[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:stajic:zh":3,"public-menus:all":38,"post:what-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs:zh":205,"related:post:what-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs:zh:1":2149},{"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,[148,161,175,181,193],{"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-folder","custom",[],{"id":176,"title":177,"url":179,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":180},"item-28",{"de":178,"en":178,"es":178,"fr":178,"it":178,"ru":178,"sr":178,"zh":178},"Solr Suggester","\u002Fportfolio\u002Fsolr-fuzzy-suggester-und-solr-infix-suggester-abfrage-ueber-ajax-und-filterung",[],{"id":182,"title":183,"url":191,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":192},"item-27",{"de":184,"en":185,"es":186,"fr":187,"it":188,"ru":189,"sr":190,"zh":185},"Firmenwebseite SEO","Company Website SEO","Sitio web corporativo SEO","Site web d’entreprise SEO","Sito web aziendale SEO","Корпоративный сайт SEO","Пословна веб-страница SEO","\u002Fportfolio\u002Fseo-sem-branding-mobile-webseite-muenchen",[],{"id":194,"title":195,"url":203,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":204},"item-31",{"de":196,"en":197,"es":198,"fr":199,"it":200,"ru":201,"sr":202,"zh":197},"Digitalisierungsportal","Digitalization Portal","Portal de digitalización","Portail de numérisation","Portale di digitalizzazione","Портал цифровизации","Портал за дигитализацију","\u002Fportfolio\u002Fdigitalisierungsportal-archiv-museum-bibliothek-ead-lido-mets-mods",[],{"statusCode":4,"data":206,"message":2148},{"id":207,"title":208,"slug":209,"content":210,"contentJson":211,"excerpt":1070,"featuredImage":1071,"featuredImageAlt":1072,"featuredImageCaption":10,"featuredImageTitle":10,"featuredImageCopyright":10,"featuredImageAuthor":10,"featuredImageSourceUrl":10,"featuredImageLicense":10,"featuredImageIsAiGenerated":43,"status":1073,"publishedAt":1074,"createdAt":1075,"updatedAt":1076,"seoLocalePaths":1077,"categories":1086,"author":1103,"translations":1108},"483","什么是AI解决方案架构师？系统边界、职责与权衡","what-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u003Cp>\u003Cstrong>AI 解决方案架构师\u003C\u002Fstrong>将业务或产品需求转化为具体 AI 赋能解决方案的架构。该角色定义系统边界，以及在应用逻辑、权威数据、检索与上下文、模型与提供商、工具或智能体、身份与权限、安全、运行时与部署、可观测性、评估、成本与运营行为方面的重大选择。这不仅仅是模型选择或提示工程：架构职责是使整个解决方案可实施、可治理、可测试且可运营。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">直接回答\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>AI 解决方案架构师设计完整的 AI 赋能解决方案，而不仅仅是 AI 模型。\u003C\u002Fstrong>该角色将需求和非功能性需求与架构决策连接起来，组合必要的应用\u002F数据\u002F模型\u002F工具\u002F运行时层，明确信任和故障边界，并定义已实施系统将如何被验证和运营。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">术语说明\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>AI 解决方案架构师是一个实用的角色标签，并非普遍标准化的职位名称。\u003C\u002Fstrong>ISO\u002FIEC\u002FIEEE 42010:2022 标准化了架构描述的概念；它并未定义这一职位角色。组织可以将这些职责分配给多个人。在本文中，该术语指针对一个具体 AI 赋能解决方案或工作负载的架构职责。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">当前来源说明 — 2026 年 10 月 8 日\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">这里的架构原则有意保持供应商中立，同时将当前供应商指南用作实施证据。NIST AI RMF 1.0 目前正在修订中；NIST AI 600-1 仍是已发布的生成式 AI 配置文件。下文引用的 Microsoft 和 AWS 指南反映了当前生产关注点，例如身份、数据边界、模型抽象、安全、可观测性、评估、可靠性和成本。\u003C\u002Fdiv>\u003C\u002Faside>\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-6\" class=\"editorjs-toc__link\">AI 解决方案架构师实际架构什么？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-10\" class=\"editorjs-toc__link\">最简单的例子\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-14\" class=\"editorjs-toc__link\">简单示例止步之处\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-17\" class=\"editorjs-toc__link\">架构职责图\u003C\u002Fa>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-1\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-20\" class=\"editorjs-toc__link\">1. 将产品需求转化为架构需求\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-23\" class=\"editorjs-toc__link\">2. 设计权威数据、检索和上下文\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-26\" class=\"editorjs-toc__link\">3. 将模型和提供商视为依赖项，而非整个系统\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-29\" class=\"editorjs-toc__link\">4. 架构工具、行动和代理边界\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-32\" class=\"editorjs-toc__link\">5. 明确信任边界和权限\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-35\" class=\"editorjs-toc__link\">6. 决定系统实际运行在哪里\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-38\" class=\"editorjs-toc__link\">7. 定义评估、可观测性和运营验收\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-41\" class=\"editorjs-toc__link\">该角色应产出什么？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-44\" class=\"editorjs-toc__link\">工作主要是权衡，而不是“最佳实践”选择\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-47\" class=\"editorjs-toc__link\">这与相邻角色有何不同？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-51\" class=\"editorjs-toc__link\">实现证据：这些边界如何出现在我自己的工作中\u003C\u002Fa>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-1\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-53\" class=\"editorjs-toc__link\">SenseFlow：需求 → 要求 → 架构 → 验证\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-57\" class=\"editorjs-toc__link\">Aaasaasa AI 客户端：在集成之前分离概念\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-61\" class=\"editorjs-toc__link\">当前架构框架如何支持这一更广泛的范围\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-65\" class=\"editorjs-toc__link\">常见误解\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-67\" class=\"editorjs-toc__link\">AI 解决方案架构师应预防的失败模式\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-69\" class=\"editorjs-toc__link\">实用的决策顺序\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-71\" class=\"editorjs-toc__link\">边缘情况和角色限制\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-75\" 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-78\" class=\"editorjs-toc__link\">AI 解决方案架构师检查清单\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-80\" class=\"editorjs-toc__link\">结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-85\" class=\"editorjs-toc__link\">相关规范知识\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-88\" class=\"editorjs-toc__link\">主要来源和当前架构指南\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-6\">AI 解决方案架构师实际架构什么？