12T06:13:54.457307Z",{"displayName":41,"href":45,"priority":46,"menuName":8},"\u002Farchives",2,{"displayName":41,"href":45},[],{"displayName":50,"metadata":51,"spec":53,"status":56,"children":57},"项目实战",{"name":16,"creationTimestamp":52},"2026-09-14T00:35:08.138170Z",{"displayName":50,"href":54,"target":36,"priority":55,"menuName":8},"\u002Fcolumn",3,{"displayName":50,"href":54},[],{"displayName":59,"metadata":60,"spec":62,"status":65,"children":66},"在线工具",{"name":17,"creationTimestamp":61},"2026-09-14T00:35:08.171474Z",{"displayName":59,"href":63,"target":36,"priority":64,"menuName":8},"\u002Ftools",4,{"displayName":59,"href":63},[],{"displayName":68,"metadata":69,"spec":71,"status":74,"children":75},"知识星球",{"name":18,"creationTimestamp":70},"2026-09-14T00:35:08.199145Z",{"displayName":68,"href":72,"target":36,"priority":73,"menuName":8},"\u002Fzsxq",5,{"displayName":68,"href":72},[],{"displayName":77,"metadata":78,"spec":80,"status":83,"children":84},"特价云服务器",{"name":19,"creationTimestamp":79},"2026-09-25T15:30:59.201614Z",{"displayName":77,"href":81,"target":36,"priority":82,"menuName":8},"\u002Fserver",6,{"displayName":77,"href":81},[],{"metadata":86,"spec":91,"status":97,"postCount":73},{"name":87,"creationTimestamp":88,"labels":89},"rag","2026-09-13T09:13:25.943986Z",{"haloweb.section":90},"interview",{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},"RAG 面试题","","RAG 检索增强生成面试题：向量检索、召回与重排、知识库构建、效果评测与常见坑。",17,false,{"permalink":98,"postCount":73,"visiblePostCount":73},"\u002Fcategories\u002Frag",[100,124,145,166,187],{"metadata":101,"spec":104,"status":113,"categories":118},{"name":102,"creationTimestamp":103},"post-rag-optimize-3be551ba","2026-09-13T10:01:17.497938Z",{"title":105,"slug":106,"cover":93,"excerpt":107,"publishTime":109,"categories":110,"tags":111,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},"RAG 系统的优化手段？","rag-optimize",{"raw":93,"autoGenerate":108},true,"2026-09-13T10:01:17.709819Z",[87],[],"PUBLIC",{"permalink":114,"excerpt":115,"lastModifyTime":116,"phase":117},"\u002Farchives\u002Frag-optimize","一、检索优化 分块优化：合理的块大小和重叠。 Embedding 优化：用更适合的模型，微调。 混合检索：向量 + 关键词，提升召回。 重排（Rerank）：用 Cross-Encoder 对检索结果重排序，提升精度。 查询改写：用 LLM 改写用户问题，提升检索效果。 多查询：生成多个相关问题分别","2026-09-16T09:16:46.845496Z","PUBLISHED",[119],{"metadata":120,"spec":122,"status":123,"postCount":73},{"name":87,"creationTimestamp":88,"labels":121},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"metadata":125,"spec":128,"status":135,"categories":139},{"name":126,"creationTimestamp":127},"post-rag-evaluation-3343ec81","2026-09-13T10:01:17.338928Z",{"title":129,"slug":130,"cover":93,"excerpt":131,"publishTime":132,"categories":133,"tags":134,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},"如何评估 RAG 系统的效果？","rag-evaluation",{"raw":93,"autoGenerate":108},"2026-09-13T10:01:17.382544Z",[87],[],{"permalink":136,"excerpt":137,"lastModifyTime":138,"phase":117},"\u002Farchives\u002Frag-evaluation","一、评估维度 检索质量：检索到的文档是否相关。 生成质量：回答是否准确、完整、流畅。 端到端效果：整体回答是否满足用户。 二、检索评估指标 Recall（召回率）：相关文档中被检索到的比例。 Precision（精确率）：检索到的文档中相关的比例。 MRR：第一个相关文档的排名倒数。 NDCG：考虑","2026-09-16T09:17:11.351718Z",[140],{"metadata":141,"spec":143,"status":144,"postCount":73},{"name":87,"creationTimestamp":88,"labels":142},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"metadata":146,"spec":149,"status":156,"categories":160},{"name":147,"creationTimestamp":148},"post-rag-chunking-bbf97db4","2026-09-13T10:01:17.143825Z",{"title":150,"slug":151,"cover":93,"excerpt":152,"publishTime":153,"categories":154,"tags":155,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},"RAG 中的文档分块策略？","rag-chunking",{"raw":93,"autoGenerate":108},"2026-09-13T10:01:17.186951Z",[87],[],{"permalink":157,"excerpt":158,"lastModifyTime":159,"phase":117},"\u002Farchives\u002Frag-chunking","一、为什么要分块 文档太长无法直接向量化（Embedding 有 token 限制），且 LLM 上下文有限。分块后每块独立检索。 二、分块方法 1. 固定长度分块 按 token 数切分，如每块 500 token，重叠 50 token。 - 优点：简单。 - 缺点：可能切断语义。 2. 