1:"$Sreact.fragment" 2:I[56271,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],"ThemeProviders"] 3:I[41402,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],""] 4:I[91367,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],"KBarSearchProvider"] 5:I[52619,["2619","static/chunks/2619-eccd988d1e161df3.js","1356","static/chunks/1356-354ecf51aaaf5203.js","6909","static/chunks/app/blog/%5B...slug%5D/page-d416e20a09ba8ecb.js"],""] 6:I[51350,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],"KBarButton"] 10:I[57150,[],""] :HL["/_next/static/css/29ff447a95e95257.css","style"] :HL["/_next/static/css/77af622156c5e021.css","style"] :HL["/_next/static/css/ddf3d904beac0ed7.css","style"] 0:{"P":null,"b":"FUTryh0Sm2lrVLTvF5oMH","p":"","c":["","blog","rag-vs-agent-retrieval-decision-framework"],"i":false,"f":[[["",{"children":["blog",{"children":[["slug","rag-vs-agent-retrieval-decision-framework","c"],{"children":["__PAGE__",{}]}]}]},"$undefined","$undefined",true],["",["$","$1","c",{"children":[[["$","link","0",{"rel":"stylesheet","href":"/_next/static/css/29ff447a95e95257.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","link","1",{"rel":"stylesheet","href":"/_next/static/css/77af622156c5e021.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}]],["$","html",null,{"lang":"zh-cn","className":"scroll-smooth","suppressHydrationWarning":true,"children":[["$","link",null,{"rel":"apple-touch-icon","sizes":"76x76","href":"/static/favicons/apple-touch-icon.png"}],["$","link",null,{"rel":"icon","type":"image/png","sizes":"32x32","href":"/static/favicons/favicon-32x32.png"}],["$","link",null,{"rel":"icon","type":"image/png","sizes":"16x16","href":"/static/favicons/favicon-16x16.png"}],["$","link",null,{"rel":"manifest","href":"/static/favicons/site.webmanifest"}],["$","link",null,{"rel":"mask-icon","href":"/static/favicons/safari-pinned-tab.svg","color":"#5bbad5"}],["$","meta",null,{"name":"msapplication-TileColor","content":"#4f46e5"}],["$","meta",null,{"name":"theme-color","media":"(prefers-color-scheme: light)","content":"#ffffff"}],["$","meta",null,{"name":"theme-color","media":"(prefers-color-scheme: dark)","content":"#0f172a"}],["$","link",null,{"rel":"alternate","type":"application/rss+xml","href":"/feed.xml"}],["$","body",null,{"className":"bg-white pl-[calc(100vw-100%)] text-black antialiased dark:bg-gray-950 dark:text-white","children":["$","$L2",null,{"children":[["$undefined","$undefined","$undefined",["$","$L3",null,{"async":true,"defer":true,"src":"https://analytics.umami.is/script.js"}],"$undefined","$undefined"],["$","section",null,{"className":"mx-auto max-w-3xl px-4 sm:px-6 xl:max-w-5xl xl:px-0","children":[["$","$L4",null,{"kbarConfig":{"searchDocumentsPath":"/search.json"},"children":[["$","header",null,{"className":"flex items-center w-full bg-white dark:bg-gray-950 justify-between py-10","children":[["$","$L5",null,{"className":"break-words","href":"/","aria-label":"TechLog","children":["$","div",null,{"className":"flex items-center justify-between","children":[["$","div",null,{"className":"mr-3","children":["$","svg",null,{"xmlns":"http://www.w3.org/2000/svg","xmlnsXlink":"http://www.w3.org/1999/xlink","width":53.87,"height":43.61,"viewBox":"344.564 330.278 111.737 91.218","children":[["$","defs",null,{"children":[["$","linearGradient",null,{"id":"logo_svg__b","x1":420.97,"x2":420.97,"y1":331.28,"y2":418.5,"gradientUnits":"userSpaceOnUse","children":[["$","stop",null,{"offset":"0%","style":{"stopColor":"#06b6d4","stopOpacity":1}}],["$","stop",null,{"offset":"100%","style":{"stopColor":"#67e8f9","stopOpacity":1}}]]}],["$","linearGradient",null,{"id":"logo_svg__d","x1":377.89,"x2":377.89,"y1":331.28,"y2":418.5,"gradientUnits":"userSpaceOnUse","children":[["$","stop",null,{"offset":"0%","style":{"stopColor":"#06b6d4","stopOpacity":1}}],["$","stop",null,{"offset":"100%","style":{"stopColor":"#67e8f9","stopOpacity":1}}]]}],["$","path",null,{"id":"logo_svg__a","d":"M453.3 