\u003C\u002Fh2>\n\u003Cp>工作的对象是\u003Cstrong>解决方案\u003C\u002Fstrong>：将需求转化为有用、受控行为的完整社会技术系统。模型可能是该系统的核心，但它仍然只是一个依赖项。同一个模型可以参与安全的内部搜索助手、不安全的过度授权智能体、低延迟客户功能，或无法经济运营的高成本原型。架构决定了这些差异。\u003C\u002Fp>\n\u003Cp>因此，一个有用的边界是：\u003Cstrong>业务成果 → 需求 → 系统职责 → 架构决策 → 实施 → 验证 → 运营\u003C\u002Fstrong>。AI 解决方案架构师在这条链上工作，同时与产品、工程、数据、安全、基础设施、治理和领域专家协作。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">解决方案比模型更广泛\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">以模型为中心的问题\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">解决方案架构问题\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">能力\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Which model can generate or reason well enough?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Which combination of model, data, application logic, retrieval, tools and controls produces the required behavior?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">数据\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">What context can fit in the prompt?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">What is authoritative, who may access it, how is it retrieved, versioned, filtered and cited?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">安全\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Does the provider offer security features?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">What are the trust boundaries, identities, permissions, secrets, data flows and failure containment mechanisms?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">运营\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">What is the token latency?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">How is the complete workload deployed, observed, evaluated, recovered, versioned and cost-controlled?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">变更\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Can we switch models?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Which dependencies are abstracted, what changes require an ADR, and how do we validate that a replacement still meets requirements?\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-10\">最简单的例子\u003C\u002Fh2>\n\u003Cp>假设一家公司想要一个内部助手，根据维护手册和操作程序回答技术人员的问题。可见功能听起来很简单：输入问题并获得带来源的答案。\u003C\u002Fp>\n\u003Cp>架构问题要大得多。哪些文档是权威的？用户如何认证？检索是否必须遵守部门或站点权限？答案是否只允许使用检索到的证据？对于该数据分类，哪个模型可接受？云提供商是否可以接收内容？当检索没有找到任何内容时会发生什么？引用如何生成？答案质量如何评估？可接受的延迟和成本是多少？谁可以查看日志，日志中可以存储什么？\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">从需求到可运营的 AI 解决方案\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 定义成果\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">明确用户、业务价值、任务边界，以及成功的答案或行动意味着什么。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 捕获需求\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">明确功能需求、非功能性需求、约束、数据规则、风险容忍度和验收标准。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 建立边界\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">识别用户、身份、应用、权威数据、模型\u002F提供商依赖、工具、外部系统和信任区域。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 设计架构\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">选择数据\u002F检索、模型、编排、工具、权限、运行时、部署、回退和可观测性模式。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 记录重大决策\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">保留架构选择、替代方案、权衡和后果，以便后续变更仍然可理解。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. 实施与集成\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">将架构转化为应用代码、API、策略、基础设施、工作流和运营控制。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">7\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">7. 验证与运营\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">测试质量、安全、可靠性、成本和用户成果；监控真实工作负载并将证据反馈到决策中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-14\">简单示例止步之处\u003C\u002Fh2>\n\u003Cp>概念验证通常可以跳过生产环境无法跳过的架构。开发者可能硬编码一个提供商、使用共享 API 密钥、将所有文档放在一个索引中、在没有用户上下文过滤的情况下运行检索、逐字记录提示，并手动判断质量。这可以证明可行性，但并不能建立生产架构。\u003C\u002Fp>\n\u003Cp>生产环境引入了相互作用的约束：租户或用户隔离、隐私、数据驻留、吞吐量、延迟、成本、提供商配额、回退行为、可审计性、模型版本变更、检索质量、工具权限、事件响应和部署生命周期。架构师的工作不是同时最大化每一项质量；而是明确权衡，并设计一个满足实际优先级集合的解决方案。\u003C\u002Fp>\n\u003Ch2 id=\"section-17\">架构职责图\u003C\u002Fh2>\n\u003Cp>确切的职责划分因组织而异，但以下图谱概括了解决方案级 AI 架构中反复出现的职责。架构师可能不会亲自实施每一层；其职责是使各层协调一致地契合，并保持关键决策可追溯。\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\">架构领域\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">AI解决方案架构师必须解决的问题\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">典型输出\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">成果与范围\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">用户是谁？范围内包含什么任务？系统不得做什么？什么构成成功？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">解决方案上下文、能力边界、验收标准\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">需求与非功能需求\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">适用哪些质量、安全、可用性、延迟、成本、驻留和合规约束？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">需求映射、非功能需求、约束、验证标准\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">应用与编排\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">确定性应用逻辑在哪里结束，AI行为从哪里开始？工作流如何协调？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">组件模型、API、编排边界、故障路径\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">权威数据与检索\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">什么是事实来源？