[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:stajic:zh":3,"public-menus:all":38,"post:why-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer:zh":221,"related:post:why-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer:zh:1":1335},{"statusCode":4,"data":5,"message":37},200,{"tenantId":6,"lang":7,"defaultLang":8,"siteUrl":9,"contactEmail":10,"brandName":11,"logoUrl":12,"siteName":11,"siteDescription":13,"ogImage":10,"robotsIndex":14,"socialLinks":10,"reservedSlugs":10,"seoPolicy":15},"stajic","zh","de","https:\u002F\u002Fstajic.de",null,"Stajic Platform","\u002FLogo_Planet.svg","Stajic Portal",true,{"branding":16,"relatedContent":17,"crossDomainLinks":18},{"logoUrl":12},{"enabled":14},[19,22,25,28,31,34],{"url":20,"label":21,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Ffigure.rocks","figure.rocks",{"url":23,"label":24,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Floving.rocks","loving.rocks",{"url":26,"label":27,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.com","bazify.com",{"url":29,"label":30,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.de","bazify.de",{"url":32,"label":33,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.at","bazify.at",{"url":35,"label":36,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.ba","bazify.ba","Portal settings resolved",[39,45],{"id":40,"name":41,"location":42,"isActive":14,"isDefault":43,"items":44},1,"main-navigation","header",false,[],{"id":46,"name":47,"location":48,"isActive":14,"isDefault":14,"items":49},4,"main-menu","sidebar",[50,66,79,93,103,118,133],{"id":51,"title":52,"url":60,"target":61,"icon":62,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":64,"portfolioId":10,"children":65},"item-18",{"de":53,"en":54,"es":55,"fr":56,"it":54,"ru":57,"sr":58,"zh":59},"Startseite","Home","Inicio","Accueil","Главная","Почетна","首页","\u002Ffull-stack-web-developer-munich-performance-seo-and-maintainable-builds","_self","i-lucide-home","page",111,[],{"id":67,"title":68,"url":75,"target":61,"icon":76,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":77,"portfolioId":10,"children":78},"item-22",{"de":69,"en":69,"es":70,"fr":69,"it":71,"ru":72,"sr":73,"zh":74},"Vision","Visión","Visione","Видение","Визија","想象","\u002Fueber-uns-webdesign-muenchen-webaplikation","i-lucide-eye",113,[],{"id":80,"title":81,"url":89,"target":61,"icon":90,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":91,"portfolioId":10,"children":92},"item-19",{"de":82,"en":83,"es":84,"fr":83,"it":85,"ru":86,"sr":87,"zh":88},"Leistungen","Services","Servicios","Servizi","Услуги","Услуге","服务","\u002Fservices-dienstleistungen-muenchen","i-lucide-wrench",116,[],{"id":94,"title":95,"url":99,"target":61,"icon":100,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":101,"portfolioId":10,"children":102},"item-23",{"de":96,"en":96,"es":96,"fr":96,"it":96,"ru":97,"sr":97,"zh":98},"Blog","Блог","博客","\u002Fblog","i-lucide-book-open",112,[],{"id":104,"title":105,"url":114,"target":61,"icon":115,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":116,"portfolioId":10,"children":117},"item-32",{"de":106,"en":107,"es":108,"fr":109,"it":110,"ru":111,"sr":112,"zh":113},"Neue Technologien","New Technologies","Nuevas tecnologías","Nouvelles technologies","Nuove tecnologie","Новые технологии","Нове технологије","新技术！","\u002Fneue-webtechnologien","i-lucide-sparkles",122,[],{"id":119,"title":120,"url":129,"target":61,"icon":130,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":131,"portfolioId":10,"children":132},"item-20",{"de":121,"en":122,"es":123,"fr":124,"it":125,"ru":126,"sr":127,"zh":128},"Kontakt","Contact us!","Contacto","Contact","Contatto","Контакт","Контактирајте нас","联系我们！","\u002Fcontact","i-lucide-mail",115,[],{"id":134,"title":135,"url":144,"target":61,"icon":145,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":147},"item-21",{"de":136,"en":137,"es":138,"fr":139,"it":140,"ru":141,"sr":142,"zh":143},"Unsere Arbeit","Our Work","Nuestro trabajo","Nos réalisations","I nostri lavori","Наши работы","Наши радови","文件夹","\u002Fportfolio","i-lucide-briefcase",114,[148,161,175,191,206],{"id":149,"title":150,"url":144,"target":61,"icon":159,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":160},"item-24",{"de":151,"en":152,"es":153,"fr":154,"it":155,"ru":156,"sr":157,"zh":158},"Alle Projekte","All Projects","Todos los proyectos","Tous les projets","Tutti i progetti","Все проекты","Сви пројекти","所有项目","i-lucide-grid-3x3",[],{"id":162,"title":163,"url":171,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":174},"item-29",{"de":164,"en":165,"es":166,"fr":167,"it":168,"ru":169,"sr":170,"zh":143},"Local Roots, Global Reach","Local Roots - Global Reach","Empresa local ","Entreprise locale","Azienda locale","Местная компания","Локално предузеће глобално тржиште","\u002Fportfolio\u002Flocal-roots-global-reach-communication-media-systems-for-modern-business","i-lucide-globe-2","custom",[],{"id":176,"title":177,"url":186,"target":61,"icon":187,"isActive":14,"type":188,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":189,"children":190},"menu-2916d060-8681-482a-af9a-c107a8bc30d8",{"de":178,"en":179,"es":180,"fr":181,"it":182,"ru":183,"sr":184,"zh":185},"Von Rohdaten zu verifizierbarem KI-Kontext: Eine praktische Architektur für agentenbereite Informationen","From Raw Data to Verifiable AI Context:","De datos sin procesar a contexto de IA verificable: una arquitectura práctica para información lista para agentes","Des données brutes au contexte d'IA vérifiable : une architecture pratique pour des informations prêtes pour les agents","Dai dati grezzi a un contesto IA verificabile: Un'architettura pratica per informazioni pronte per gli agenti","От необработанных данных к верифицируемому контексту ИИ: практическая архитектура информации, готовой для агентов","Od sirovih podataka do proverljivog AI konteksta: Praktična arhitektura za informacije spremne za agente","从原始数据到可验证的 AI 上下文：面向智能体就绪信息的实用架构","\u002Fportfolio\u002Ffrom-raw-data-to-verifiable-ai-context-a-practical-architecture-for-agent-ready-information","i-lucide-database","portfolio",22,[],{"id":192,"title":193,"url":202,"target":61,"icon":203,"isActive":14,"type":188,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":204,"children":205},"menu-e2061cd7-ff53-49d5-a0e8-3d49090577c0",{"de":194,"en":195,"es":196,"fr":197,"it":198,"ru":199,"sr":200,"zh":201},"Präsenzplattformen — Unternehmenskommunikation","Presence Platforms — Enterprise Communication","Plataformas de presencia — Comunicación empresarial","Plateformes de présence — Communication d'entreprise","Presence Platforms — Comunicazione aziendale","Платформы присутствия — Корпоративные коммуникации","Presence platforme — Poslovna komunikacija","Presence 平台 — 企业通信","\u002Fportfolio\u002Fpresence-platforms-enterprise-communication-that-wins-without-competing","i-lucide-badge-percent",21,[],{"id":207,"title":208,"url":217,"target":61,"icon":218,"isActive":14,"type":188,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":219,"children":220},"menu-a781582e-e90d-418e-b8c9-b070bc37c37e",{"de":209,"en":210,"es":211,"fr":212,"it":213,"ru":214,"sr":215,"zh":216},"Maximierung des POS-ROI durch Konsistenz","Maximizing POS ROI Through Consistent","Maximizar el ROI del POS a través de constante","Maximiser le ROI du PDV grâce à la cohérence","Massimizzare il ROI del POS attraverso costante","Максимизация ROI POS за счет последовательного","Maksimizacija POS ROI-ja kroz dosledno","通过持续最大化 POS 投资回报率","\u002Fportfolio\u002Ffrom-vision-to-value-maximizing-pos-roi-through-consistent-ci-creation-and-strategic-branding","i-lucide-briefcase-business",11,[],{"statusCode":4,"data":222,"message":1334},{"id":223,"title":224,"slug":225,"content":226,"contentJson":227,"excerpt":684,"featuredImage":685,"featuredImageAlt":686,"featuredImageCaption":10,"featuredImageTitle":10,"featuredImageCopyright":10,"featuredImageAuthor":10,"featuredImageSourceUrl":10,"featuredImageLicense":10,"featuredImageIsAiGenerated":43,"status":687,"publishedAt":688,"createdAt":689,"updatedAt":690,"seoLocalePaths":691,"categories":700,"author":709,"translations":713},"496","为什么AI会检索某些来源而忽略其他来源？——检索权威层","why-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u003Ch2 id=\"section-1\">问题\u003C\u002Fh2>\n\u003Cp>为什么AI检索系统会选择其中一个来源，而忽略另一个看似讨论完全相同主题的来源？\u003C\u002Fp>\n\u003Cp>语义相关性只是答案的一部分。两个文档可能与查询同样相似，但在权威性、时效性、来源、范围、权限和适用性方面存在根本差异。\u003C\u002Fp>\n\u003Cp>本文使用术语“检索权威层”来指代位于发现潜在相关信息与将该信息视为合适证据之间的来源选择边界。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">检索权威层是一组来源选择规则和来源属性，用于确定哪些可检索信息值得成为特定问题的证据。\u003Ccite class=\"block mt-2 text-sm\">— 工作定义——本文使用的系统概念，并非标准化研究术语。\u003C\u002Fcite>\u003C\u002Fblockquote>\n\u003Cnav class=\"editorjs-toc\" data-editorjs-toc=\"true\" aria-label=\"目录\">\u003Cstrong class=\"editorjs-toc__title\">目录\u003C\u002Fstrong>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-0\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-1\" class=\"editorjs-toc__link\">问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-7\" class=\"editorjs-toc__link\">这真正意味着什么\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-13\" class=\"editorjs-toc__link\">最简单的例子\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-19\" class=\"editorjs-toc__link\">示例失效之处\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-23\" class=\"editorjs-toc__link\">直接回答\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-27\" class=\"editorjs-toc__link\">为什么仅靠相似性不够\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-31\" class=\"editorjs-toc__link\">六个权威信号\u003C\u002Fa>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-1\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-32\" class=\"editorjs-toc__link\">1. 