331.28v28.57l-64.66 58.65v-30.08z"}],["$","path",null,{"id":"logo_svg__c","d":"M410.23 331.28v28.57l-64.67 58.65v-30.08z"}]]}],["$","use",null,{"xlinkHref":"#logo_svg__a","fill":"url(#logo_svg__b)"}],["$","use",null,{"xlinkHref":"#logo_svg__c","fill":"url(#logo_svg__d)"}]]}]}],["$","div",null,{"className":"hidden h-6 text-2xl font-semibold sm:block","children":"TechLog"}]]}]}],["$","div",null,{"className":"flex items-center space-x-4 leading-5 sm:-mr-6 sm:space-x-6","children":[["$","div",null,{"className":"no-scrollbar hidden max-w-40 items-center gap-x-4 overflow-x-auto sm:flex md:max-w-72 lg:max-w-96","children":[["$","$L5","博客",{"className":"hover:text-primary-500 dark:hover:text-primary-400 m-1 font-medium text-gray-900 dark:text-gray-100","href":"/blog.html","children":"博客"}],["$","$L5","标签",{"className":"hover:text-primary-500 dark:hover:text-primary-400 m-1 font-medium text-gray-900 dark:text-gray-100","href":"/tags.html","children":"标签"}],["$","$L5","项目",{"className":"hover:text-primary-500 dark:hover:text-primary-400 m-1 font-medium text-gray-900 dark:text-gray-100","href":"/projects.html","children":"项目"}],["$","$L5","关于",{"className":"hover:text-primary-500 dark:hover:text-primary-400 m-1 font-medium text-gray-900 dark:text-gray-100","href":"/about.html","children":"关于"}]]}],["$","$L6",null,{"aria-label":"Search","children":"$L7"}],"$L8","$L9"]}]]}],"$La"]}],"$Lb"]}]]}]}]]}]]}],{"children":["blog","$Lc",{"children":[["slug","rag-vs-agent-retrieval-decision-framework","c"],"$Ld",{"children":["__PAGE__","$Le",{},null,false]},null,false]},null,false]},null,false],"$Lf",false]],"m":"$undefined","G":["$10",[]],"s":false,"S":true} 11:I[49217,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],"default"] 12:I[394,["2545","static/chunks/c16f53c3-391a1f5a73546b11.js","2619","static/chunks/2619-eccd988d1e161df3.js","3353","static/chunks/3353-35c19838f8cf7b3b.js","7177","static/chunks/app/layout-cb4429a7e3e88a9f.js"],"default"] 13:I[9766,[],""] 14:I[98924,[],""] 16:I[24431,[],"OutletBoundary"] 18:I[15278,[],"AsyncMetadataOutlet"] 1a:I[24431,[],"ViewportBoundary"] 1c:I[24431,[],"MetadataBoundary"] 1d:"$Sreact.suspense" 7:["$","svg",null,{"xmlns":"http://www.w3.org/2000/svg","fill":"none","viewBox":"0 0 24 24","strokeWidth":1.5,"stroke":"currentColor","className":"hover:text-primary-500 dark:hover:text-primary-400 h-6 w-6 text-gray-900 dark:text-gray-100","children":["$","path",null,{"strokeLinecap":"round","strokeLinejoin":"round","d":"M21 21l-5.197-5.197m0 0A7.5 7.5 0 105.196 5.196a7.5 7.5 0 0010.607 10.607z"}]}] 8:["$","$L11",null,{}] 9:["$","$L12",null,{}] a:["$","main",null,{"className":"mb-auto","children":["$","$L13",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L14",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":[["$","div",null,{"className":"flex flex-col items-start justify-start md:mt-24 md:flex-row md:items-center md:justify-center md:space-x-6","children":[["$","div",null,{"className":"space-x-2 pt-6 pb-8 md:space-y-5","children":["$","h1",null,{"className":"text-6xl leading-9 font-extrabold tracking-tight text-gray-900 md:border-r-2 md:px-6 md:text-8xl md:leading-14 dark:text-gray-100","children":"404"}]}],["$","div",null,{"className":"max-w-md","children":[["$","p",null,{"className":"mb-4 text-xl leading-normal font-bold md:text-2xl","children":"Sorry we couldn't find this page."