数据如何被摄取、授权、检索、过滤、排序和引用？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">数据流、检索架构、元数据和授权规则\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型与提供商层\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">需要哪些能力？哪些提供商\u002F运行时约束重要？应抽象什么？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型\u002F提供商决策、路由\u002F回退策略、抽象边界\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">工具与代理\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">系统可以采取哪些行动？哪些行动需要批准？工具身份和权限如何强制执行？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">工具契约、代理边界、批准和最小权限规则\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">身份与安全\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">存在哪些人类和机器身份？密钥保存在哪里？跨越了哪些信任边界？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">威胁\u002F信任边界模型、身份传播、密钥和授权设计\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运行时与部署\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">组件在哪里执行？什么是本地、云、边缘或混合？存在哪些网络和可用性假设？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部署视图、运行时拓扑、环境和连接决策\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">评估与可观测性\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">发布前后如何衡量质量？需要哪些追踪、指标、日志和证据？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">评估计划、遥测、审计追踪、发布门禁\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运营与变更\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型\u002F提示\u002F配置\u002F数据版本如何变更、回滚和支持？\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运营模型、生命周期控制、ADR、运行手册、变更规则\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch3 id=\"section-20\">1. 将产品需求转化为架构需求\u003C\u002Fh3>\n\u003Cp>AI架构在模型选择之前就开始了。架构师首先确定解决方案预期实现什么以及在哪些约束下实现。这包括功能行为，也包括缩小设计空间的非功能需求和策略：安全、可靠性、延迟、隐私、驻留、可维护性、成本和运营支持。\u003C\u002Fp>\n\u003Cp>这就是A02区分重要的地方：诸如“未经授权的用户不得检索受限文档”这样的需求不是架构决策。它是一个驱动因素。关于身份传播、索引分区、元数据过滤、API边界和授权执行的决策是架构响应，之后必须进行验证。\u003C\u002Fp>\n\u003Ch3 id=\"section-23\">2. 设计权威数据、检索和上下文\u003C\u002Fh3>\n\u003Cp>AI系统常常在模型行为与企业事实之间的边界上失败。架构师必须定义哪些来源是权威的，新鲜度和来源意味着什么，访问控制如何到达检索，以及检索到的证据如何成为模型上下文。向量数据库、嵌入模型或RAG库本身并不是架构。\u003C\u002Fp>\n\u003Cp>微软当前的AI工作负载指南明确表达了同样的分离：应用代码不应绕过数据访问边界；用户或租户上下文应传播到检索和过滤中；基础数据必须为可搜索性而设计，同时仍满足安全和合规要求。\u003C\u002Fp>\n\u003Ch3 id=\"section-26\">3. 将模型和提供商视为依赖项，而非整个系统\u003C\u002Fh3>\n\u003Cp>模型选择很重要，但它应由所需能力和约束驱动。架构师考虑推理或生成质量、模态、上下文限制、延迟、数据处理、部署位置、提供商可用性、成本、可观测性和替换风险。\u003C\u002Fp>\n\u003Cp>提供商抽象并不自动意味着“更好的架构”。它增加工程成本，并可能隐藏提供商特定的能力。当可移植性、回退、策略分离或多提供商路由是明确需求时，它才是合理的。否则，直接集成可能是更好的决策。关键在于让权衡是有意为之的。\u003C\u002Fp>\n\u003Ch3 id=\"section-29\">4. 架构工具、行动和代理边界\u003C\u002Fh3>\n\u003Cp>当AI系统可以调用工具、修改数据、发送消息、运行代码或操作业务系统时，架构风险就发生了变化。工具访问需要自己的身份和授权模型。模型请求行动的能力与执行该行动的权限并不相同。\u003C\u002Fp>\n\u003Cp>对于代理工作负载，当前AWS指南强调了额外维度，如代理身份、工具访问、编排、人工监督、追踪、故障处理和迭代推理循环的成本。即使框架隐藏了一些实现机制，这些也是解决方案关注点。\u003C\u002Fp>\n\u003Ch3 id=\"section-32\">5. 明确信任边界和权限\u003C\u002Fh3>\n\u003Cp>生产AI解决方案有多个信任边界：浏览器或客户端、应用后端、AI编排、检索\u002F数据服务、模型提供商、工具API、本地运行时和外部系统。每个边界都应回答：谁在调用，代表谁，使用什么凭证，针对哪个资源，有什么审计追踪，以及有什么故障遏制？\u003C\u002Fp>\n\u003Cp>安全不能推迟到模型周围的“护栏”。微软的AI工作负载指南明确将安全置于所有架构层，并要求身份\u002F访问管理、数据保护、内容控制和生命周期安全。NIST同样将治理和风险管理视为贯穿AI生命周期的持续活动。\u003C\u002Fp>\n\u003Ch3 id=\"section-35\">6. 决定系统实际运行在哪里\u003C\u002Fh3>\n\u003Cp>“本地AI”、“云AI”和“混合AI”只有在执行和数据路径精确时才是架构陈述。本地桌面进程仍然可以调用云模型。云托管应用可以从本地数据源检索。气隙解决方案具有完全不同的更新、模型分发和可观测性约束。\u003C\u002Fp>\n\u003Cp>因此，架构师将\u003Cstrong>运行时位置\u003C\u002Fstrong>、\u003Cstrong>推理位置\u003C\u002Fstrong>、\u003Cstrong>数据位置\u003C\u002Fstrong>和\u003Cstrong>控制平面\u003C\u002Fstrong>分开。将它们混为一谈会产生错误的安全性和部署假设。\u003C\u002Fp>\n\u003Ch3 id=\"section-38\">7. 定义评估、可观测性和运营验收\u003C\u002Fh3>\n\u003Cp>AI 行为具有部分非确定性，因此发布定义不能仅依赖传统的单元测试。架构需要可衡量的验收标准：任务成功率、在相关情况下的有据可依性或引用正确性、拒绝行为、工具安全性、延迟、成本、可靠性和安全测试。具体指标取决于用例。\u003C\u002Fp>\n\u003Cp>微软当前的 Well-Architected AI 指南将监控视为持续性的，并将其应用于模型行为、提示\u002F补全、异常、安全和生产质量门禁。AWS 同样将可观测性、生命周期管理以及模型\u002F提示可追溯性视为运营架构关注点。\u003C\u002Fp>\n\u003Ch2 id=\"section-41\">该角色应产出什么？\u003C\u002Fh2>\n\u003Cp>架构不是幻灯片。有用的产出是那些能让工程、安全、产品和运营做出一致决策，并在之后理解系统为何以当前形式存在的工件。\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\">工件\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">目的\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">解决方案上下文和边界\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">展示用户、外部系统、主要职责以及范围之外的内容\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">需求\u002F非功能需求映射\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">将产品需求和约束与架构工作及验证联系起来\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">组件和数据流视图\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">展示应用、数据\u002F检索、模型、工具、身份和运行时交互\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">信任和权限模型\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">明确身份、密钥、授权、敏感数据和高风险操作\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">架构决策记录\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">保留重要选择、替代方案、权衡、状态和后果\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">评估和验收计划\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">定义声称解决方案满足质量和安全期望所需的证据\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部署和运营视图\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">定义环境、运行时位置、可观测性、回滚、事件和生命周期职责\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">可追溯性链接\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">连接需求、决策、实现工作、测试和运营证据\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-44\">工作主要是权衡，而不是“最佳实践”选择\u003C\u002Fh2>\n\u003Cp>架构之所以存在，是因为理想的品质会相互冲突。成本较低的模型可能会降低质量。能力更强的模型可能会增加延迟或数据治理约束。激进的缓存可以提高成本和速度，但会使新鲜度复杂化。更自主的代理可以减少人力投入，但会增加影响范围和审计要求。