来源身份\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-34\" class=\"editorjs-toc__link\">2. 来源出处\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-36\" class=\"editorjs-toc__link\">3. 新鲜度\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-38\" class=\"editorjs-toc__link\">4. 适用性\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-40\" class=\"editorjs-toc__link\">5. 证据特异性\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-42\" class=\"editorjs-toc__link\">6. 访问合法性\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-45\" class=\"editorjs-toc__link\">一个最小权限模型\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-49\" class=\"editorjs-toc__link\">与代理就绪信息的连接\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-53\" class=\"editorjs-toc__link\">诊断\u002F决策方法\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-57\" class=\"editorjs-toc__link\">常见失败模式\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-62\" class=\"editorjs-toc__link\">边缘情况：相互冲突的权威来源\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-66\" class=\"editorjs-toc__link\">局限性\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-70\" class=\"editorjs-toc__link\">什么会改变这个答案？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-73\" class=\"editorjs-toc__link\">结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-78\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-7\">这真正意味着什么\u003C\u002Fh2>\n\u003Cp>检索器回答的是一个相对狭窄的问题：哪些条目看起来与此查询相关？\u003C\u002Fp>\n\u003Cp>生产级AI系统通常需要回答一个更难的问题：这些相关条目中，哪些应被允许影响答案？\u003C\u002Fp>\n\u003Cp>第二个问题引入了来源身份、版本、权威性、有效期、来源和范围。一份相关的旧政策、复制的博客文章或非官方摘要可能排名极高，但对于该决策而言仍然是错误的证据。\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>Question\n   ↓\nRetrieval Trigger\n   ↓\nCandidate Sources\n   ↓\nRETRIEVAL AUTHORITY LAYER\n   ↓\nQualified Evidence\n   ↓\nContext \u002F Reasoning\n   ↓\nAnswer Validity Boundary\n   ↓\nAnswer\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">AI何时应停止信任自身知识？——检索触发器\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">检索触发器决定何时需要外部证据。检索权威层则决定哪些来源应提供证据。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ch2 id=\"section-13\">最简单的例子\u003C\u002Fh2>\n\u003Cp>假设一名员工问企业助手：员工有多少天年假？\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">候选\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">内容\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">语义相关性\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">A\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">当前HR政策，已批准版本6.2\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">高\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">B\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">从版本4.8复制的旧入职指南\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">高\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>两段文字可能包含几乎相同的措辞。因此，相似度分数可能将它们排得非常接近。\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>Semantic relevance\n        ≠\nSource authority\n        ≠\nApplicability\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>当前已批准的HR政策通常应占主导地位，因为它是该主张的管辖来源。较旧的指南仍然是相关文本，但不应具有同等的证据权重。\u003C\u002Fp>\n\u003Ch2 id=\"section-19\">示例失效之处\u003C\u002Fh2>\n\u003Cp>权威并不总是一个每次都由某一来源胜出的普遍层级。\u003C\u002Fp>\n\u003Cp>事务型数据库可能是订单当前状态的权威来源。产品手册可能是预期操作的权威来源。已签署的合同可能控制商业义务。原始科学论文可能是其实验的最佳来源，而后续的系统综述可能对更广泛的结论更有力。\u003C\u002Fp>\n\u003Cp>因此，权威是相对于主张而言的。正确的来源类别取决于所问的问题。\u003C\u002Fp>\n\u003Ch2 id=\"section-23\">直接回答\u003C\u002Fh2>\n\u003Cp>当某一来源不仅相关，而且对于所回答的具体主张而言足够权威、及时、适用、可追溯且可访问时，它才应成为优先的检索证据。\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>Useful retrieval evidence\n=\nrelevance\n+ authority\n+ provenance\n+ freshness\n+ applicability\n+ accessibility\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>这些属性并不保证真实性。它们为决定哪些检索到的信息值得进入推理过程提供了更强的基础。\u003C\u002Fp>\n\u003Ch2 id=\"section-27\">为什么仅靠相似性不够\u003C\u002Fh2>\n\u003Cp>向量检索非常有用，因为即使措辞不同，它也能找到语义相关的信息。但语义相似性本身并不知道一份文档是否已过时、被取代、非官方、未经授权或在另一司法管辖区有效。\u003C\u002Fp>\n\u003Cp>最初的RAG架构展示了外部非参数记忆如何增强模型知识。生产系统必须在该检索过程周围添加特定于应用的控制。\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhere-does-an-llm-get-its-data-rag-data-sources-in-python\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">LLM从哪里获取数据？Python中的RAG数据源\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">文件、数据库、API、搜索和向量索引都可以提供运行时知识。检索方法应与来源和问题相匹配。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ch2 id=\"section-31\">六个权威信号\u003C\u002Fh2>\n\u003Ch3 id=\"section-32\">1. 来源身份\u003C\u002Fh3>\n\u003Cp>系统应知道来源实际上是什么：政策、合同、API、数据库记录、研究论文、文档页面、用户文件或第三方评论。\u003C\u002Fp>\n\u003Ch3 id=\"section-34\">2. 来源出处\u003C\u002Fh3>\n\u003Cp>证据应保留其来源、如何进入系统，以及如果该主张受到质疑，可以检查哪个原始对象或文档。\u003C\u002Fp>\n\u003Ch3 id=\"section-36\">3. 新鲜度\u003C\u002Fh3>\n\u003Cp>当底层事实发生变化时，时效性至关重要。一个当前的API响应和一篇五年前的文章，可能对查询的词语同样相关，但对其时间要求却并非如此。\u003C\u002Fp>\n\u003Ch3 id=\"section-38\">4. 适用性\u003C\u002Fh3>\n\u003Cp>一个事实可能是正确的，但适用于错误的产品版本、国家、租户、客户群、合同、软件发布版本或时间段。\u003C\u002Fp>\n\u003Ch3 id=\"section-40\">5. 证据特异性\u003C\u002Fh3>\n\u003Cp>直接支持该主张的原始证据通常比仅仅重复该主张的一系列摘要更有力。\u003C\u002Fp>\n\u003Ch3 id=\"section-42\">6. 访问合法性\u003C\u002Fh3>\n\u003Cp>一份技术上可检索的文档并不一定是被授权的上下文。检索权限必须在租户、权限、机密性和数据最小化的边界内运作。\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhat-is-context-engineering-what-the-model-receives-before-it-answers\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">什么是上下文工程？模型在回答前接收什么\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">上下文组装也是一个有效性和安全性边界。检索到的信息必须保留范围、权限和访问约束。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ch2 id=\"section-45\">一个最小权限模型\u003C\u002Fh2>\n\u003Cp>一个实用的系统可以显式地表示这些属性，而不是寄希望于相似度排序能够捕捉它们。\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>{\n  "source": "hr-policy-6.2",\n  "sourceType": "approved_policy",\n  "effectiveFrom": "2026-01-01",\n  "supersedes": "hr-policy-5.9",\n  "jurisdiction": "DE",\n  "authority": "primary",\n  "accessScope": "employee",\n  "provenance": "hr-policy-system"\n}\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>检索器仍然可以使用语义或词汇相关性。权限层添加了仅靠相关性无法提供的约束和重排序信号。\u003C\u002Fp>\n\u003Ch2 id=\"section-49\">与代理就绪信息的连接\u003C\u002Fh2>\n\u003Cp>当消费者不是阅读文档的人类，而是做出决策的AI代理时，这一点变得尤为重要。\u003C\u002Fp>\n\u003Cp>一个代理就绪的信息对象应在检索阶段之前保留身份、来源、冲突、未知项和有效性条件。否则，系统会试图在信息已经被扁平化为匿名文本之后重建权限。\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fportfolio\u002Ffrom-raw-data-to-verifiable-ai-context-a-practical-architecture-for-agent-ready-information\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">从原始数据到可验证的AI上下文：代理就绪信息的实用架构\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">一种投资组合架构，用于将异构数据转化为具有身份、来源、冲突和有效性的受控信息对象。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ch2 id=\"section-53\">诊断\u002F决策方法\u003C\u002Fh2>\n\u003Cp>对于每一个高价值的检索主张，问六个问题：\u003C\u002Fp>\n\u003Col>\u003Cli>这一证据究竟来自哪个确切来源？\u003C\u002Fli>\u003Cli>该来源对此类主张是否具有权威性？\u003C\u002Fli>\u003Cli>该信息是否仍然是最新的？\u003C\u002Fli>\u003Cli>它是否适用于该用户、版本、司法管辖区和时间段？\u003C\u002Fli>\u003Cli>该主张能否追溯到原始证据？\u003C\u002Fli>\u003Cli>用户或代理是否被授权检索它？\u003C\u002Fli>\u003C\u002Fol>\n\u003Cp>如果系统无法回答这些问题，问题就不再仅仅是检索质量。这是一个证据治理问题。\u003C\u002Fp>\n\u003Ch2 id=\"section-57\">常见失败模式\u003C\u002Fh2>\n\u003Cp>相似度得分最高者胜出。这把语义接近性与证据权威性混为一谈。\u003C\u002Fp>\n\u003Cp>最新的来源总是胜出。较新的评论可能不如较旧的管辖合同或标准具有权威性。\u003C\u002Fp>\n\u003Cp>官方意味着普遍适用。官方来源仍可能属于另一个司法管辖区、版本或范围。\u003C\u002Fp>\n\u003Cp>公开意味着可以安全检索。信息可以是公开的，却不适合该任务，而私人信息只有在授权情境内才可能是正当的。\u003C\u002Fp>\n\u003Ch2 id=\"section-62\">边缘情况：相互冲突的权威来源\u003C\u002Fh2>\n\u003Cp>困难的情况不是权威来源与明显薄弱来源之间的对比。而是两个看似权威却相互矛盾的来源。\u003C\u002Fp>\n\u003Cp>此时，权威层应保留冲突，而不是将其悄然合并为一个答案。\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>Retrieve\n   ↓\nQualify source authority\n   ↓\nDetect conflict\n   ↓\nPreserve disagreement\n   ↓\nReason about scope \u002F version \u002F provenance\n   ↓\nAnswer only inside the supported boundary\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Ch2 id=\"section-66\">局限性\u003C\u002Fh2>\n\u003Cp>检索权威层是一种概念架构，而不是通用排名算法。\u003C\u002Fp>\n\u003Cp>不同领域需要不同的权威模型。法律、科学、商业、运营和个人数据不能都使用相同的来源层级。\u003C\u002Fp>\n\u003Cp>它也无法保证外部搜索或AI系统会对发布者的页面进行排名。外部排名系统仍不受发布者控制。\u003C\u002Fp>\n\u003Ch2 id=\"section-70\">什么会改变这个答案？\u003C\u002Fh2>\n\u003Cp>检索系统可能越来越多地自动学习来源可靠性，将结构化来源与神经排名相结合，并通过代理协议协商访问。\u003C\u002Fp>\n\u003Cp>这可能会改变权威信号的实现方式。它并不会消除找到相关信息与判断该信息是否值得作为证据权重之间的区别。\u003C\u002Fp>\n\u003Ch2 id=\"section-73\">结论\u003C\u002Fh2>\n\u003Cp>检索触发器回答：我需要外部证据吗？\u003C\u002Fp>\n\u003Cp>检索权威层回答下一个问题：哪个外部来源值得成为证据？\u003C\u002Fp>\n\u003Cpre class=\"code-block\">\u003Ccode>Model Knowledge\n   ↓\nRetrieval Trigger\n   ↓\nCandidate Sources\n   ↓\nRetrieval Authority Layer\n   ↓\nQualified Evidence\n   ↓\nReasoning\n   ↓\nAnswer\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>这一区分将RAG从文档检索转向受控的证据检索。\u003C\u002Fp>\n\u003Ch2 id=\"section-78\">主要来源\u003C\u002Fh2>\n\u003Cp>Patrick Lewis 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">面向知识密集型NLP任务的检索增强生成\u003C\u002Fa>（2020年）。