}],["$","p",null,{"className":"mb-8","children":"But dont worry, you can find plenty of other things on our homepage."}],["$","$L5",null,{"className":"focus:shadow-outline-blue inline rounded-lg border border-transparent bg-blue-600 px-4 py-2 text-sm leading-5 font-medium text-white shadow-xs transition-colors duration-150 hover:bg-blue-700 focus:outline-hidden dark:hover:bg-blue-500","href":"/","children":"Back to homepage"}]]}]]}],[]],"forbidden":"$undefined","unauthorized":"$undefined"}]}] b:["$","footer",null,{"children":["$","div",null,{"className":"mt-16 flex flex-col items-center","children":[["$","div",null,{"className":"mb-3 flex space-x-4","children":[["$","a",null,{"className":"text-sm text-gray-500 transition hover:text-gray-600","target":"_blank","rel":"noopener noreferrer","href":"mailto:birenwenmou@qq.com","children":[["$","span",null,{"className":"sr-only","children":"mail"}],["$","svg",null,{"xmlns":"http://www.w3.org/2000/svg","viewBox":"0 0 20 20","className":"hover:text-primary-500 dark:hover:text-primary-400 fill-current text-gray-700 dark:text-gray-200 h-6 w-6","children":[["$","title",null,{"children":"Mail"}],["$","path",null,{"d":"M2.003 5.884L10 9.882l7.997-3.998A2 2 0 0016 4H4a2 2 0 00-1.997 1.884z"}],["$","path",null,{"d":"M18 8.118l-8 4-8-4V14a2 2 0 002 2h12a2 2 0 002-2V8.118z"}]]}]]}],["$","a",null,{"className":"text-sm text-gray-500 transition hover:text-gray-600","target":"_blank","rel":"noopener noreferrer","href":"https://github.com/joker1point","children":[["$","span",null,{"className":"sr-only","children":"github"}],["$","svg",null,{"xmlns":"http://www.w3.org/2000/svg","viewBox":"0 0 24 24","className":"hover:text-primary-500 dark:hover:text-primary-400 fill-current text-gray-700 dark:text-gray-200 h-6 w-6","children":[["$","title",null,{"children":"GitHub"}],["$","path",null,{"d":"M12 .297c-6.63 0-12 5.373-12 12 0 5.303 3.438 9.8 8.205 11.385.6.113.82-.258.82-.577 0-.285-.01-1.04-.015-2.04-3.338.724-4.042-1.61-4.042-1.61C4.422 18.07 3.633 17.7 3.633 17.7c-1.087-.744.084-.729.084-.729 1.205.084 1.838 1.236 1.838 1.236 1.07 1.835 2.809 1.305 3.495.998.108-.776.417-1.305.76-1.605-2.665-.3-5.466-1.332-5.466-5.93 0-1.31.465-2.38 1.235-3.22-.135-.303-.54-1.523.105-3.176 0 0 1.005-.322 3.3 1.23.96-.267 1.98-.399 3-.405 1.02.006 2.04.138 3 .405 2.28-1.552 3.285-1.23 3.285-1.23.645 1.653.24 2.873.12 3.176.765.84 1.23 1.91 1.23 3.22 0 4.61-2.805 5.625-5.475 5.92.42.36.81 1.096.81 2.22 0 1.606-.015 2.896-.015 3.286 0 .315.21.69.825.57C20.565 22.092 24 17.592 24 12.297c0-6.627-5.373-12-12-12"}]]}]]}],null,null,null,null,null,null,null,null,null]}],["$","div",null,{"className":"mb-2 flex space-x-2 text-sm text-gray-500 dark:text-gray-400","children":[["$","div",null,{"children":"眠月"}],["$","div",null,{"children":" • "}],["$","div",null,{"children":"© 2026"}],["$","div",null,{"children":" • "}],["$","$L5",null,{"className":"break-words","href":"/","children":"TechLog"}]]}],["$","div",null,{"className":"mb-8 text-sm text-gray-500 dark:text-gray-400","children":["$","a",null,{"className":"break-words","target":"_blank","rel":"noopener noreferrer","href":"https://github.com/joker1point","children":"GitHub"}]}]]}]}] c:["$","$1","c",{"children":[null,["$","$L13",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L14",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}] d:["$","$1","c",{"children":[null,["$","$L13",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L14",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}] e:["$","$1","c",{"children":["$L15",[["$","link","0",{"rel":"stylesheet","href":"/_next/static/css/ddf3d904beac0ed7.