\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\">决策\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">潜在收益\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">潜在成本\u002F风险\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">架构问题\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">托管云模型\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">快速采用、强大的托管能力\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">外部依赖、数据和成本约束\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">工作负载是否允许该提供商\u002F数据路径并满足弹性需求？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">本地\u002F自托管推理\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">控制、离线\u002F私有选项\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">硬件、运营、模型生命周期负担\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">控制收益是否值得承担运营责任？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">单一提供商集成\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">实现更简单、完整的提供商功能\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">更高的切换\u002F故障集中度\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">是否确实需要可移植性或回退？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">提供商抽象\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">可移植性、路由和策略分离\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">最低公分母风险、更多代码\u002F测试\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">哪些差异必须保持可见而不是被抽象掉？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">大上下文\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">每次请求更多信息\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">延迟、成本、注意力稀释、泄露面\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">是否应检索\u002F过滤数据，而不是始终注入？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">强大的工具\u002F自主性\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">更多端到端自动化\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">更高权限和故障影响范围\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">哪些操作需要最小权限、确认或人工批准？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">严格验证和日志记录\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">更好的证据和运营\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">延迟、存储、隐私和复杂性成本\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">此风险级别需要哪些证据？\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-47\">这与相邻角色有何不同？\u003C\u002Fh2>\n\u003Cp>各公司的职位名称高度重叠。有用的区别在于\u003Cstrong>架构职责范围\u003C\u002Fstrong>，而不是人力资源标签。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">相邻角色回答不同的主要问题\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">角色\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">主要架构关注点\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">AI 解决方案架构师\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">One concrete AI-enabled solution\u002Fworkload\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">How requirements, data, models, tools, security, runtime and operations fit together to deliver the target outcome\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">AI 平台架构师\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Reusable AI platform capabilities across many solutions\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Shared provider gateways, model access, identity, evaluation, retrieval services, observability, deployment patterns and developer experience\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">企业 AI 架构师\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Organization\u002Fportfolio-level target architecture\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Capability landscape, governance, integration principles, shared platforms, standards, sourcing and strategic constraints across domains\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">AI \u002F ML 工程师\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Implementation of AI\u002FML behavior and pipelines\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Models, data, inference, evaluation, application logic and engineering tasks within the architecture\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">安全架构师\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Security architecture across systems\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Threats, identity, authorization, data protection, controls, assurance and compliance boundaries\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">产品\u002F交付负责人\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Outcome, scope, prioritization and delivery system\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">Why\u002Fwhat to build, sequencing, stakeholders, milestones, acceptance and value realization\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Cp>在小型产品团队中，一个人可能覆盖其中多个范围。在大型企业中，它们可能是具有正式评审委员会的独立角色。当职位名称改变时，架构职责并不会消失。\u003C\u002Fp>\n\u003Ch2 id=\"section-51\">实现证据：这些边界如何出现在我自己的工作中\u003C\u002Fh2>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">实现证据，而非通用规则\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">以下示例是\u003Cstrong>原始实现\u002F项目证据\u003C\u002Fstrong>。它们展示了我在真实项目工作中如何分离产品需求、需求、架构、运行时、模型\u002F提供商、权限和验证。它们并不声称每个组织都必须使用相同结构，也不意味着客户采用或企业级部署。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch3 id=\"section-53\">SenseFlow：需求 → 要求 → 架构 → 验证\u003C\u002Fh3>\n\u003Cp>在 SenseFlow 项目 Source of Truth 中，技术明确从属于产品愿景。开发结构从问题和产品愿景出发，经过用户需求、价值、范围、史诗、故事和验收标准，进入架构、实现、验证和迭代。\u003C\u002Fp>\n\u003Cp>需求被设计为可从产品目标 → 能力 → 史诗 → 用户故事 → 验收标准 → 技术任务进行追溯。在可行的情况下，它们包括功能需求、非功能需求、依赖关系、风险、假设、验收标准和验证方法。重大决策保留决策、原因、备选方案、权衡、状态和日期\u002F版本。\u003C\u002Fp>\n\u003Cp>这是在选择特定 AI 框架或模型之前进行的架构工作：它保护产品意图与技术决策之间的联系，并使后续变更可审查而非隐式。\u003C\u002Fp>\n\u003Ch3 id=\"section-57\">Aaasaasa AI 客户端：在集成之前分离概念\u003C\u002Fh3>\n\u003Cp>Aaasaasa AI 客户端提供了一个更接近实现层面的示例。其 AI Hub 有意分离了\u003Cstrong>代理\u002F客户端\u003C\u002Fstrong>、\u003Cstrong>提供者\u003C\u002Fstrong>、\u003Cstrong>模型\u003C\u002Fstrong>、\u003Cstrong>连接\u002F运行时位置\u003C\u002Fstrong>、\u003Cstrong>权限\u003C\u002Fstrong>和\u003Cstrong>Web 客户端\u003C\u002Fstrong>。