\u003C\u002Fp>\n\u003Cp>Akari Asai 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\" target=\"_blank\">Self-RAG：通过自我反思学习检索、生成和批判\u003C\u002Fa>（2023年）。\u003C\u002Fp>\n\u003Cp>Soyeong Jeong 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\" target=\"_blank\">Adaptive-RAG：通过问题复杂度学习自适应检索增强大语言模型\u003C\u002Fa>（2024年）。\u003C\u002Fp>",{"time":228,"blocks":229,"version":683},1792072664509,[230,236,242,247,252,260,267,272,277,282,287,293,304,309,314,331,336,341,346,351,356,361,366,371,376,381,386,391,396,401,410,415,420,425,430,435,440,445,450,455,460,465,470,475,484,489,494,499,504,509,514,519,528,533,538,553,558,563,568,573,578,583,588,593,598,603,608,613,618,623,628,633,638,643,648,653,658,663,668,673,678],{"id":231,"data":232,"type":42,"tunes":235},"XpcNQW15-B",{"text":233,"level":234},"问题",2,{},{"id":237,"data":238,"type":240,"tunes":241},"ren2ISkSca",{"text":239},"为什么AI检索系统会选择其中一个来源，而忽略另一个看似讨论完全相同主题的来源？","paragraph",{},{"id":243,"data":244,"type":240,"tunes":246},"X_bzzHXG4I",{"text":245},"语义相关性只是答案的一部分。两个文档可能与查询同样相似，但在权威性、时效性、来源、范围、权限和适用性方面存在根本差异。",{},{"id":248,"data":249,"type":240,"tunes":251},"elm6ePGXbF",{"text":250},"本文使用术语“检索权威层”来指代位于发现潜在相关信息与将该信息视为合适证据之间的来源选择边界。",{},{"id":253,"data":254,"type":258,"tunes":259},"TeJ97szQOn",{"text":255,"caption":256,"alignment":257},"检索权威层是一组来源选择规则和来源属性，用于确定哪些可检索信息值得成为特定问题的证据。","工作定义——本文使用的系统概念，并非标准化研究术语。","left","quote",{},{"id":261,"data":262,"type":265,"tunes":266},"JfN6zo2EAW",{"title":263,"maxLevel":264,"minLevel":234},"目录",3,"tableOfContents",{},{"id":268,"data":269,"type":42,"tunes":271},"RoTT8hQShH",{"text":270,"level":234},"这真正意味着什么",{},{"id":273,"data":274,"type":240,"tunes":276},"Cns53IeLiV",{"text":275},"检索器回答的是一个相对狭窄的问题：哪些条目看起来与此查询相关？",{},{"id":278,"data":279,"type":240,"tunes":281},"2qvTGNQESy",{"text":280},"生产级AI系统通常需要回答一个更难的问题：这些相关条目中，哪些应被允许影响答案？",{},{"id":283,"data":284,"type":240,"tunes":286},"KneZrq8Dgr",{"text":285},"第二个问题引入了来源身份、版本、权威性、有效期、来源和范围。一份相关的旧政策、复制的博客文章或非官方摘要可能排名极高，但对于该决策而言仍然是错误的证据。",{},{"id":288,"data":289,"type":291,"tunes":292},"DYVDznB9AW",{"code":290},"Question\n   ↓\nRetrieval Trigger\n   ↓\nCandidate Sources\n   ↓\nRETRIEVAL AUTHORITY LAYER\n   ↓\nQualified Evidence\n   ↓\nContext \u002F Reasoning\n   ↓\nAnswer Validity Boundary\n   ↓\nAnswer","code",{},{"id":294,"data":295,"type":302,"tunes":303},"mBPGlmAz5D",{"link":296,"meta":297},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger",{"image":298,"title":300,"description":301},{"url":299},"","AI何时应停止信任自身知识？——检索触发器","检索触发器决定何时需要外部证据。检索权威层则决定哪些来源应提供证据。","linkTool",{},{"id":305,"data":306,"type":42,"tunes":308},"zuuDXruzRL",{"text":307,"level":234},"最简单的例子",{},{"id":310,"data":311,"type":240,"tunes":313},"QP0qUL8Ciw",{"text":312},"假设一名员工问企业助手：员工有多少天年假？",{},{"id":315,"data":316,"type":329,"tunes":330},"2Ev2t8opWr",{"content":317,"stretched":43,"withHeadings":14},[318,322,326],[319,320,321],"候选","内容","语义相关性",[323,324,325],"A","当前HR政策，已批准版本6.2","高",[327,328,325],"B","从版本4.8复制的旧入职指南","table",{},{"id":332,"data":333,"type":240,"tunes":335},"wtPrU1Dnqd",{"text":334},"两段文字可能包含几乎相同的措辞。因此，相似度分数可能将它们排得非常接近。",{},{"id":337,"data":338,"type":291,"tunes":340},"v97zTkFBzt",{"code":339},"Semantic relevance\n        ≠\nSource authority\n        ≠\nApplicability",{},{"id":342,"data":343,"type":240,"tunes":345},"kAKa32Hh_B",{"text":344},"当前已批准的HR政策通常应占主导地位，因为它是该主张的管辖来源。较旧的指南仍然是相关文本，但不应具有同等的证据权重。",{},{"id":347,"data":348,"type":42,"tunes":350},"gUE1_pfNG9",{"text":349,"level":234},"示例失效之处",{},{"id":352,"data":353,"type":240,"tunes":355},"3c0r-JJC4s",{"text":354},"权威并不总是一个每次都由某一来源胜出的普遍层级。",{},{"id":357,"data":358,"type":240,"tunes":360},"q1E1rsEJq0",{"text":359},"事务型数据库可能是订单当前状态的权威来源。产品手册可能是预期操作的权威来源。已签署的合同可能控制商业义务。原始科学论文可能是其实验的最佳来源，而后续的系统综述可能对更广泛的结论更有力。",{},{"id":362,"data":363,"type":240,"tunes":365},"oCUaHjt58Z",{"text":364},"因此，权威是相对于主张而言的。正确的来源类别取决于所问的问题。",{},{"id":367,"data":368,"type":42,"tunes":370},"4CxlnDa9Cc",{"text":369,"level":234},"直接回答",{},{"id":372,"data":373,"type":240,"tunes":375},"QW5_w33xWI",{"text":374},"当某一来源不仅相关，而且对于所回答的具体主张而言足够权威、及时、适用、可追溯且可访问时，它才应成为优先的检索证据。",{},{"id":377,"data":378,"type":291,"tunes":380},"owyjAlbwY4",{"code":379},"Useful retrieval evidence\n=\nrelevance\n+ authority\n+ provenance\n+ freshness\n+ applicability\n+ accessibility",{},{"id":382,"data":383,"type":240,"tunes":385},"IBCawe0zx3",{"text":384},"这些属性并不保证真实性。它们为决定哪些检索到的信息值得进入推理过程提供了更强的基础。",{},{"id":387,"data":388,"type":42,"tunes":390},"4vZIEUjrdq",{"text":389,"level":234},"为什么仅靠相似性不够",{},{"id":392,"data":393,"type":240,"tunes":395},"1PFBw4V0ni",{"text":394},"向量检索非常有用，因为即使措辞不同，它也能找到语义相关的信息。但语义相似性本身并不知道一份文档是否已过时、被取代、非官方、未经授权或在另一司法管辖区有效。",{},{"id":397,"data":398,"type":240,"tunes":400},"vfaNbSsKCP",{"text":399},"最初的RAG架构展示了外部非参数记忆如何增强模型知识。生产系统必须在该检索过程周围添加特定于应用的控制。",{},{"id":402,"data":403,"type":302,"tunes":409},"RKi9NxoiRb",{"link":404,"meta":405},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhere-does-an-llm-get-its-data-rag-data-sources-in-python",{"image":406,"title":407,"description":408},{"url":299},"LLM从哪里获取数据？Python中的RAG数据源","文件、数据库、API、搜索和向量索引都可以提供运行时知识。检索方法应与来源和问题相匹配。",{},{"id":411,"data":412,"type":42,"tunes":414},"hULFRV_NCz",{"text":413,"level":234},"六个权威信号",{},{"id":416,"data":417,"type":42,"tunes":419},"7lJ3RudSld",{"text":418,"level":264},"1. 来源身份",{},{"id":421,"data":422,"type":240,"tunes":424},"J2JcDKRmnb",{"text":423},"系统应知道来源实际上是什么：政策、合同、API、数据库记录、研究论文、文档页面、用户文件或第三方评论。",{},{"id":426,"data":427,"type":42,"tunes":429},"bKliUrdBTI",{"text":428,"level":264},"2. 来源出处",{},{"id":431,"data":432,"type":240,"tunes":434},"LRJssGYfAY",{"text":433},"证据应保留其来源、如何进入系统，以及如果该主张受到质疑，可以检查哪个原始对象或文档。",{},{"id":436,"data":437,"type":42,"tunes":439},"N-KXk6Cp2H",{"text":438,"level":264},"3. 新鲜度",{},{"id":441,"data":442,"type":240,"tunes":444},"FJKVbQVgOr",{"text":443},"当底层事实发生变化时，时效性至关重要。一个当前的API响应和一篇五年前的文章，可能对查询的词语同样相关，但对其时间要求却并非如此。",{},{"id":446,"data":447,"type":42,"tunes":449},"8Wxcbhl365",{"text":448,"level":264},"4. 适用性",{},{"id":451,"data":452,"type":240,"tunes":454},"8w01u-EYq6",{"text":453},"一个事实可能是正确的，但适用于错误的产品版本、国家、租户、客户群、合同、软件发布版本或时间段。",{},{"id":456,"data":457,"type":42,"tunes":459},"NYQm_3y_PX",{"text":458,"level":264},"5. 证据特异性",{},{"id":461,"data":462,"type":240,"tunes":464},"uqqpEuQdxC",{"text":463},"直接支持该主张的原始证据通常比仅仅重复该主张的一系列摘要更有力。",{},{"id":466,"data":467,"type":42,"tunes":469},"GggJkQwr1C",{"text":468,"level":264},"6. 访问合法性",{},{"id":471,"data":472,"type":240,"tunes":474},"spnPYg1KRU",{"text":473},"一份技术上可检索的文档并不一定是被授权的上下文。检索权限必须在租户、权限、机密性和数据最小化的边界内运作。",{},{"id":476,"data":477,"type":302,"tunes":483},"yvkNCwWFiC",{"link":478,"meta":479},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhat-is-context-engineering-what-the-model-receives-before-it-answers",{"image":480,"title":481,"description":482},{"url":299},"什么是上下文工程？模型在回答前接收什么","上下文组装也是一个有效性和安全性边界。检索到的信息必须保留范围、权限和访问约束。",{},{"id":485,"data":486,"type":42,"tunes":488},"b_9T6lmCVJ",{"text":487,"level":234},"一个最小权限模型",{},{"id":490,"data":491,"type":240,"tunes":493},"VpHVLljSPR",{"text":492},"一个实用的系统可以显式地表示这些属性，而不是寄希望于相似度排序能够捕捉它们。",{},{"id":495,"data":496,"type":291,"tunes":498},"L4W7ZQSTYs",{"code":497},"{\n  \"source\": \"hr-policy-6.2\",\n  \"sourceType\": \"approved_policy\",\n  \"effectiveFrom\": \"2026-01-01\",\n  \"supersedes\": \"hr-policy-5.9\",\n  \"jurisdiction\": \"DE\",\n  \"authority\": \"primary\",\n  \"accessScope\": \"employee\",\n  \"provenance\": \"hr-policy-system\"\n}",{},{"id":500,"data":501,"type":240,"tunes":503},"VISBdGcdtA",{"text":502},"检索器仍然可以使用语义或词汇相关性。权限层添加了仅靠相关性无法提供的约束和重排序信号。",{},{"id":505,"data":506,"type":42,"tunes":508},"GcbcUHNia6",{"text":507,"level":234},"与代理就绪信息的连接",{},{"id":510,"data":511,"type":240,"tunes":513},"CVqLmnn5CD",{"text":512},"当消费者不是阅读文档的人类，而是做出决策的AI代理时，这一点变得尤为重要。",{},{"id":515,"data":516,"type":240,"tunes":518},"MkvgpsEqJc",{"text":517},"一个代理就绪的信息对象应在检索阶段之前保留身份、来源、冲突、未知项和有效性条件。否则，系统会试图在信息已经被扁平化为匿名文本之后重建权限。",{},{"id":520,"data":521,"type":302,"tunes":527},"5KX3sugIpt",{"link":522,"meta":523},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fportfolio\u002Ffrom-raw-data-to-verifiable-ai-context-a-practical-architecture-for-agent-ready-information",{"image":524,"title":525,"description":526},{"url":299},"从原始数据到可验证的AI上下文：代理就绪信息的实用架构","一种投资组合架构，用于将异构数据转化为具有身份、来源、冲突和有效性的受控信息对象。",{},{"id":529,"data":530,"type":42,"tunes":532},"O6bYy5LenW",{"text":531,"level":234},"诊断\u002F决策方法",{},{"id":534,"data":535,"type":240,"tunes":537},"cZeNmj3Z_3",{"text":536},"对于每一个高价值的检索主张，问六个问题：",{},{"id":539,"data":540,"type":551,"tunes":552},"lWP8qNNrxw",{"meta":541,"items":543,"style":550},{"counterType":542},"numeric",[544,545,546,547,548,549],"这一证据究竟来自哪个确切来源？","该来源对此类主张是否具有权威性？","该信息是否仍然是最新的？","它是否适用于该用户、版本、司法管辖区和时间段？","该主张能否追溯到原始证据？","用户或代理是否被授权检索它？","ordered","list",{},{"id":554,"data":555,"type":240,"tunes":557},"zzgllTPRBV",{"text":556},"如果系统无法回答这些问题，问题就不再仅仅是检索质量。这是一个证据治理问题。",{},{"id":559,"data":560,"type":42,"tunes":562},"gmdM4uPGYl",{"text":561,"level":234},"常见失败模式",{},{"id":564,"data":565,"type":240,"tunes":567},"GjbXINllzx",{"text":566},"相似度得分最高者胜出。这把语义接近性与证据权威性混为一谈。",{},{"id":569,"data":570,"type":240,"tunes":572},"DapWd1cmb5",{"text":571},"最新的来源总是胜出。