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}]],["$","$L16",null,{"children":["$L17",["$","$L18",null,{"promise":"$@19"}]]}]]}] f:["$","$1","h",{"children":[null,[["$","$L1a",null,{"children":"$L1b"}],null],["$","$L1c",null,{"children":["$","div",null,{"hidden":true,"children":["$","$1d",null,{"fallback":null,"children":"$L1e"}]}]}]]}] 1f:I[62088,["2619","static/chunks/2619-eccd988d1e161df3.js","1356","static/chunks/1356-354ecf51aaaf5203.js","6909","static/chunks/app/blog/%5B...slug%5D/page-d416e20a09ba8ecb.js"],"default"] 20:I[81356,["2619","static/chunks/2619-eccd988d1e161df3.js","1356","static/chunks/1356-354ecf51aaaf5203.js","6909","static/chunks/app/blog/%5B...slug%5D/page-d416e20a09ba8ecb.js"],"Image"] 15:[["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"BlogPosting\",\"headline\":\"RAG 还必要吗:从 130 条技术评论里提炼出的选型决策框架\",\"datePublished\":\"2026-09-15T00:00:00.000Z\",\"dateModified\":\"2026-09-15T00:00:00.000Z\",\"description\":\"围绕「向量 RAG vs LLM Wiki + grep / Agent 自主检索」的四派立场梳理(谨慎派、混合派、场景决定论派、演进视角派)、四个辩论冲突点,以及一组可直接落地的决策框架:按更新频率、查询频率、数据规模、领域特性决定上不上 RAG。\",\"image\":\"/static/images/twitter-card.png\",\"url\":\"https://techlog-18328.app.workbuddy.host/blog/rag-vs-agent-retrieval-decision-framework\",\"author\":[{\"@type\":\"Person\",\"name\":\"眠月\"}]}"}}],["$","section",null,{"className":"mx-auto max-w-3xl px-4 sm:px-6 xl:max-w-5xl xl:px-0","children":[["$","$L1f",null,{}],["$","article",null,{"children":["$","div",null,{"className":"xl:divide-y xl:divide-gray-200 xl:dark:divide-gray-700","children":[["$","header",null,{"className":"pt-6 xl:pb-6","children":["$","div",null,{"className":"space-y-1 text-center","children":[["$","dl",null,{"className":"space-y-10","children":[["$","div",null,{"children":[["$","dt",null,{"className":"sr-only","children":"Published on"}],["$","dd",null,{"className":"text-base leading-6 font-medium text-gray-500 dark:text-gray-400","children":["$","time",null,{"dateTime":"2026-09-15T00:00:00.000Z","children":"2026年9月15日星期二"}]}]]}],["$","div",null,{"children":[["$","dt",null,{"className":"sr-only","children":"Word count"}],["$","dd",null,{"className":"text-sm text-gray-400 dark:text-gray-500","children":["1,878"," 字"," · 阅读约 7 分钟"]}]]}]]}],["$","div",null,{"children":["$","h1",null,{"className":"text-3xl leading-9 font-extrabold tracking-tight text-gray-900 sm:text-4xl sm:leading-10 md:text-5xl md:leading-14 dark:text-gray-100","children":"RAG 还必要吗:从 130 条技术评论里提炼出的选型决策框架"}]}]]}]}],["$","div",null,{"className":"grid-rows-[auto_1fr] divide-y divide-gray-200 pb-8 xl:grid xl:grid-cols-4 xl:gap-x-6 xl:divide-y-0 dark:divide-gray-700","children":[["$","dl",null,{"className":"pt-6 pb-10 xl:border-b xl:border-gray-200 xl:pt-11 xl:dark:border-gray-700","children":[["$","dt",null,{"className":"sr-only","children":"Authors"}],["$","dd",null,{"children":["$","ul",null,{"className":"flex flex-wrap justify-center gap-4 sm:space-x-12 xl:block xl:space-y-8 xl:space-x-0","children":[["$","li","眠月",{"className":"flex items-center space-x-2","children":[["$","$L20",null,{"src":"/static/images/avatar.png","width":38,"height":38,"alt":"avatar","className":"h-10 w-10 rounded-full"}],["$","dl",null,{"className":"text-sm leading-5 font-medium whitespace-nowrap","children":[["$","dt",null,{"className":"sr-only","children":"Name"}],["$","dd",null,{"className":"text-gray-900 dark:text-gray-100","children":"眠月"}],["$","dt",null,{"className":"sr-only","children":"Twitter"}],["$","dd",null,{"children":""}]]}]]}]]}]}]]}],["$","div",null,{"className":"divide-y divide-gray-200 xl:col-span-3 xl:row-span-2 xl:pb-0 dark:divide-gray-700","children":[["$","div",null,{"className":"prose dark:prose-invert max-w-none pt-10 pb-8","children":[["$","p",null,{"children":"基于对 130 条相关技术评论的筛选与分析,已过滤玩梗、灌水、无实质内容的评论(约占一半以上),以下为结构化笔记。"