本地运行时并不假定意味着本地推理，权限被视为运行时\u002F工具策略，而非模型的属性。\u003C\u002Fp>\n\u003Cp>桌面架构还定义了信任边界：Nuxt 渲染器相对于 Electron 主进程是不可信的。一个狭窄的预加载脚本和经过验证的 IPC 调解对 AI 服务、设置、加密密钥、工作区\u002F数据服务和运行时的访问。云凭证保留在特权主进程中；渲染器代码接收规范化状态，而非原始密钥或无限制的操作系统访问。\u003C\u002Fp>\n\u003Cp>路由决策同样是架构性的。实现不会从本地路由静默回退到付费云推理；云路由需要明确确认。直接聊天默认没有文件系统或 shell 工具，而代理执行应用选定的工作区和权限配置文件。这些是关于信任、成本、执行和用户期望的解决方案级决策——而非模型特性。\u003C\u002Fp>\n\u003Ch2 id=\"section-61\">当前架构框架如何支持这一更广泛的范围\u003C\u002Fh2>\n\u003Cp>ISO\u002FIEC\u002FIEEE 42010:2022 为跨软件、系统和企业的架构描述提供了一般性规范。它有意比 AI 更广泛，并且不规定单一的架构方法或职位名称。这使得它在此处作为边界很有用：AI 解决方案架构仍然是架构，具有利益相关者关注点、多个视图和必须清晰表达的重要关系。\u003C\u002Fp>\n\u003Cp>NIST AI RMF 1.0 通过\u003Cstrong>治理、映射、测量和管理\u003C\u002Fstrong>来构建 AI 风险管理框架，并强调风险管理应在 AI 系统生命周期中持续进行。生成式 AI 配置文件（NIST AI 600-1）将该框架适配到 GAI 风险和组织优先级。这强化了架构不能止步于功能模型性能。\u003C\u002Fp>\n\u003Cp>微软当前的 Azure Well-Architected AI 指南分离了应用程序设计、应用平台、训练数据、基础数据和数据平台关注点，并反复将它们与可靠性、安全性、卓越运营、性能和成本联系起来。AWS 的生成式 AI 和代理式 AI 透镜同样将可观测性、安全性、可靠性、模型\u002F工具生命周期、成本和人工监督视为架构关注点。\u003C\u002Fp>\n\u003Ch2 id=\"section-65\">常见误解\u003C\u002Fh2>\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\">误解\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">纠正\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“架构师选择 LLM。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型选择是更大解决方案架构中的一个决策。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“提示工程就是架构。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">提示影响行为，但它们不定义身份、数据访问、信任边界、部署、工具权限或运营。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“RAG 解决企业知识。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">检索只是一个子系统；授权、来源、新鲜度、证据、索引、评估和源治理仍然需要设计。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“本地运行时意味着私有\u002F本地 AI。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运行时、推理、数据和控制平面位置是独立的架构属性。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“如果供应商提供护栏，安全性就覆盖了。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">安全性涵盖身份、授权、密钥、数据流、工具、日志记录、部署、人工审批和提供者边界。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“架构师必须编写每个组件。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">动手实现可以提高架构质量，但角色由集成决策责任定义，而非亲自编写每一层代码。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“架构图证明生产就绪。”\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">就绪需要跨质量、安全、运营和业务验收的实施控制和验证证据。\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-67\">AI 解决方案架构师应预防的失败模式\u003C\u002Fh2>\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\">失败模式\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">发生原因\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">架构纠正\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型优先设计\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">一个有前景的模型演示成为系统蓝图\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">从结果、约束和验证开始；在该框架内选择模型\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">生产环境中的原型权限\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">共享凭证和广泛访问在 PoC 后仍然存在\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">尽早定义身份传播、最小权限、工具范围和审批边界\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">未经授权的检索\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">搜索质量在设计数据访问规则之前就被设计\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">将用户\u002F租户上下文带入检索，并在数据访问边界强制执行授权\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">静默的提供者\u002F运行时假设\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">“本地”、“云”和“离线”使用不精确\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">分别记录运行时、推理、数据和控制平面位置\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">无失败契约\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">设计了快乐路径，但未设计拒绝\u002F回退\u002F错误行为\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">指定检索为空、模型不可用、工具失败和策略拒绝的行为\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">实现后评估\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">在发布前手动判断质量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">在架构冻结之前定义可衡量的验收和代表性评估集\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">不可追溯的变更\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型、提示、检索或权限变更没有架构历史\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">对关键配置进行版本控制，并记录重大决策\u002F验证证据\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运营仅视为基础设施\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部署后 AI 行为不可观测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">将跟踪、质量指标、安全事件、成本遥测和回滚一起设计\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-69\">实用的决策顺序\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">AI 解决方案架构决策顺序\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">结果\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">定义用户\u002F业务结果和明确的非目标。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">证据和约束\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">识别权威数据、策略、非功能需求、风险和验收条件。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">系统边界\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">映射用户、身份、应用程序、数据、模型\u002F提供者、工具和外部系统。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">架构选项\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">比较检索、模型访问、编排、部署、权限、评估和可观测性的模式。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">权衡决策\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">选择重要选项并保留理由、备选方案和后果。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">实现契约\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">将决策转化为 API、模式、权限规则、部署定义和工程任务。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">7\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">验证\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">根据原始功能和非功能需求测试已实现的系统。