按结构","2026-09-16T09:16:42.573515Z",[161],{"metadata":162,"spec":164,"status":165,"postCount":73},{"name":87,"creationTimestamp":88,"labels":163},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"metadata":167,"spec":170,"status":177,"categories":181},{"name":168,"creationTimestamp":169},"post-rag-embedding-027aa771","2026-09-13T10:01:16.978069Z",{"title":171,"slug":172,"cover":93,"excerpt":173,"publishTime":174,"categories":175,"tags":176,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},"RAG 中的 Embedding 和向量数据库？","rag-embedding",{"raw":93,"autoGenerate":108},"2026-09-13T10:01:17.022874Z",[87],[],{"permalink":178,"excerpt":179,"lastModifyTime":180,"phase":117},"\u002Farchives\u002Frag-embedding","一、Embedding（嵌入） 把文本转换为高维向量（如 768\u002F1024 维），语义相近的文本向量距离近。 二、向量数据库 专门存储和检索向量的数据库，支持相似性搜索。 Milvus：开源，高性能，分布式。 Pinecone：云服务。 Weaviate：开源，支持混合检索。 Chroma：轻量，适","2026-09-16T09:16:46.368492Z",[182],{"metadata":183,"spec":185,"status":186,"postCount":73},{"name":87,"creationTimestamp":88,"labels":184},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"metadata":188,"spec":191,"status":198,"categories":202},{"name":189,"creationTimestamp":190},"post-rag-what-c4de36f1","2026-09-13T10:01:16.820767Z",{"title":192,"slug":193,"cover":93,"excerpt":194,"publishTime":195,"categories":196,"tags":197,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},"什么是 RAG？","rag-what",{"raw":93,"autoGenerate":108},"2026-09-13T10:01:16.868881Z",[87],[],{"permalink":199,"excerpt":200,"lastModifyTime":201,"phase":117},"\u002Farchives\u002Frag-what","一、定义 RAG（Retrieval-Augmented Generation，检索增强生成）把外部知识检索和大语言模型生成结合起来。先从知识库检索相关文档，再让 LLM 基于检索结果生成回答。 二、解决的问题 LLM 知识截止，不知道最新信息。 LLM 幻觉，编造内容。 LLM 不知道企业私有知识","2026-09-16T09:16:45.934790Z",[203],{"metadata":204,"spec":206,"status":207,"postCount":73},{"name":87,"creationTimestamp":88,"labels":205},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"metadata":209,"spec":210,"status":214,"categories":215,"content":221},{"name":168,"creationTimestamp":169},{"title":171,"slug":172,"cover":93,"excerpt":211,"publishTime":174,"categories":212,"tags":213,"priority":27,"pinned":96,"deleted":96,"visible":112,"allowComment":108},{"raw":93,"autoGenerate":108},[87],[],{"permalink":178,"excerpt":179,"lastModifyTime":180,"phase":117},[216],{"metadata":217,"spec":219,"status":220,"postCount":73},{"name":87,"creationTimestamp":88,"labels":218},{"haloweb.section":90},{"displayName":92,"slug":87,"cover":93,"description":94,"priority":95,"hideFromList":96,"preventParentPostCascadeQuery":96},{"permalink":98,"postCount":73,"visiblePostCount":73},{"content":222,"raw":222},"\u003Ch2>一、Embedding（嵌入）\u003C\u002Fh2>\n\u003Cp>把文本转换为高维向量（如 768\u002F1024 维），语义相近的文本向量距离近。\u003C\u002Fp>\n\u003Ch2>二、向量数据库\u003C\u002Fh2>\n\u003Cp>专门存储和检索向量的数据库，支持相似性搜索。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Milvus\u003C\u002Fstrong>：开源，高性能，分布式。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Pinecone\u003C\u002Fstrong>：云服务。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Weaviate\u003C\u002Fstrong>：开源，支持混合检索。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Chroma\u003C\u002Fstrong>：轻量，适合原型。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>FAISS\u003C\u002Fstrong>：Meta 开源的向量检索库，非数据库。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>三、相似度计算\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>余弦相似度：方向相近，与长度无关，最常用。\u003C\u002Fli>\n\u003Cli>内积：考虑长度。\u003C\u002Fli>\n\u003Cli>欧氏距离：空间距离。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>四、检索方式\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>向量检索\u003C\u002Fstrong>：语义相似，能找到同义表达。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>关键词检索（BM25）\u003C\u002Fstrong>：精确词匹配。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>混合检索\u003C\u002Fstrong>：向量 + 关键词，召回更全面。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>五、Embedding 模型\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>OpenAI text-embedding-3。\u003C\u002Fli>\n\u003Cli>BGE（智源）：中文效果好。\u003C\u002Fli>\n\u003Cli>M3E、GTE：中文开源。\u003C\u002Fli>\n\u003Cli>选择时考虑语言、维度、性能。\u003C\u002Fli>\n\u003C\u002Ful>"]