}],["$","hr",null,{}],["$","h3",null,{"className":"content-header","id":"话题主题","children":[["$","a",null,{"className":"break-words","href":"#话题主题","aria-hidden":"true","tabIndex":"-1","children":["$","span",null,{"className":"content-header-link","children":["$","svg",null,{"className":"h-5 linkicon w-5","fill":"currentColor","viewBox":"0 0 20 20","xmlns":"http://www.w3.org/2000/svg","children":[["$","path",null,{"d":"M12.232 4.232a2.5 2.5 0 0 1 3.536 3.536l-1.225 1.224a.75.75 0 0 0 1.061 1.06l1.224-1.224a4 4 0 0 0-5.656-5.656l-3 3a4 4 0 0 0 .225 5.865.75.75 0 0 0 .977-1.138 2.5 2.5 0 0 1-.142-3.667l3-3Z"}],["$","path",null,{"d":"M11.603 7.963a.75.75 0 0 0-.977 1.138 2.5 2.5 0 0 1 .142 3.667l-3 3a2.5 2.5 0 0 1-3.536-3.536l1.225-1.224a.75.75 0 0 0-1.061-1.06l-1.224 1.224a4 4 0 1 0 5.656 5.656l3-3a4 4 0 0 0-.225-5.865Z"}]]}]}]}],["$","strong",null,{"children":"话题主题"}]]}],["$","p",null,{"children":"大模型长上下文与 Agentic 能力增强的背景下,传统 RAG(检索增强生成)是否仍然必要?核心争议围绕「向量 RAG vs LLM Wiki + grep / Agent 自主检索」的技术路线选择,以及 RAG 在不同场景下的适用边界。"}],"$L21","$L22","$L23","$L24","$L25","$L26","$L27","$L28","$L29","$L2a","$L2b","$L2c","$L2d","$L2e","$L2f","$L30","$L31","$L32","$L33","$L34","$L35","$L36","$L37","$L38","$L39"]}],"$L3a","$L3b"]}],"$L3c"]}]]}]}]]}]] 3d:I[50528,["2619","static/chunks/2619-eccd988d1e161df3.js","1356","static/chunks/1356-354ecf51aaaf5203.js","6909","static/chunks/app/blog/%5B...slug%5D/page-d416e20a09ba8ecb.js"],"default"] 21:["$","hr",null,{}] 22:["$","h3",null,{"className":"content-header","id":"不同立场--核心观点","children":[["$","a",null,{"className":"break-words","href":"#不同立场--核心观点","aria-hidden":"true","tabIndex":"-1","children":["$","span",null,{"className":"content-header-link","children":["$","svg",null,{"className":"h-5 linkicon w-5","fill":"currentColor","viewBox":"0 0 20 20","xmlns":"http://www.w3.org/2000/svg","children":[["$","path",null,{"d":"M12.232 4.232a2.5 2.5 0 0 1 3.536 3.536l-1.225 1.224a.75.75 0 0 0 1.061 1.06l1.224-1.224a4 4 0 0 0-5.656-5.656l-3 3a4 4 0 0 0 .225 5.865.75.75 0 0 0 .977-1.138 2.5 2.5 0 0 1-.142-3.667l3-3Z"}],["$","path",null,{"d":"M11.603 7.963a.75.75 0 0 0-.977 1.138 2.5 2.5 0 0 1 .142 3.667l-3 3a2.5 2.5 0 0 1-3.536-3.536l1.225-1.224a.75.75 0 0 0-1.061-1.06l-1.224 1.224a4 4 0 1 0 5.656 5.656l3-3a4 4 0 0 0-.225-5.865Z"}]]}]}]}],["$","strong",null,{"children":"不同立场 & 核心观点"}]]}] 23:["$","p",null,{"children":["$","strong",null,{"children":"立场 A:RAG 谨慎使用派 / Agent 检索派(评论区声量最大)"}]}] 24:["$","p",null,{"children":"核心观点:个人及中小规模知识库场景下,LLM Wiki + grep(或 Agent 自主检索)优于传统向量 RAG。"}] 25:["$","p",null,{"children":"关键论据:"}] 26:["$","ul",null,{"children":[["$","li",null,{"children":"RAG 落地门槛高:需搭建向量库、设计分块策略,Embedding 与 Rerank 均为黑盒;语料变大后难以评估效果,召回不佳时无法定位问题出在分块、向量化、召回还是重排环节,调优缺乏可操作性。"}],["$","li",null,{"children":"grep 路线检索过程透明、可验证、可调试,符合人类「关键词定位 → 阅读上下文 → 沿文档结构扩展」的直觉。"}],["$","li",null,{"children":"RAG 优化近年除稀疏检索和图数据库外近乎停滞;主流模型 Agentic 能力增强后,仅凭 grep + shell 即可高效检索;优化提示词比调 RAG 参数效果更稳定。"}],["$","li",null,{"children":"多位评论者引用 Karpathy 的 LLM Wiki 思路佐证「非必要不上 RAG」,认为「按目录翻书」式检索返璞归真。"}]]}] 27:["$","p",null,{"children":["$","strong",null,{"children":"立场 B:RAG 改造 / 混合路线派"}]}] 28:["$","p",null,{"children":"核心观点:RAG 本身没死,但老套的「用户原文 → 向量召回」模式应被淘汰,知识库的存在价值与查询方式要分开讨论。"}] 29:["$","p",null,{"children":"关键论据:"}] 2a:["$","ul",null,{"children":[["$","li",null,{"children":"拿用户提问原文(即使改写)做召回是低效的;应提供搜索工具,由 Agent 自主决定搜什么、用什么方式和参数搜——「怎么查」比「用什么结构搭知识库」更重要。"}],["$","li",null,{"children":"Cursor 实践案例:用 Agent 实际 grep 到的代码数据训练 Embedding,构建「向量 RAG + grep」混合方案,官方宣称同模型下效果优于 codex、claude code——「用 Agent 检索结果反哺 Embedding 训练」可能是结合两者优势的方向。"}],["$","li",null,{"children":"视频方观点(经 AI 总结评论转述):RAG 未过时,成本、延迟、上下文稀释问题使全量塞入不可行;推荐语义/命题分块、领域专用 Embedding、混合检索 + 重排序作为默认方案、自适应路由(简单问题走 pipeline,复杂问题走 Agent)。"