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">8\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">运营反馈\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">使用生产证据、事件、质量指标和成本\u002F安全信号触发受控变更。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-71\">边缘情况和角色限制\u003C\u002Fh2>\n\u003Cp>一些 AI 产品以模型训练、科学实验或专用硬件为主。在这些情况下，模型\u002F数据科学和 ML 系统架构可能比此处显示的解决方案级映射深入得多。AI 解决方案架构师仍然需要集成和运营边界，但专家架构可能拥有训练平台本身。\u003C\u002Fp>\n\u003Cp>在另一个极端，简单的 SaaS 集成可能并不需要专职架构师。一位高级工程师或技术产品负责人也能承担同样的架构职责。有用的检验标准不是头衔，而是是否在有意地做出并验证重大的跨层决策。\u003C\u002Fp>\n\u003Cp>受监管、主权、气隙隔离、安全关键、高度自治或多租户系统也会改变重心。身份、隔离、驻留、保证、更新机制、人工监督和可审计性在架构中可能比模型质量更为重要。\u003C\u002Fp>\n\u003Ch2 id=\"section-75\">什么会改变这个答案？\u003C\u002Fh2>\n\u003Cp>当架构从一个应用转向可复用平台或企业级目标架构时，确切的职责边界会发生变化。这就是为什么 \u003Cstrong>AI 平台架构师\u003C\u002Fstrong> 和 \u003Cstrong>企业 AI 架构\u003C\u002Fstrong> 值得单独进行规范处理，而不是并入这个角色。\u003C\u002Fp>\n\u003Cp>技术变化也很重要。新的模型能力、协议、本地运行时和托管服务可以消除一些实现工作，同时创造新的信任或运营边界。稳定的职责是将这些变化理解为系统变化——而不是把新框架当作架构的替代品。\u003C\u002Fp>\n\u003Ch2 id=\"section-78\">AI 解决方案架构师检查清单\u003C\u002Fh2>\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\">检查项\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">问题\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">成果\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">用户\u002F业务结果和非目标边界是否明确？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">需求\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">功能需求、非功能需求、约束和验收标准是否可追溯？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">数据\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">权威来源、来源追溯、时效性、保留和访问规则是否已定义？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">检索\u002F上下文\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">授权是否延伸到检索和上下文构建？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型\u002F提供商\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">模型\u002F提供商选择是否基于能力和约束，而非偏好？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">工具\u002F智能体\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">操作边界、权限、审批和失败行为是否明确？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">身份\u002F安全\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">人类\u002F机器身份、密钥和信任边界是否已定义？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运行时\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运行时、推理、数据和控制平面的位置是否已区分？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">评估\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">是否有可衡量的证据证明质量、安全性和验收？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">可观测性\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">能否调查生产行为、故障、成本和安全事件？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">变更\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">重大架构决策和替换是否可追溯？\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">运营\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部署、回滚、事件和生命周期的归属是否清晰？\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-80\">结论\u003C\u002Fh2>\n\u003Cp>AI 解决方案架构师是将 AI 机会转化为连贯技术系统的人或架构职能。关键技能不是知道最多的模型名称，而是连接产品需求、需求规格、数据、应用架构、AI 能力、安全、运行时、交付和验证，同时不丢失它们之间的边界。\u003C\u002Fp>\n\u003Cp>因此，一个强大的 AI 解决方案架构可以总结为：\u003Cstrong>定义目标 → 建立需求和约束 → 设计系统边界 → 明确重大权衡 → 通过清晰的契约实现 → 依据证据进行验证 → 有意识地运营和演进。\u003C\u002Fstrong> 模型很重要。解决方案才是产品。\u003C\u002Fp>\n\u003Csection class=\"editorjs-faq my-6 rounded-xl border border-gray-200 p-5 dark:border-gray-700\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">AI 解决方案架构师 — 常见问题\u003C\u002Fh3>\u003Cdiv id=\"faq1\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">什么是 AI 解决方案架构师？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">AI 解决方案架构师将业务或产品需求转化为具体 AI 赋能解决方案的架构，定义应用逻辑、数据\u002F检索、模型、工具、身份、安全、运行时、评估和运营如何协同工作。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq2\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 解决方案架构师和 AI 工程师是同一个角色吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">不是。这些角色可能重叠，尤其是在小团队中，但 AI 工程师主要是实现角色，而解决方案架构师负责或协调完整工作负载的跨层架构决策和权衡。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq3\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 解决方案架构师需要编码吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">从定义上讲不需要，但动手实现知识非常有价值，因为 AI 架构跨越 API、数据、检索、安全、运行时和运营行为。这个角色由架构职责定义，而不是由亲自编写每个组件定义。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq4\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">选择 LLM 是主要工作吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">不是。模型选择只是一个决策。生产架构还需要数据和检索边界、权限、工具、提供商\u002F运行时选择、可观测性、评估、可靠性、成本和生命周期设计。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq5\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 解决方案架构师和 AI 平台架构师有什么区别？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">AI 解决方案架构师专注于一个具体的解决方案或工作负载。AI 平台架构师专注于支持多个解决方案的可复用 AI 能力和护栏。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq6\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 解决方案架构师和企业 AI 架构师有什么区别？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">解决方案架构师在应用\u002F工作负载范围内工作。企业 AI 架构跨组织组合、目标架构、治理、共享能力、集成原则和战略约束工作。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq7\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">RAG 和智能体适合放在哪里？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">当需求证明其合理性时，它们是解决方案内部的架构模式或子系统。RAG 处理基于检索的上下文；智能体增加了规划\u002F工具执行，因此带来额外的身份、权限、编排和运营关注点。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq8\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">什么能证明架构有效？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">实现加上验证证据：功能测试、评估结果、安全\u002F授权测试、性能和可靠性测量、可观测性、运营演练以及针对原始需求的验收。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Csection class=\"editorjs-glossary my-6 rounded-xl border border-gray-200 dark:border-gray-700 p-5\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">核心术语\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"ai-solution-architect\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 解决方案架构师\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">对一个具体 AI 赋能解决方案或工作负载的架构职责，将产品需求与应用、数据、模型、工具、安全、运行时和运营设计相结合。