较新的评论可能不如较旧的管辖合同或标准具有权威性。",{},{"id":574,"data":575,"type":240,"tunes":577},"VVGLrdBjxS",{"text":576},"官方意味着普遍适用。官方来源仍可能属于另一个司法管辖区、版本或范围。",{},{"id":579,"data":580,"type":240,"tunes":582},"vBrSvczrCq",{"text":581},"公开意味着可以安全检索。信息可以是公开的，却不适合该任务，而私人信息只有在授权情境内才可能是正当的。",{},{"id":584,"data":585,"type":42,"tunes":587},"gN24OPQDE2",{"text":586,"level":234},"边缘情况：相互冲突的权威来源",{},{"id":589,"data":590,"type":240,"tunes":592},"qcST_7WVLw",{"text":591},"困难的情况不是权威来源与明显薄弱来源之间的对比。而是两个看似权威却相互矛盾的来源。",{},{"id":594,"data":595,"type":240,"tunes":597},"E9gsL-vSes",{"text":596},"此时，权威层应保留冲突，而不是将其悄然合并为一个答案。",{},{"id":599,"data":600,"type":291,"tunes":602},"CBZ5ctc6_X",{"code":601},"Retrieve\n   ↓\nQualify source authority\n   ↓\nDetect conflict\n   ↓\nPreserve disagreement\n   ↓\nReason about scope \u002F version \u002F provenance\n   ↓\nAnswer only inside the supported boundary",{},{"id":604,"data":605,"type":42,"tunes":607},"xeQqLNvjCb",{"text":606,"level":234},"局限性",{},{"id":609,"data":610,"type":240,"tunes":612},"raTFpBhsAR",{"text":611},"检索权威层是一种概念架构，而不是通用排名算法。",{},{"id":614,"data":615,"type":240,"tunes":617},"G6tXMIP-u1",{"text":616},"不同领域需要不同的权威模型。法律、科学、商业、运营和个人数据不能都使用相同的来源层级。",{},{"id":619,"data":620,"type":240,"tunes":622},"SWCg4MmSIb",{"text":621},"它也无法保证外部搜索或AI系统会对发布者的页面进行排名。外部排名系统仍不受发布者控制。",{},{"id":624,"data":625,"type":42,"tunes":627},"j_xI-ClxxP",{"text":626,"level":234},"什么会改变这个答案？",{},{"id":629,"data":630,"type":240,"tunes":632},"xrctSdgP3p",{"text":631},"检索系统可能越来越多地自动学习来源可靠性，将结构化来源与神经排名相结合，并通过代理协议协商访问。",{},{"id":634,"data":635,"type":240,"tunes":637},"Ka8dOdJHXO",{"text":636},"这可能会改变权威信号的实现方式。它并不会消除找到相关信息与判断该信息是否值得作为证据权重之间的区别。",{},{"id":639,"data":640,"type":42,"tunes":642},"E6SLB3m5CI",{"text":641,"level":234},"结论",{},{"id":644,"data":645,"type":240,"tunes":647},"0_oDIxt-xd",{"text":646},"检索触发器回答：我需要外部证据吗？",{},{"id":649,"data":650,"type":240,"tunes":652},"mCMupYycYk",{"text":651},"检索权威层回答下一个问题：哪个外部来源值得成为证据？",{},{"id":654,"data":655,"type":291,"tunes":657},"rSohO6bowX",{"code":656},"Model Knowledge\n   ↓\nRetrieval Trigger\n   ↓\nCandidate Sources\n   ↓\nRetrieval Authority Layer\n   ↓\nQualified Evidence\n   ↓\nReasoning\n   ↓\nAnswer",{},{"id":659,"data":660,"type":240,"tunes":662},"f8GMr_eelV",{"text":661},"这一区分将RAG从文档检索转向受控的证据检索。",{},{"id":664,"data":665,"type":42,"tunes":667},"hOWp43J5Dw",{"text":666,"level":234},"主要来源",{},{"id":669,"data":670,"type":240,"tunes":672},"evqsbllTgC",{"text":671},"Patrick Lewis 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">面向知识密集型NLP任务的检索增强生成\u003C\u002Fa>（2020年）。",{},{"id":674,"data":675,"type":240,"tunes":677},"gz0XrNndq4",{"text":676},"Akari Asai 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\" target=\"_blank\">Self-RAG：通过自我反思学习检索、生成和批判\u003C\u002Fa>（2023年）。",{},{"id":679,"data":680,"type":240,"tunes":682},"NgkPqz5Poz",{"text":681},"Soyeong Jeong 等人，\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\" target=\"_blank\">Adaptive-RAG：通过问题复杂度学习自适应检索增强大语言模型\u003C\u002Fa>（2024年）。",{},"2.31","找到相关文档还不够。AI系统还需要区分哪些来源是当前的、权威的、适用的和可追溯的，哪些来源只是看起来相似。检索权威层定义了这种来源选择的边界。","\u002Fuploads\u002F2026\u002F10\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer-1792072410616-0evp1z.webp","why-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer-1792072410616-0evp1z","PUBLISHED","2026-10-15T07:51:00.000Z","2026-10-15T13:51:04.579Z","2026-10-15T20:37:39.839Z",{"en":692,"de":693,"sr":694,"es":695,"fr":696,"it":697,"ru":698,"zh":699},"\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fde\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fsr\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fes\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Ffr\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fit\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fru\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer","\u002Fzh\u002Fblog\u002Fwhy-does-ai-retrieve-some-sources-and-ignore-others-the-retrieval-authority-layer",[701,705],{"id":702,"name":703,"slug":704},65,"主题权威体系","topical-authority",{"id":706,"name":707,"slug":708},106,"AI回滚运行手册","ai-rollback",{"id":710,"login":711,"email":299,"displayName":712},"20","rooth8233","Aleksandar Stajić",[714,1074],{"lang":715,"title":716,"content":717,"contentJson":718,"excerpt":1073},"en","Why Does AI Retrieve Some Sources and Ignore Others? — The Retrieval Authority Layer","{\"time\":1792072370632,\"blocks\":[{\"id\":\"XpcNQW15-B\",\"data\":{\"text\":\"Question\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"ren2ISkSca\",\"data\":{\"text\":\"Why does an AI retrieval system choose one source while ignoring another source that appears to discuss exactly the same subject?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"X_bzzHXG4I\",\"data\":{\"text\":\"Semantic relevance is only part of the answer. Two documents can be equally similar to a query while differing radically in authority, freshness, provenance, scope, permissions and applicability.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"elm6ePGXbF\",\"data\":{\"text\":\"This article uses the term Retrieval Authority Layer for the source-selection boundary that sits between finding potentially relevant information and treating that information as suitable evidence.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"TeJ97szQOn\",\"data\":{\"text\":\"A Retrieval Authority Layer is the set of source-selection rules and source properties that determine which retrievable information deserves to become evidence for a particular question.\",\"caption\":\"Working definition — a systems concept used in this article, not a standardized research term.\",\"alignment\":\"left\"},\"type\":\"quote\",\"tunes\":{}},{\"id\":\"JfN6zo2EAW\",\"data\":{\"title\":\"Contents\",\"maxLevel\":3,\"minLevel\":2},\"type\":\"tableOfContents\",\"tunes\":{}},{\"id\":\"RoTT8hQShH\",\"data\":{\"text\":\"What This Really Means\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"Cns53IeLiV\",\"data\":{\"text\":\"A retriever answers a relatively narrow question: Which items appear relevant to this query?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"2qvTGNQESy\",\"data\":{\"text\":\"A production AI system usually needs to answer a harder one: Which of those relevant items should be allowed to influence the answer?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"KneZrq8Dgr\",\"data\":{\"text\":\"That second question introduces source identity, version, authority, validity period, provenance and scope. A relevant old policy, copied blog post or unofficial summary can rank extremely well while still being the wrong evidence for the decision.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"DYVDznB9AW\",\"data\":{\"code\":\"Question\\n   ↓\\nRetrieval Trigger\\n   ↓\\nCandidate Sources\\n   ↓\\nRETRIEVAL AUTHORITY LAYER\\n   ↓\\nQualified Evidence\\n   ↓\\nContext \u002F Reasoning\\n   ↓\\nAnswer Validity Boundary\\n   ↓\\nAnswer\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"mBPGlmAz5D\",\"data\":{\"link\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"When Should an AI Stop Trusting Its Own Knowledge? — The Retrieval Trigger\",\"description\":\"The Retrieval Trigger decides when external evidence is required. The Retrieval Authority Layer asks which sources should provide it.\"}},\"type\":\"linkTool\",\"tunes\":{}},{\"id\":\"zuuDXruzRL\",\"data\":{\"text\":\"Simplest Example\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"QP0qUL8Ciw\",\"data\":{\"text\":\"Imagine an employee asks an enterprise assistant: How many vacation days do employees receive?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"2Ev2t8opWr\",\"data\":{\"content\":[[\"Candidate\",\"Content\",\"Semantic relevance\"],[\"A\",\"Current HR policy, approved version 6.2\",\"High\"],[\"B\",\"Old onboarding guide copied from version 4.8\",\"High\"]],\"stretched\":false,\"withHeadings\":true},\"type\":\"table\",\"tunes\":{}},{\"id\":\"wtPrU1Dnqd\",\"data\":{\"text\":\"Both passages can contain nearly identical wording. A similarity score may therefore place them very close together.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"v97zTkFBzt\",\"data\":{\"code\":\"Semantic relevance\\n        ≠\\nSource authority\\n        ≠\\nApplicability\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"kAKa32Hh_B\",\"data\":{\"text\":\"The current approved HR policy should normally dominate because it is the governing source for the claim. The older guide remains relevant text, but it should not carry the same evidential weight.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"gUE1_pfNG9\",\"data\":{\"text\":\"Where the Example Stops Working\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"3c0r-JJC4s\",\"data\":{\"text\":\"Authority is not always a universal hierarchy in which one source wins every time.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"q1E1rsEJq0\",\"data\":{\"text\":\"A transactional database may be authoritative for the current status of an order. A product manual may be authoritative for intended operation. A signed contract may control a commercial obligation. A primary scientific paper may be the best source for its experiment, while a later systematic review may be stronger for a broader conclusion.