}]]}] 2b:["$","p",null,{"children":["$","strong",null,{"children":"立场 C:场景决定论派"}]}] 2c:["$","p",null,{"children":"核心观点:不存在通用答案,技术选型由数据特征和场景决定。"}] 2d:["$","p",null,{"children":"关键论据:"}] 2e:["$","ul",null,{"children":[["$","li",null,{"children":"数据更新频率 vs 查询频率:RAG 当前主要价值是控制成本,仅当被查询内容的更新速度远低于查询次数时才划算;更新频繁的场景(如个人笔记)不如直接准备目录 + 分节文档让 Agent 自己翻两轮。"}],["$","li",null,{"children":"数据规模:百万 token 级上下文窗口下,小体量数据根本不需要 RAG 系统;RAG 应对的是稳定可靠、有可靠来源的基础数据集,而非对话产生的数据。"}],["$","li",null,{"children":"领域特性:代码、法律等场景的核心需求是「关键字在哪里被定义/使用」的精确索引,语法分析器和索引器比相似度检出的 RAG 更合适。"}],["$","li",null,{"children":"长文本创作(如小说):上下文长度确实是硬约束需要 RAG,但纯 RAG 缺少实体提取和关系链,需要 graph 类记忆补充,且可借此剥离记忆阶段与输出阶段以控制成本。"}]]}] 2f:["$","p",null,{"children":["$","strong",null,{"children":"立场 D:演进视角派(RAG 是更大系统的组件)"}]}] 30:["$","p",null,{"children":"核心观点:Memory System 尚处研究探索期,RAG 只是其中一个基础组件;LLM Wiki 是「没有好 memory system 时存全文、相信大模型」的过渡形态——当前有效,但非最终形态。另有评论指出 RAG 流程与当下 Agent Memory 流程高度相似,暗示两者正在合流。"}] 31:["$","hr",null,{}] 32:["$","h3",null,{"className":"content-header","id":"辩论冲突点","children":[["$","a",null,{"className":"break-words","href":"#辩论冲突点","aria-hidden":"true","tabIndex":"-1","children":["$","span",null,{"className":"content-header-link","children":["$","svg",null,{"className":"h-5 linkicon w-5","fill":"currentColor","viewBox":"0 0 20 20","xmlns":"http://www.w3.org/2000/svg","children":[["$","path",null,{"d":"M12.232 4.232a2.5 2.5 0 0 1 3.536 3.536l-1.225 1.224a.75.75 0 0 0 1.061 1.06l1.224-1.224a4 4 0 0 0-5.656-5.656l-3 3a4 4 0 0 0 .225 5.865.75.75 0 0 0 .977-1.138 2.5 2.5 0 0 1-.142-3.667l3-3Z"}],["$","path",null,{"d":"M11.603 7.963a.75.75 0 0 0-.977 1.138 2.5 2.5 0 0 1 .142 3.667l-3 3a2.5 2.5 0 0 1-3.536-3.536l1.225-1.224a.75.75 0 0 0-1.061-1.06l-1.224 1.224a4 4 0 1 0 5.656 5.656l3-3a4 4 0 0 0-.225-5.865Z"}]]}]}]}],["$","strong",null,{"children":"辩论冲突点"}]]}] 33:["$","ol",null,{"children":[["$","li",null,{"children":[["$","strong",null,{"children":"成本论的对撞"}],":RAG 支持方认为其核心价值是压缩上下文成本;反方认为固定上下文基本走缓存、费用已很低,成本理由不成立;第三方补充「1M 窗口用满会致幻」,即成本之外还有上下文稀释带来的质量问题。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"黑盒可调性"}],":反方指控 Embedding/Rerank 是玄学黑盒、无法定位问题;混合派以 Cursor 案例回应——黑盒可以用 Agent 行为数据去训练和打开。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"查询发起权"}],":传统派默认用户原文召回;改造派主张 Agent 自主构造查询。但有实践者反馈反面证据:不细致给定搜索词时,AI 自己想的搜索词召回率低,小数据库场景 token 消耗甚至超过全量发送——Agent 自主检索并非免费午餐。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"语义模糊查询的归属"}],":grep 派的论证默认用户知道关键词或文档结构良好,但对「用户自己说不清要找什么」的语义模糊场景,向量检索的不可替代性在反方论证中基本缺席。"]}]]}] 34:["$","hr",null,{}] 35:["$","h3",null,{"className":"content-header","id":"逻辑疑点--待思考问题","children":[["$","a",null,{"className":"break-words","href":"#逻辑疑点--待思考问题","aria-hidden":"true","tabIndex":"-1","children":["$","span",null,{"className":"content-header-link","children":["$","svg",null,{"className":"h-5 linkicon w-5","fill":"currentColor","viewBox":"0 0 20 20","xmlns":"http://www.w3.org/2000/svg","children":[["$","path",null,{"d":"M12.232 4.232a2.5 2.5 0 0 1 3.536 3.536l-1.225 1.224a.75.75 0 0 0 1.061 1.06l1.224-1.224a4 4 0 0 0-5.656-5.656l-3 3a4 4 0 0 0 .225 5.865.75.75 0 0 0 .977-1.138 2.5 2.5 0 0 1-.142-3.667l3-3Z"}],["$","path",null,{"d":"M11.603 7.963a.75.75 0 0 0-.977 1.138 2.5 2.5 0 0 1 .142 3.667l-3 3a2.5 2.5 0 0 1-3.536-3.536l1.225-1.224a.75.75 0 0 0-1.061-1.06l-1.224 1.224a4 4 0 1 0 5.656 5.656l3-3a4 4 0 0 0-.225-5.865Z"}]]}]}]}],["$","strong",null,{"children":"逻辑疑点 / 待思考问题"}]]}] 36:["$","ul",null,{"children":[["$","li",null,{"children":"「优化提示词即可解决一批同类问题、效果稳定」与「RAG 调优是玄学」两个主张都缺乏量化证据,双方举证标准不对称。"}],["$","li",null,{"children":"Cursor 混合方案的效果宣称来自官方自述,评论区自身也存疑(「不知道靠不靠谱」),尚无第三方独立验证。"