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"system-boundary\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">系统边界\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">解决方案所属范围与其交互的用户、系统、提供商、数据源和环境之间的明确分隔。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"trust-boundary\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">信任边界\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">数据、身份或控制在不同信任假设的组件之间跨越的点，因此需要明确的安全控制。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"grounding\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">接地\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">向 AI 模型提供相关外部信息或证据，使其响应可以基于模型参数之外的来源。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"provider-abstraction\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">提供商抽象\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">将解决方案的部分与某个模型\u002F提供商接口解耦的应用边界。当路由、可移植性或策略需求证明其合理性时有用，但并非没有权衡。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"evaluation\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">评估\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">根据定义的验收标准对 AI 工作负载行为进行结构化测量，包括任务质量以及相关的安全、安保、性能和运营属性。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"ai-platform-architect\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">AI 平台架构师\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">专注于多个解决方案使用的可复用 AI 平台能力而非单个工作负载架构的架构角色。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"enterprise-ai-architecture\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">企业 AI 架构\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">组织级架构，协调整个组合中的 AI 能力、平台、治理、集成和战略约束。\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>\n\u003Ch2 id=\"section-85\">相关规范知识\u003C\u002Fh2>\n\u003Cp>本文属于 AI 架构基础集群。其直接基础是 \u003Cstrong>生成式 AI 解释：模型、检索、工具和应用不是同一回事\u003C\u002Fstrong> 和 \u003Cstrong>ADR 与 NFR：架构决策和系统质量不是同一回事\u003C\u002Fstrong>。相邻的规范节点包括 \u003Cstrong>智能体 AI 解释\u003C\u002Fstrong>、\u003Cstrong>AI 系统中的真相来源\u003C\u002Fstrong>、\u003Cstrong>向量数据库、嵌入和重排序\u003C\u002Fstrong>、\u003Cstrong>什么是上下文工程？\u003C\u002Fstrong>、\u003Cstrong>RBAC 与租户隔离\u003C\u002Fstrong>、\u003Cstrong>AI 平台架构师\u003C\u002Fstrong>、\u003Cstrong>企业 AI 架构\u003C\u002Fstrong> 和 \u003Cstrong>AI 治理\u003C\u002Fstrong>。在这些节点尚未发布的地方，URL 有意不进行虚构。\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works\" 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 RAG？对其工作原理的最简单解释\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">stajic.de 现有的关于检索增强生成的规范解释，对 AI 解决方案架构的检索\u002F接地部分有用。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ch2 id=\"section-88\">主要来源和当前架构指南\u003C\u002Fh2>\n\u003Cp>以下外部来源支持一般架构主张；SenseFlow 和 Aaasaasa AI Client 部分是明确的原创项目\u002F实现证据。当前状态参考已于 2026 年 10 月 8 日核对。NIST 指出 AI RMF 1.0 正在修订中，因此当后续版本发布时，对版本敏感的治理参考应重新核对。\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fwww.iso.org\u002Fstandard\u002F74393.html\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">ISO\u002FIEC\u002FIEEE 42010:2022 — 架构描述\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于架构描述的结构和表达的现行国际标准。它将架构与其描述区分开来，并且不规定单一的架构方法、工具或记录格式。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nist.gov\u002Fitl\u002Fai-risk-management-framework\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">NIST AI 风险管理框架\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">NIST 的 AI RMF 资源页面。截至 2026 年 10 月，该页面指出 AI RMF 1.0 正在修订中，并链接了生成式 AI 配置文件及相关资源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fairc.nist.gov\u002Fairmf-resources\u002Fairmf\u002F5-sec-core\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">NIST AI RMF 核心 — 治理、映射、测量、管理\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">NIST AIRC 对 AI RMF 1.0 核心的官方介绍，包括四大功能以及面向生命周期的风险管理框架。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nist.gov\u002Fpublications\u002Fartificial-intelligence-risk-management-framework-generative-artificial-intelligence\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">NIST AI 600-1 — 生成式 AI 配置文件\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">面向 AI RMF 1.0 的跨行业生成式 AI 配置文件，于 2024 年 7 月 26 日发布，并由 NIST 于 2026 年更新。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fget-started\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">Microsoft Azure 架构完善框架 — AI 工作负载\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">当前的工作负载级架构指南，涵盖 AI 应用程序设计、应用程序平台、训练数据、基础数据、数据平台以及生产就绪相关事项。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fapplication-design\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">Microsoft — AI 工作负载的应用程序设计\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于模型\u002F工具抽象、数据访问边界、身份传播、授权以及客户端、智能、知识和工具层分离的指南。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fdesign-principles\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">Microsoft — AI 工作负载的设计原则\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">当前 AI 工作负载设计原则，涵盖可靠性、安全性、成本、运营卓越和性能，包括身份和数据保护责任。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fmlops-genaiops\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">Microsoft — 面向 AI 工作负载的 MLOps 和 GenAIOps\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">生产生命周期指南，涵盖监控、质量门禁、模型\u002F提示行为、安全性和运营度量。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fgenerative-ai-lens\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">AWS 架构完善框架生成式 AI 透镜\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">面向生成式 AI 工作负载的 AWS 架构指南，涵盖运营卓越、安全性、可靠性、性能效率、成本优化和可持续性。