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"oCUaHjt58Z\",\"data\":{\"text\":\"Authority is therefore claim-relative. The correct source class depends on what is being asked.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"4CxlnDa9Cc\",\"data\":{\"text\":\"Direct Answer\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"QW5_w33xWI\",\"data\":{\"text\":\"A source should become preferred retrieval evidence when it is not merely relevant, but also sufficiently authoritative, current, applicable, traceable and accessible for the specific claim being answered.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"owyjAlbwY4\",\"data\":{\"code\":\"Useful retrieval evidence\\n=\\nrelevance\\n+ authority\\n+ provenance\\n+ freshness\\n+ applicability\\n+ accessibility\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"IBCawe0zx3\",\"data\":{\"text\":\"These properties do not guarantee truth. They create a stronger basis for deciding which retrieved information deserves to enter the reasoning process.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"4vZIEUjrdq\",\"data\":{\"text\":\"Why Similarity Alone Is Not Enough\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"1PFBw4V0ni\",\"data\":{\"text\":\"Vector retrieval is extremely useful because it can find semantically related information even when the wording differs. But semantic similarity does not know by itself whether a document is obsolete, superseded, unofficial, unauthorized or valid for another jurisdiction.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"vfaNbSsKCP\",\"data\":{\"text\":\"The original RAG architecture showed how external non-parametric memory can augment model knowledge. Production systems must add application-specific controls around that retrieval process.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"RKi9NxoiRb\",\"data\":{\"link\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhere-does-an-llm-get-its-data-rag-data-sources-in-python\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Where Does an LLM Get Its Data? RAG Data Sources in Python\",\"description\":\"Files, databases, APIs, search and vector indexes can all provide runtime knowledge. The retrieval method should match the source and the question.\"}},\"type\":\"linkTool\",\"tunes\":{}},{\"id\":\"hULFRV_NCz\",\"data\":{\"text\":\"The Six Authority Signals\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"7lJ3RudSld\",\"data\":{\"text\":\"1. Source Identity\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"J2JcDKRmnb\",\"data\":{\"text\":\"The system should know what the source actually is: policy, contract, API, database record, research paper, documentation page, user file or third-party commentary.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"bKliUrdBTI\",\"data\":{\"text\":\"2. Provenance\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"LRJssGYfAY\",\"data\":{\"text\":\"Evidence should preserve where it came from, how it entered the system and which original object or document can be inspected if the claim is challenged.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"N-KXk6Cp2H\",\"data\":{\"text\":\"3. Freshness\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"FJKVbQVgOr\",\"data\":{\"text\":\"Freshness matters when the underlying fact changes. A current API response and a five-year-old article may be equally relevant to the words of a query but not to its temporal requirement.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"8Wxcbhl365\",\"data\":{\"text\":\"4. Applicability\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"8w01u-EYq6\",\"data\":{\"text\":\"A fact may be correct but applicable to the wrong product version, country, tenant, customer group, contract, software release or time period.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"NYQm_3y_PX\",\"data\":{\"text\":\"5. Evidence Specificity\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"uqqpEuQdxC\",\"data\":{\"text\":\"Primary evidence that directly supports the claim is generally stronger than a chain of summaries that merely repeat it.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"GggJkQwr1C\",\"data\":{\"text\":\"6. Access Legitimacy\",\"level\":3},\"type\":\"header\",\"tunes\":{}},{\"id\":\"spnPYg1KRU\",\"data\":{\"text\":\"A technically retrievable document is not necessarily authorized context. Retrieval authority must operate inside tenant, permission, confidentiality and data-minimization boundaries.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"yvkNCwWFiC\",\"data\":{\"link\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhat-is-context-engineering-what-the-model-receives-before-it-answers\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"What Is Context Engineering? What the Model Receives Before It Answers\",\"description\":\"Context assembly is also a validity and security boundary. Retrieved information must preserve scope, authority and access constraints.\"}},\"type\":\"linkTool\",\"tunes\":{}},{\"id\":\"b_9T6lmCVJ\",\"data\":{\"text\":\"A Minimal Authority Model\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"VpHVLljSPR\",\"data\":{\"text\":\"A practical system can represent these properties explicitly rather than hoping that similarity ranking captures them.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"L4W7ZQSTYs\",\"data\":{\"code\":\"{\\n  \\\"source\\\": \\\"hr-policy-6.2\\\",\\n  \\\"sourceType\\\": \\\"approved_policy\\\",\\n  \\\"effectiveFrom\\\": \\\"2026-01-01\\\",\\n  \\\"supersedes\\\": \\\"hr-policy-5.9\\\",\\n  \\\"jurisdiction\\\": \\\"DE\\\",\\n  \\\"authority\\\": \\\"primary\\\",\\n  \\\"accessScope\\\": \\\"employee\\\",\\n  \\\"provenance\\\": \\\"hr-policy-system\\\"\\n}\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"VISBdGcdtA\",\"data\":{\"text\":\"The retriever can still use semantic or lexical relevance. The authority layer adds constraints and reranking signals that relevance alone cannot provide.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"GcbcUHNia6\",\"data\":{\"text\":\"Connection to Agent-Ready Information\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"CVqLmnn5CD\",\"data\":{\"text\":\"This becomes especially important when the consumer is not a human reading documents but an AI agent making decisions.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"MkvgpsEqJc\",\"data\":{\"text\":\"An agent-ready information object should preserve identity, provenance, conflicts, unknowns and validity conditions before the retrieval stage. Otherwise the system tries to reconstruct authority after information has already been flattened into anonymous text.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"5KX3sugIpt\",\"data\":{\"link\":\"https:\u002F\u002Fstajic.de\u002Fportfolio\u002Ffrom-raw-data-to-verifiable-ai-context-a-practical-architecture-for-agent-ready-information\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"From Raw Data to Verifiable AI Context: A Practical Architecture for Agent-Ready Information\",\"description\":\"A portfolio architecture for turning heterogeneous data into controlled information objects with identity, provenance, conflicts and validity.\"}},\"type\":\"linkTool\",\"tunes\":{}},{\"id\":\"O6bYy5LenW\",\"data\":{\"text\":\"Diagnostic \u002F Decision Method\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"cZeNmj3Z_3\",\"data\":{\"text\":\"For every high-value retrieved claim, ask six questions:\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"lWP8qNNrxw\",\"data\":{\"meta\":{\"counterType\":\"numeric\"},\"items\":[\"What exact source produced this evidence?\",\"Is this source authoritative for this type of claim?\",\"Is the information still current?\",\"Does it apply to this user, version, jurisdiction and time period?\",\"Can the claim be traced back to the original evidence?\",\"Was the user or agent authorized to retrieve it?\"],\"style\":\"ordered\"},\"type\":\"list\",\"tunes\":{}},{\"id\":\"zzgllTPRBV\",\"data\":{\"text\":\"If the system cannot answer these questions, the problem is no longer only retrieval quality. It is an evidence-governance problem.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"gmdM4uPGYl\",\"data\":{\"text\":\"Common Failure Modes\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"GjbXINllzx\",\"data\":{\"text\":\"The highest similarity score wins. This confuses semantic closeness with evidential authority.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"DapWd1cmb5\",\"data\":{\"text\":\"The newest source always wins. A newer commentary may be less authoritative than an older governing contract or standard.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"VVGLrdBjxS\",\"data\":{\"text\":\"Official means universally applicable. An official source may still belong to another jurisdiction, version or scope.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"vBrSvczrCq\",\"data\":{\"text\":\"Public means safe to retrieve. Information can be public yet inappropriate for the task, while private information can be legitimate only inside an authorized context.