}],["$","li",null,{"children":"grep 派的隐性前提——语料结构良好、查询含明确关键词、模型 Agentic 能力足够强——在语料规模极大或结构混乱时是否成立,未被正方回应。"}],["$","li",null,{"children":"LLM Wiki 存全文的方案,其成本与维护开销随规模增长的可持续性缺少讨论。"}],["$","li",null,{"children":"若 RAG 只是 Memory System 的基础组件、且与 Agent Memory 流程趋同,则「RAG 过时论」是否存在偷换概念——过时的究竟是 RAG 这项技术,还是「朴素向量召回」这一种实现?"}],["$","li",null,{"children":"动态学习路线(模型自主补充知识边界)衍生出数据安全、治理、千人千面问题,评论区指出这「好像又回到了 RAG 范畴」,形成循环论证,无人给出解法。"}]]}] 37:["$","hr",null,{}] 38:["$","h3",null,{"className":"content-header","id":"关键信息摘要","children":[["$","a",null,{"className":"break-words","href":"#关键信息摘要","aria-hidden":"true","tabIndex":"-1","children":["$","span",null,{"className":"content-header-link","children":["$","svg",null,{"className":"h-5 linkicon w-5","fill":"currentColor","viewBox":"0 0 20 20","xmlns":"http://www.w3.org/2000/svg","children":[["$","path",null,{"d":"M12.232 4.232a2.5 2.5 0 0 1 3.536 3.536l-1.225 1.224a.75.75 0 0 0 1.061 1.06l1.224-1.224a4 4 0 0 0-5.656-5.656l-3 3a4 4 0 0 0 .225 5.865.75.75 0 0 0 .977-1.138 2.5 2.5 0 0 1-.142-3.667l3-3Z"}],["$","path",null,{"d":"M11.603 7.963a.75.75 0 0 0-.977 1.138 2.5 2.5 0 0 1 .142 3.667l-3 3a2.5 2.5 0 0 1-3.536-3.536l1.225-1.224a.75.75 0 0 0-1.061-1.06l-1.224 1.224a4 4 0 1 0 5.656 5.656l3-3a4 4 0 0 0-.225-5.865Z"}]]}]}]}],["$","strong",null,{"children":"关键信息摘要"}]]}] 39:["$","ul",null,{"children":[["$","li",null,{"children":[["$","strong",null,{"children":"决策框架(可复用)"}],":更新频率 > 查询频率 → 目录 + Agent 翻检;更新稳定 + 高频查询 → RAG;数据量 < 上下文窗口有效容量 → 直接全量;精确符号定位需求(代码/法律)→ 语法索引器优先于向量。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"工程共识"}],":分块与排序是 RAG 永恒的难点(「切片和排序永远是大敌」获高共鸣);混合检索 + Rerank 是当前性价比最高的默认方案。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"值得关注的技术动向"}],":Agent 检索行为数据反哺 Embedding 训练(Cursor 模式);自适应路由(轻量分类器分流 pipeline / Agent 模式);Graph 类记忆补足 RAG 缺失的实体关系链。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"实践教训"}],":小数据规模下自建 RAG(哪怕用免费 Embedding API + Qdrant)可能得不偿失,token 消耗与召回率双输。"]}],["$","li",null,{"children":[["$","strong",null,{"children":"社区情绪"}],":对「老套 RAG」普遍悲观,对「知识库 + Agent 检索」普遍乐观,但对 Agent 检索的可靠性同样存在一线实践者的警惕。"]}]]}] 3a:["$","div",null,{"className":"pt-6 pb-6 text-sm text-gray-700 dark:text-gray-300","children":[["$","a",null,{"className":"break-words","target":"_blank","rel":"nofollow","href":"https://mobile.twitter.com/search?q=https%3A%2F%2Ftechlog-18328.app.workbuddy.host%2Fblog%2Frag-vs-agent-retrieval-decision-framework","children":"Discuss on Twitter"}]," • ",["$","a",null,{"className":"break-words","target":"_blank","rel":"noopener noreferrer","href":"https://github.com/joker1point/techlog/blob/main/data/blog/rag-vs-agent-retrieval-decision-framework.mdx","children":"View on GitHub"}]]}] 3b:["$","div",null,{"className":"pt-6 pb-6 text-center text-gray-700 dark:text-gray-300","id":"comment","children":["$","$L3d",null,{"slug":"rag-vs-agent-retrieval-decision-framework"}]}] 3c:["$","footer",null,{"children":[["$","div",null,{"className":"divide-gray-200 text-sm leading-5 font-medium xl:col-start-1 xl:row-start-2 xl:divide-y dark:divide-gray-700","children":[["$","div",null,{"className":"py-4 xl:py-8","children":[["$","h2",null,{"className":"text-xs tracking-wide text-gray-500 uppercase dark:text-gray-400","children":"Tags"}],["$","div",null,{"className":"flex flex-wrap","children":[["$","$L5","rag",{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400 mr-3 text-sm font-medium uppercase","href":"/tags/rag.html","children":"rag"}],["$","$L5","agent",{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400 mr-3 text-sm font-medium uppercase","href":"/tags/agent.html","children":"agent"}],["$","$L5","ai",{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400 mr-3 text-sm font-medium uppercase","href":"/tags/ai.html","children":"ai"}],["$","$L5","research",{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400 mr-3 text-sm font-medium uppercase","href":"/tags/research.html","children":"research"}]]}]]}],["$","div",null,{"className":"flex justify-between py-4 xl:block xl:space-y-8 xl:py-8","children":[["$","div",null,{"children":[["$","h2",null,{"className":"text-xs tracking-wide text-gray-500 uppercase dark:text-gray-400","children":"Previous Article"}],["$","div",null,{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400","children":["$","$L5",null,{"className":"break-words","href":"/blog/agent-lessons-precipitation.html","children":"如何沉淀 Agent 每次的踩坑经验"}]}]]}],["$","div",null,{"children":[["$","h2",null,{"className":"text-xs tracking-wide text-gray-500 uppercase dark:text-gray-400","children":"Next Article"}],["$","div",null,{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400","children":["$","$L5",null,{"className":"break-words","href":"/blog/mpt-pr-1291-contribution-review.html","children":"给 119k star 项目提 PR:8 小时从评论到合并的完整复盘"}]}]]}]]}]]}],["$","div",null,{"className":"pt-4 xl:pt-8","children":["$","$L5",null,{"className":"text-primary-500 hover:text-primary-600 dark:hover:text-primary-400","href":"/blog.html","aria-label":"Back to the blog","children":"← Back to the blog"}]}]]}] 1b:[["$","meta","0",{"charSet":"utf-8"}],["$","meta","1",{"name":"viewport","content":"width=device-width, initial-scale=1"}]] 17:null 19:{"metadata":[["$","title","0",{"children":"RAG 还必要吗:从 130 条技术评论里提炼出的选型决策框架 | TechLog"}],["$","meta","1",{"name":"description","content":"围绕「向量 RAG vs LLM Wiki + grep / Agent 自主检索」的四派立场梳理(谨慎派、混合派、场景决定论派、演进视角派)、四个辩论冲突点,以及一组可直接落地的决策框架:按更新频率、查询频率、数据规模、领域特性决定上不上 RAG。"}],["$","meta","2",{"name":"robots","content":"index, follow"}],["$","meta","3",{"name":"googlebot","content":"index, follow, max-video-preview:-1, max-image-preview:large, max-snippet:-1"}],["$","link","4",{"rel":"canonical","href":"https://techlog-18328.app.workbuddy.host/blog/rag-vs-agent-retrieval-decision-framework"}],["$","link","5",{"rel":"alternate","type":"application/rss+xml","href":"https://techlog-18328.app.workbuddy.host/feed.xml"}],["$","meta","6",{"property":"og:title","content":"RAG 还必要吗:从 130 条技术评论里提炼出的选型决策框架"}],["$","meta","7",{"property":"og:description","content":"围绕「向量 RAG vs LLM Wiki + grep / Agent 自主检索」的四派立场梳理(谨慎派、混合派、场景决定论派、演进视角派)、四个辩论冲突点,以及一组可直接落地的决策框架:按更新频率、查询频率、数据规模、领域特性决定上不上 RAG。"}],["$","meta","8",{"property":"og:url","content":"https://techlog-18328.app.workbuddy.host/blog/rag-vs-agent-retrieval-decision-framework"}],["$","meta","9",{"property":"og:site_name","content":"TechLog"}],["$","meta","10",{"property":"og:locale","content":"en_US"}],["$","meta","11",{"property":"og:image","content":"https://techlog-18328.app.workbuddy.host/static/images/twitter-card.png"}],["$","meta","12",{"property":"og:type","content":"article"}],["$","meta","13",{"property":"article:published_time","content":"2026-09-15T00:00:00.000Z"}],["$","meta","14",{"property":"article:modified_time","content":"2026-09-15T00:00:00.000Z"}],["$","meta","15",{"property":"article:author","content":"眠月"}],["$","meta","16",{"name":"twitter:card","content":"summary_large_image"}],["$","meta","17",{"name":"twitter:title","content":"RAG 还必要吗:从 130 条技术评论里提炼出的选型决策框架"}],["$","meta","18",{"name":"twitter:description","content":"围绕「向量 RAG vs LLM Wiki + grep / Agent 自主检索」的四派立场梳理(谨慎派、混合派、场景决定论派、演进视角派)、四个辩论冲突点,以及一组可直接落地的决策框架:按更新频率、查询频率、数据规模、领域特性决定上不上 RAG。"}],["$","meta","19",{"name":"twitter:image","content":"https://techlog-18328.app.workbuddy.host/static/images/twitter-card.png"}]],"error":null,"digest":"$undefined"} 1e:"$19:metadata"