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"gN24OPQDE2\",\"data\":{\"text\":\"Edge Case: Conflicting Authoritative Sources\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"qcST_7WVLw\",\"data\":{\"text\":\"The difficult case is not an authoritative source versus an obviously weak source. It is two apparently authoritative sources that disagree.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"E9gsL-vSes\",\"data\":{\"text\":\"At that point, the authority layer should preserve the conflict instead of silently collapsing it into one answer.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"CBZ5ctc6_X\",\"data\":{\"code\":\"Retrieve\\n   ↓\\nQualify source authority\\n   ↓\\nDetect conflict\\n   ↓\\nPreserve disagreement\\n   ↓\\nReason about scope \u002F version \u002F provenance\\n   ↓\\nAnswer only inside the supported boundary\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"xeQqLNvjCb\",\"data\":{\"text\":\"Limitations\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"raTFpBhsAR\",\"data\":{\"text\":\"The Retrieval Authority Layer is a conceptual architecture, not a universal ranking algorithm.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"G6tXMIP-u1\",\"data\":{\"text\":\"Different domains require different authority models. Legal, scientific, commercial, operational and personal data cannot all use the same source hierarchy.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"SWCg4MmSIb\",\"data\":{\"text\":\"It also cannot guarantee that an external search or AI system will rank a publisher's page. External ranking systems remain outside the publisher's control.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"j_xI-ClxxP\",\"data\":{\"text\":\"What Would Change This Answer?\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"xrctSdgP3p\",\"data\":{\"text\":\"Retrieval systems may increasingly learn source reliability automatically, combine structured provenance with neural ranking and negotiate access through agent protocols.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"Ka8dOdJHXO\",\"data\":{\"text\":\"That may change how authority signals are implemented. It does not remove the distinction between finding relevant information and deciding whether that information deserves evidential weight.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"E6SLB3m5CI\",\"data\":{\"text\":\"Conclusion\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"0_oDIxt-xd\",\"data\":{\"text\":\"The Retrieval Trigger answers: Do I need external evidence?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"mCMupYycYk\",\"data\":{\"text\":\"The Retrieval Authority Layer answers the next question: Which external source deserves to become evidence?\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"rSohO6bowX\",\"data\":{\"code\":\"Model Knowledge\\n   ↓\\nRetrieval Trigger\\n   ↓\\nCandidate Sources\\n   ↓\\nRetrieval Authority Layer\\n   ↓\\nQualified Evidence\\n   ↓\\nReasoning\\n   ↓\\nAnswer\"},\"type\":\"code\",\"tunes\":{}},{\"id\":\"f8GMr_eelV\",\"data\":{\"text\":\"That distinction moves RAG from document retrieval toward controlled evidence retrieval.\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"hOWp43J5Dw\",\"data\":{\"text\":\"Primary Sources\",\"level\":2},\"type\":\"header\",\"tunes\":{}},{\"id\":\"evqsbllTgC\",\"data\":{\"text\":\"Patrick Lewis et al., \u003Ca href=\\\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\\\" target=\\\"_blank\\\">Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks\u003C\u002Fa> (2020).\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"gz0XrNndq4\",\"data\":{\"text\":\"Akari Asai et al., \u003Ca href=\\\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\\\" target=\\\"_blank\\\">Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection\u003C\u002Fa> (2023).\"},\"type\":\"paragraph\",\"tunes\":{}},{\"id\":\"NgkPqz5Poz\",\"data\":{\"text\":\"Soyeong Jeong et al., \u003Ca href=\\\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\\\" target=\\\"_blank\\\">Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity\u003C\u002Fa> (2024).\"},\"type\":\"paragraph\",\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":719,"blocks":720,"version":1072},1792072370632,[721,725,729,733,737,742,746,750,754,758,762,765,773,777,781,794,798,801,805,809,813,817,821,825,829,832,836,840,844,848,856,860,864,868,872,876,880,884,888,892,896,900,904,908,916,920,924,927,931,935,939,943,951,955,959,970,974,978,982,986,990,994,998,1002,1006,1009,1013,1017,1021,1025,1029,1033,1037,1041,1045,1049,1052,1056,1060,1064,1068],{"id":231,"data":722,"type":42,"tunes":724},{"text":723,"level":234},"Question",{},{"id":237,"data":726,"type":240,"tunes":728},{"text":727},"Why does an AI retrieval system choose one source while ignoring another source that appears to discuss exactly the same subject?",{},{"id":243,"data":730,"type":240,"tunes":732},{"text":731},"Semantic relevance is only part of the answer. Two documents can be equally similar to a query while differing radically in authority, freshness, provenance, scope, permissions and applicability.",{},{"id":248,"data":734,"type":240,"tunes":736},{"text":735},"This article uses the term Retrieval Authority Layer for the source-selection boundary that sits between finding potentially relevant information and treating that information as suitable evidence.",{},{"id":253,"data":738,"type":258,"tunes":741},{"text":739,"caption":740,"alignment":257},"A Retrieval Authority Layer is the set of source-selection rules and source properties that determine which retrievable information deserves to become evidence for a particular question.","Working definition — a systems concept used in this article, not a standardized research term.",{},{"id":261,"data":743,"type":265,"tunes":745},{"title":744,"maxLevel":264,"minLevel":234},"Contents",{},{"id":268,"data":747,"type":42,"tunes":749},{"text":748,"level":234},"What This Really Means",{},{"id":273,"data":751,"type":240,"tunes":753},{"text":752},"A retriever answers a relatively narrow question: Which items appear relevant to this query?",{},{"id":278,"data":755,"type":240,"tunes":757},{"text":756},"A production AI system usually needs to answer a harder one: Which of those relevant items should be allowed to influence the answer?",{},{"id":283,"data":759,"type":240,"tunes":761},{"text":760},"That second question introduces source identity, version, authority, validity period, provenance and scope. A relevant old policy, copied blog post or unofficial summary can rank extremely well while still being the wrong evidence for the decision.",{},{"id":288,"data":763,"type":291,"tunes":764},{"code":290},{},{"id":294,"data":766,"type":302,"tunes":772},{"link":767,"meta":768},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger",{"image":769,"title":770,"description":771},{"url":299},"When Should an AI Stop Trusting Its Own Knowledge? — The Retrieval Trigger","The Retrieval Trigger decides when external evidence is required. The Retrieval Authority Layer asks which sources should provide it.",{},{"id":305,"data":774,"type":42,"tunes":776},{"text":775,"level":234},"Simplest Example",{},{"id":310,"data":778,"type":240,"tunes":780},{"text":779},"Imagine an employee asks an enterprise assistant: How many vacation days do employees receive?",{},{"id":315,"data":782,"type":329,"tunes":793},{"content":783,"stretched":43,"withHeadings":14},[784,788,791],[785,786,787],"Candidate","Content","Semantic relevance",[323,789,790],"Current HR policy, approved version 6.2","High",[327,792,790],"Old onboarding guide copied from version 4.8",{},{"id":332,"data":795,"type":240,"tunes":797},{"text":796},"Both passages can contain nearly identical wording. A similarity score may therefore place them very close together.",{},{"id":337,"data":799,"type":291,"tunes":800},{"code":339},{},{"id":342,"data":802,"type":240,"tunes":804},{"text":803},"The current approved HR policy should normally dominate because it is the governing source for the claim. The older guide remains relevant text, but it should not carry the same evidential weight.",{},{"id":347,"data":806,"type":42,"tunes":808},{"text":807,"level":234},"Where the Example Stops Working",{},{"id":352,"data":810,"type":240,"tunes":812},{"text":811},"Authority is not always a universal hierarchy in which one source wins every time.",{},{"id":357,"data":814,"type":240,"tunes":816},{"text":815},"A transactional database may be authoritative for the current status of an order. A product manual may be authoritative for intended operation. A signed contract may control a commercial obligation. A primary scientific paper may be the best source for its experiment, while a later systematic review may be stronger for a broader conclusion.",{},{"id":362,"data":818,"type":240,"tunes":820},{"text":819},"Authority is therefore claim-relative. The correct source class depends on what is being asked.",{},{"id":367,"data":822,"type":42,"tunes":824},{"text":823,"level":234},"Direct Answer",{},{"id":372,"data":826,"type":240,"tunes":828},{"text":827},"A source should become preferred retrieval evidence when it is not merely relevant, but also sufficiently authoritative, current, applicable, traceable and accessible for the specific claim being answered.",{},{"id":377,"data":830,"type":291,"tunes":831},{"code":379},{},{"id":382,"data":833,"type":240,"tunes":835},{"text":834},"These properties do not guarantee truth. They create a stronger basis for deciding which retrieved information deserves to enter the reasoning process.",{},{"id":387,"data":837,"type":42,"tunes":839},{"text":838,"level":234},"Why Similarity Alone Is Not Enough",{},{"id":392,"data":841,"type":240,"tunes":843},{"text":842},"Vector retrieval is extremely useful because it can find semantically related information even when the wording differs. But semantic similarity does not know by itself whether a document is obsolete, superseded, unofficial, unauthorized or valid for another jurisdiction.",{},{"id":397,"data":845,"type":240,"tunes":847},{"text":846},"The original RAG architecture showed how external non-parametric memory can augment model knowledge. Production systems must add application-specific controls around that retrieval process.",{},{"id":402,"data":849,"type":302,"tunes":855},{"link":850,"meta":851},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhere-does-an-llm-get-its-data-rag-data-sources-in-python",{"image":852,"title":853,"description":854},{"url":299},"Where Does an LLM Get Its Data? RAG Data Sources in Python","Files, databases, APIs, search and vector indexes can all provide runtime knowledge. The retrieval method should match the source and the question.",{},{"id":411,"data":857,"type":42,"tunes":859},{"text":858,"level":234},"The Six Authority Signals",{},{"id":416,"data":861,"type":42,"tunes":863},{"text":862,"level":264},"1. Source Identity",{},{"id":421,"data":865,"type":240,"tunes":867},{"text":866},"The system should know what the source actually is: policy, contract, API, database record, research paper, documentation page, user file or third-party commentary.",{},{"id":426,"data":869,"type":42,"tunes":871},{"text":870,"level":264},"2. Provenance",{},{"id":431,"data":873,"type":240,"tunes":875},{"text":874},"Evidence should preserve where it came from, how it entered the system and which original object or document can be inspected if the claim is challenged.",{},{"id":436,"data":877,"type":42,"tunes":879},{"text":878,"level":264},"3. Freshness",{},{"id":441,"data":881,"type":240,"tunes":883},{"text":882},"Freshness matters when the underlying fact changes. A current API response and a five-year-old article may be equally relevant to the words of a query but not to its temporal requirement.",{},{"id":446,"data":885,"type":42,"tunes":887},{"text":886,"level":264},"4. Applicability",{},{"id":451,"data":889,"type":240,"tunes":891},{"text":890},"A fact may be correct but applicable to the wrong product version, country, tenant, customer group, contract, software release or time period.",{},{"id":456,"data":893,"type":42,"tunes":895},{"text":894,"level":264},"5. Evidence Specificity",{},{"id":461,"data":897,"type":240,"tunes":899},{"text":898},"Primary evidence that directly supports the claim is generally stronger than a chain of summaries that merely repeat it.",{},{"id":466,"data":901,"type":42,"tunes":903},{"text":902,"level":264},"6. Access Legitimacy",{},{"id":471,"data":905,"type":240,"tunes":907},{"text":906},"A technically retrievable document is not necessarily authorized context. Retrieval authority must operate inside tenant, permission, confidentiality and data-minimization boundaries.",{},{"id":476,"data":909,"type":302,"tunes":915},{"link":910,"meta":911},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhat-is-context-engineering-what-the-model-receives-before-it-answers",{"image":912,"title":913,"description":914},{"url":299},"What Is Context Engineering? What the Model Receives Before It Answers","Context assembly is also a validity and security boundary. Retrieved information must preserve scope, authority and access constraints.",{},{"id":485,"data":917,"type":42,"tunes":919},{"text":918,"level":234},"A Minimal Authority Model",{},{"id":490,"data":921,"type":240,"tunes":923},{"text":922},"A practical system can represent these properties explicitly rather than hoping that similarity ranking captures them.",{},{"id":495,"data":925,"type":291,"tunes":926},{"code":497},{},{"id":500,"data":928,"type":240,"tunes":930},{"text":929},"The retriever can still use semantic or lexical relevance. The authority layer adds constraints and reranking signals that relevance alone cannot provide.",{},{"id":505,"data":932,"type":42,"tunes":934},{"text":933,"level":234},"Connection to Agent-Ready Information",{},{"id":510,"data":936,"type":240,"tunes":938},{"text":937},"This becomes especially important when the consumer is not a human reading documents but an AI agent making decisions.",{},{"id":515,"data":940,"type":240,"tunes":942},{"text":941},"An agent-ready information object should preserve identity, provenance, conflicts, unknowns and validity conditions before the retrieval stage. Otherwise the system tries to reconstruct authority after information has already been flattened into anonymous text.",{},{"id":520,"data":944,"type":302,"tunes":950},{"link":945,"meta":946},"https:\u002F\u002Fstajic.de\u002Fportfolio\u002Ffrom-raw-data-to-verifiable-ai-context-a-practical-architecture-for-agent-ready-information",{"image":947,"title":948,"description":949},{"url":299},"From Raw Data to Verifiable AI Context: A Practical Architecture for Agent-Ready Information","A portfolio architecture for turning heterogeneous data into controlled information objects with identity, provenance, conflicts and validity.",{},{"id":529,"data":952,"type":42,"tunes":954},{"text":953,"level":234},"Diagnostic \u002F Decision Method",{},{"id":534,"data":956,"type":240,"tunes":958},{"text":957},"For every high-value retrieved claim, ask six questions:",{},{"id":539,"data":960,"type":551,"tunes":969},{"meta":961,"items":962,"style":550},{"counterType":542},[963,964,965,966,967,968],"What exact source produced this evidence?","Is this source authoritative for this type of claim?","Is the information still current?","Does it apply to this user, version, jurisdiction and time period?","Can the claim be traced back to the original evidence?","Was the user or agent authorized to retrieve it?",{},{"id":554,"data":971,"type":240,"tunes":973},{"text":972},"If the system cannot answer these questions, the problem is no longer only retrieval quality. It is an evidence-governance problem.",{},{"id":559,"data":975,"type":42,"tunes":977},{"text":976,"level":234},"Common Failure Modes",{},{"id":564,"data":979,"type":240,"tunes":981},{"text":980},"The highest similarity score wins. This confuses semantic closeness with evidential authority.",{},{"id":569,"data":983,"type":240,"tunes":985},{"text":984},"The newest source always wins. A newer commentary may be less authoritative than an older governing contract or standard.",{},{"id":574,"data":987,"type":240,"tunes":989},{"text":988},"Official means universally applicable. An official source may still belong to another jurisdiction, version or scope.",{},{"id":579,"data":991,"type":240,"tunes":993},{"text":992},"Public means safe to retrieve. Information can be public yet inappropriate for the task, while private information can be legitimate only inside an authorized context.",{},{"id":584,"data":995,"type":42,"tunes":997},{"text":996,"level":234},"Edge Case: Conflicting Authoritative Sources",{},{"id":589,"data":999,"type":240,"tunes":1001},{"text":1000},"The difficult case is not an authoritative source versus an obviously weak source. It is two apparently authoritative sources that disagree.",{},{"id":594,"data":1003,"type":240,"tunes":1005},{"text":1004},"At that point, the authority layer should preserve the conflict instead of silently collapsing it into one answer.",{},{"id":599,"data":1007,"type":291,"tunes":1008},{"code":601},{},{"id":604,"data":1010,"type":42,"tunes":1012},{"text":1011,"level":234},"Limitations",{},{"id":609,"data":1014,"type":240,"tunes":1016},{"text":1015},"The Retrieval Authority Layer is a conceptual architecture, not a universal ranking algorithm.",{},{"id":614,"data":1018,"type":240,"tunes":1020},{"text":1019},"Different domains require different authority models. Legal, scientific, commercial, operational and personal data cannot all use the same source hierarchy.",{},{"id":619,"data":1022,"type":240,"tunes":1024},{"text":1023},"It also cannot guarantee that an external search or AI system will rank a publisher's page. External ranking systems remain outside the publisher's control.",{},{"id":624,"data":1026,"type":42,"tunes":1028},{"text":1027,"level":234},"What Would Change This Answer?",{},{"id":629,"data":1030,"type":240,"tunes":1032},{"text":1031},"Retrieval systems may increasingly learn source reliability automatically, combine structured provenance with neural ranking and negotiate access through agent protocols.",{},{"id":634,"data":1034,"type":240,"tunes":1036},{"text":1035},"That may change how authority signals are implemented. It does not remove the distinction between finding relevant information and deciding whether that information deserves evidential weight.",{},{"id":639,"data":1038,"type":42,"tunes":1040},{"text":1039,"level":234},"Conclusion",{},{"id":644,"data":1042,"type":240,"tunes":1044},{"text":1043},"The Retrieval Trigger answers: Do I need external evidence?",{},{"id":649,"data":1046,"type":240,"tunes":1048},{"text":1047},"The Retrieval Authority Layer answers the next question: Which external source deserves to become evidence?",{},{"id":654,"data":1050,"type":291,"tunes":1051},{"code":656},{},{"id":659,"data":1053,"type":240,"tunes":1055},{"text":1054},"That distinction moves RAG from document retrieval toward controlled evidence retrieval.",{},{"id":664,"data":1057,"type":42,"tunes":1059},{"text":1058,"level":234},"Primary Sources",{},{"id":669,"data":1061,"type":240,"tunes":1063},{"text":1062},"Patrick Lewis et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.11401\" target=\"_blank\">Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks\u003C\u002Fa> (2020).",{},{"id":674,"data":1065,"type":240,"tunes":1067},{"text":1066},"Akari Asai et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.11511\" target=\"_blank\">Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection\u003C\u002Fa> (2023).",{},{"id":679,"data":1069,"type":240,"tunes":1071},{"text":1070},"Soyeong Jeong et al., \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14403\" target=\"_blank\">Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity\u003C\u002Fa> (2024).",{},"2.31.6","Finding a relevant document is not enough. AI systems also need to distinguish sources that are current, authoritative, applicable and traceable from sources that merely look similar. The Retrieval Authority Layer defines that source-selection boundary.",{"lang":7,"title":224,"content":226,"contentJson":1075,"excerpt":684},{"time":228,"blocks":1076,"version":683},[1077,1080,1083,1086,1089,1092,1095,1098,1101,1104,1107,1110,1115,1118,1121,1128,1131,1134,1137,1140,1143,1146,1149,1152,1155,1158,1161,1164,1167,1170,1175,1178,1181,1184,1187,1190,1193,1196,1199,1202,1205,1208,1211,1214,1219,1222,1225,1228,1231,1234,1237,1240,1245,1248,1251,1256,1259,1262,1265,1268,1271,1274,1277,1280,1283,1286,1289,1292,1295,1298,1301,1304,1307,1310,1313,1316,1319,1322,1325,1328,1331],{"id":231,"data":1078,"type":42,"tunes":1079},{"text":233,"level":234},{},{"id":237,"data":1081,"type":240,"tunes":1082},{"text":239},{},{"id":243,"data":1084,"type":240,"tunes":1085},{"text":245},{},{"id":248,"data":1087,"type":240,"tunes":1088},{"text":250},{},{"id":253,"data":1090,"type":258,"tunes":1091},{"text":255,"caption":256,"alignment":257},{},{"id":261,"data":1093,"type":265,"tunes":1094},{"title":263,"maxLevel":264,"minLevel":234},{},{"id":268,"data":1096,"type":42,"tunes":1097},{"text":270,"level":234},{},{"id":273,"data":1099,"type":240,"tunes":1100},{"text":275},{},{"id":278,"data":1102,"type":240,"tunes":1103},{"text":280},{},{"id":283,"data":1105,"type":240,"tunes":1106},{"text":285},{},{"id":288,"data":1108,"type":291,"tunes":1109},{"code":290},{},{"id":294,"data":1111,"type":302,"tunes":1114},{"link":296,"meta":1112},{"image":1113,"title":300,"description":301},{"url":299},{},{"id":305,"data":1116,"type":42,"tunes":1117},{"text":307,"level":234},{},{"id":310,"data":1119,"type":240,"tunes":1120},{"text":312},{},{"id":315,"data":1122,"type":329,"tunes":1127},{"content":1123,"stretched":43,"withHeadings":14},[1124,1125,1126],[319,320,321],[323,324,325],[327,328,325],{},{"id":332,"data":1129,"type":240,"tunes":1130},{"text":334},{},{"id":337,"data":1132,"type":291,"tunes":1133},{"code":339},{},{"id":342,"data":1135,"type":240,"tunes":1136},{"text":344},{},{"id":347,"data":1138,"type":42,"tunes":1139},{"text":349,"level":234},{},{"id":352,"data":1141,"type":240,"tunes":1142},{"text":354},{},{"id":357,"data":1144,"type":240,"tunes":1145},{"text":359},{},{"id":362,"data":1147,"type":240,"tunes":1148},{"text":364},{},{"id":367,"data":1150,"type":42,"tunes":1151},{"text":369,"level":234},{},{"id":372,"data":1153,"type":240,"tunes":1154},{"text":374},{},{"id":377,"data":1156,"type":291,"tunes":1157},{"code":379},{},{"id":382,"data":1159,"type":240,"tunes":1160},{"text":384},{},{"id":387,"data":1162,"type":42,"tunes":1163},{"text":389,"level":234},{},{"id":392,"data":1165,"type":240,"tunes":1166},{"text":394},{},{"id":397,"data":1168,"type":240,"tunes":1169},{"text":399},{},{"id":402,"data":1171,"type":302,"tunes":1174},{"link":404,"meta":1172},{"image":1173,"title":407,"description":408},{"url":299},{},{"id":411,"data":1176,"type":42,"tunes":1177},{"text":413,"level":234},{},{"id":416,"data":1179,"type":42,"tunes":1180},{"text":418,"level":264},{},{"id":421,"data":1182,"type":240,"tunes":1183},{"text":423},{},{"id":426,"data":1185,"type":42,"tunes":1186},{"text":428,"level":264},{},{"id":431,"data":1188,"type":240,"tunes":1189},{"text":433},{},{"id":436,"data":1191,"type":42,"tunes":1192},{"text":438,"level":264},{},{"id":441,"data":1194,"type":240,"tunes":1195},{"text":443},{},{"id":446,"data":1197,"type":42,"tunes":1198},{"text":448,"level":264},{},{"id":451,"data":1200,"type":240,"tunes":1201},{"text":453},{},{"id":456,"data":1203,"type":42,"tunes":1204},{"text":458,"level":264},{},{"id":461,"data":1206,"type":240,"tunes":1207},{"text":463},{},{"id":466,"data":1209,"type":42,"tunes":1210},{"text":468,"level":264},{},{"id":471,"data":1212,"type":240,"tunes":1213},{"text":473},{},{"id":476,"data":1215,"type":302,"tunes":1218},{"link":478,"meta":1216},{"image":1217,"title":481,"description":482},{"url":299},{},{"id":485,"data":1220,"type":42,"tunes":1221},{"text":487,"level":234},{},{"id":490,"data":1223,"type":240,"tunes":1224},{"text":492},{},{"id":495,"data":1226,"type":291,"tunes":1227},{"code":497},{},{"id":500,"data":1229,"type":240,"tunes":1230},{"text":502},{},{"id":505,"data":1232,"type":42,"tunes":1233},{"text":507,"level":234},{},{"id":510,"data":1235,"type":240,"tunes":1236},{"text":512},{},{"id":515,"data":1238,"type":240,"tunes":1239},{"text":517},{},{"id":520,"data":1241,"type":302,"tunes":1244},{"link":522,"meta":1242},{"image":1243,"title":525,"description":526},{"url":299},{},{"id":529,"data":1246,"type":42,"tunes":1247},{"text":531,"level":234},{},{"id":534,"data":1249,"type":240,"tunes":1250},{"text":536},{},{"id":539,"data":1252,"type":551,"tunes":1255},{"meta":1253,"items":1254,"style":550},{"counterType":542},[544,545,546,547,548,549],{},{"id":554,"data":1257,"type":240,"tunes":1258},{"text":556},{},{"id":559,"data":1260,"type":42,"tunes":1261},{"text":561,"level":234},{},{"id":564,"data":1263,"type":240,"tunes":1264},{"text":566},{},{"id":569,"data":1266,"type":240,"tunes":1267},{"text":571},{},{"id":574,"data":1269,"type":240,"tunes":1270},{"text":576},{},{"id":579,"data":1272,"type":240,"tunes":1273},{"text":581},{},{"id":584,"data":1275,"type":42,"tunes":1276},{"text":586,"level":234},{},{"id":589,"data":1278,"type":240,"tunes":1279},{"text":591},{},{"id":594,"data":1281,"type":240,"tunes":1282},{"text":596},{},{"id":599,"data":1284,"type":291,"tunes":1285},{"code":601},{},{"id":604,"data":1287,"type":42,"tunes":1288},{"text":606,"level":234},{},{"id":609,"data":1290,"type":240,"tunes":1291},{"text":611},{},{"id":614,"data":1293,"type":240,"tunes":1294},{"text":616},{},{"id":619,"data":1296,"type":240,"tunes":1297},{"text":621},{},{"id":624,"data":1299,"type":42,"tunes":1300},{"text":626,"level":234},{},{"id":629,"data":1302,"type":240,"tunes":1303},{"text":631},{},{"id":634,"data":1305,"type":240,"tunes":1306},{"text":636},{},{"id":639,"data":1308,"type":42,"tunes":1309},{"text":641,"level":234},{},{"id":644,"data":1311,"type":240,"tunes":1312},{"text":646},{},{"id":649,"data":1314,"type":240,"tunes":1315},{"text":651},{},{"id":654,"data":1317,"type":291,"tunes":1318},{"code":656},{},{"id":659,"data":1320,"type":240,"tunes":1321},{"text":661},{},{"id":664,"data":1323,"type":42,"tunes":1324},{"text":666,"level":234},{},{"id":669,"data":1326,"type":240,"tunes":1327},{"text":671},{},{"id":674,"data":1329,"type":240,"tunes":1330},{"text":676},{},{"id":679,"data":1332,"type":240,"tunes":1333},{"text":681},{},"Post erfolgreich abgerufen",{"items":1336,"source":1357,"manualIds":1358,"manualMatchedIds":1359},[1337,1344,1351],{"id":1338,"slug":1339,"title":1340,"excerpt":1341,"featuredImage":1342,"publishedAt":1343},"384","new-qwen-3-5-plus","全新Qwen 3.5-Plus：开源AI迈入新纪元","探索阿里巴巴Qwen 3.5-Plus的革命性特性与优势，这款为开发者打造的颠覆性开源人工智能模型。","\u002Fuploads\u002F2026\u002F02\u002Fnew-qwen-3-5-plus-1771515512741-dcbi9p.webp","2026-02-19T10:23:00.000Z",{"id":1345,"slug":1346,"title":1347,"excerpt":1348,"featuredImage":1349,"publishedAt":1350},"486","source-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from","AI系统中的真相来源：可靠知识究竟从何而来","事实来源（Source of Truth）定义了对于特定事实或状态，哪个来源具有权威性。了解它与RAG、溯源、记忆、上下文、向量数据库和记录系统有何不同。","\u002Fuploads\u002F2026\u002F10\u002Fsource-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from-1791479103235-6bq9em.webp","2026-10-08T13:02:00.000Z",{"id":1352,"slug":1353,"title":1354,"excerpt":1355,"featuredImage":1342,"publishedAt":1356},"445","qwen-3-6-in-production-release-runbook-ai-rollback-and-llmops-versioning","Qwen 3.6 生产环境部署：发布手册、AI 回滚与 LLMOps 版本管理","Qwen 3.6 不仅仅是一次模型升级。它同时是一个发布事件、一个回滚场景和一个版本管理问题。本文通过LLMOps规范、提示词与模型可追溯性、受控发布以及基于证据的回滚准备，阐述了在生产环境中应如何处理Qwen 3.6。","2026-05-04T02:49:00.000Z","fallback",[],[]]