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fagentic-ai-lens\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" 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\">AWS 架构完善框架代理式 AI 透镜\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">发布于 2026 年，涵盖代理式 AI 特有的架构关注点，包括身份、工具、编排、人工监督、可靠性、追踪和推理循环成本。\u003C\u002Fp>\u003C\u002Fa>",{"time":212,"blocks":213,"version":1069},1791477081691,[214,219,226,232,237,244,248,252,256,300,304,308,312,340,344,348,352,356,360,408,412,416,420,424,428,432,436,440,444,448,452,456,460,464,468,472,476,480,484,488,492,496,500,531,535,539,583,587,591,639,643,647,653,657,661,665,669,673,677,681,685,689,693,697,701,705,733,737,777,781,810,814,818,822,826,830,834,838,842,882,886,890,894,931,963,967,971,981,985,989,997,1005,1013,1021,1029,1037,1045,1053,1061],{"id":215,"data":216,"type":218},"intro",{"text":217},"\u003Cstrong>AI 解决方案架构师\u003C\u002Fstrong>将业务或产品需求转化为具体 AI 赋能解决方案的架构。该角色定义系统边界，以及在应用逻辑、权威数据、检索与上下文、模型与提供商、工具或智能体、身份与权限、安全、运行时与部署、可观测性、评估、成本与运营行为方面的重大选择。这不仅仅是模型选择或提示工程：架构职责是使整个解决方案可实施、可治理、可测试且可运营。","paragraph",{"id":220,"data":221,"type":225},"direct",{"body":222,"title":223,"variant":224},"\u003Cstrong>AI 解决方案架构师设计完整的 AI 赋能解决方案，而不仅仅是 AI 模型。\u003C\u002Fstrong>该角色将需求和非功能性需求与架构决策连接起来，组合必要的应用\u002F数据\u002F模型\u002F工具\u002F运行时层，明确信任和故障边界，并定义已实施系统将如何被验证和运营。","直接回答","info","callout",{"id":227,"data":228,"type":225},"role-note",{"body":229,"title":230,"variant":231},"\u003Cstrong>AI 解决方案架构师是一个实用的角色标签，并非普遍标准化的职位名称。\u003C\u002Fstrong>ISO\u002FIEC\u002FIEEE 42010:2022 标准化了架构描述的概念；它并未定义这一职位角色。组织可以将这些职责分配给多个人。在本文中，该术语指针对一个具体 AI 赋能解决方案或工作负载的架构职责。","术语说明","note",{"id":233,"data":234,"type":225},"version-note",{"body":235,"title":236,"variant":231},"这里的架构原则有意保持供应商中立，同时将当前供应商指南用作实施证据。NIST AI RMF 1.0 目前正在修订中；NIST AI 600-1 仍是已发布的生成式 AI 配置文件。下文引用的 Microsoft 和 AWS 指南反映了当前生产关注点，例如身份、数据边界、模型抽象、安全、可观测性、评估、可靠性和成本。","当前来源说明 — 2026 年 10 月 8 日",{"id":238,"data":239,"type":243},"toc",{"title":240,"maxLevel":241,"minLevel":242},"目录",3,2,"tableOfContents",{"id":245,"data":246,"type":42},"h-meaning",{"text":247,"level":242},"AI 解决方案架构师实际架构什么？",{"id":249,"data":250,"type":218},"p-meaning-1",{"text":251},"工作的对象是\u003Cstrong>解决方案\u003C\u002Fstrong>：将需求转化为有用、受控行为的完整社会技术系统。模型可能是该系统的核心，但它仍然只是一个依赖项。同一个模型可以参与安全的内部搜索助手、不安全的过度授权智能体、低延迟客户功能，或无法经济运营的高成本原型。架构决定了这些差异。",{"id":253,"data":254,"type":218},"p-meaning-2",{"text":255},"因此，一个有用的边界是：\u003Cstrong>业务成果 → 需求 → 系统职责 → 架构决策 → 实施 → 验证 → 运营\u003C\u002Fstrong>。AI 解决方案架构师在这条链上工作，同时与产品、工程、数据、安全、基础设施、治理和领域专家协作。",{"id":257,"data":258,"type":299},"solution-vs-model",{"rows":259,"title":290,"layout":291,"columns":292},[260,266,272,278,284],{"id":261,"label":262,"values":263},"m1","能力",{"model":264,"solution":265},"Which model can generate or reason well enough?","Which combination of model, data, application logic, retrieval, tools and controls produces the required behavior?",{"id":267,"label":268,"values":269},"m2","数据",{"model":270,"solution":271},"What context can fit in the prompt?","What is authoritative, who may access it, how is it retrieved, versioned, filtered and cited?",{"id":273,"label":274,"values":275},"m3","安全",{"model":276,"solution":277},"Does the provider offer security features?","What are the trust boundaries, identities, permissions, secrets, data flows and failure containment mechanisms?",{"id":279,"label":280,"values":281},"m4","运营",{"model":282,"solution":283},"What is the token latency?","How is the complete workload deployed, observed, evaluated, recovered, versioned and cost-controlled?",{"id":285,"label":286,"values":287},"m5","变更",{"model":288,"solution":289},"Can we switch models?","Which dependencies are abstracted, what changes require an ADR, and how do we validate that a replacement still meets requirements?","解决方案比模型更广泛","table",[293,296],{"id":294,"label":295},"model","以模型为中心的问题",{"id":297,"label":298},"solution","解决方案架构问题","comparison",{"id":301,"data":302,"type":42},"h-simple",{"text":303,"level":242},"最简单的例子",{"id":305,"data":306,"type":218},"p-simple-1",{"text":307},"假设一家公司想要一个内部助手，根据维护手册和操作程序回答技术人员的问题。可见功能听起来很简单：输入问题并获得带来源的答案。",{"id":309,"data":310,"type":218},"p-simple-2",{"text":311},"架构问题要大得多。哪些文档是权威的？用户如何认证？检索是否必须遵守部门或站点权限？答案是否只允许使用检索到的证据？对于该数据分类，哪个模型可接受？云提供商是否可以接收内容？当检索没有找到任何内容时会发生什么？引用如何生成？答案质量如何评估？可接受的延迟和成本是多少？谁可以查看日志，日志中可以存储什么？",{"id":313,"data":314,"type":339},"simple-flow",{"steps":315,"title":337,"orientation":338},[316,319,322,325,328,331,334],{"label":317,"description":318},"1. 定义成果","明确用户、业务价值、任务边界，以及成功的答案或行动意味着什么。",{"label":320,"description":321},"2. 捕获需求","明确功能需求、非功能性需求、约束、数据规则、风险容忍度和验收标准。",{"label":323,"description":324},"3. 建立边界","识别用户、身份、应用、权威数据、模型\u002F提供商依赖、工具、外部系统和信任区域。",{"label":326,"description":327},"4. 设计架构","选择数据\u002F检索、模型、编排、工具、权限、运行时、部署、回退和可观测性模式。",{"label":329,"description":330},"5. 记录重大决策","保留架构选择、替代方案、权衡和后果，以便后续变更仍然可理解。",{"label":332,"description":333},"6. 实施与集成","将架构转化为应用代码、API、策略、基础设施、工作流和运营控制。",{"label":335,"description":336},"7. 验证与运营","测试质量、安全、可靠性、成本和用户成果；监控真实工作负载并将证据反馈到决策中。","从需求到可运营的 AI 解决方案","auto","processFlow",{"id":341,"data":342,"type":42},"h-where-simple-stops",{"text":343,"level":242},"简单示例止步之处",{"id":345,"data":346,"type":218},"p-stop-1",{"text":347},"概念验证通常可以跳过生产环境无法跳过的架构。开发者可能硬编码一个提供商、使用共享 API 密钥、将所有文档放在一个索引中、在没有用户上下文过滤的情况下运行检索、逐字记录提示，并手动判断质量。这可以证明可行性，但并不能建立生产架构。",{"id":349,"data":350,"type":218},"p-stop-2",{"text":351},"生产环境引入了相互作用的约束：租户或用户隔离、隐私、数据驻留、吞吐量、延迟、成本、提供商配额、回退行为、可审计性、模型版本变更、检索质量、工具权限、事件响应和部署生命周期。架构师的工作不是同时最大化每一项质量；而是明确权衡，并设计一个满足实际优先级集合的解决方案。",{"id":353,"data":354,"type":42},"h-responsibility-map",{"text":355,"level":242},"架构职责图",{"id":357,"data":358,"type":218},"p-resp-intro",{"text":359},"确切的职责划分因组织而异，但以下图谱概括了解决方案级 AI 架构中反复出现的职责。架构师可能不会亲自实施每一层；其职责是使各层协调一致地契合，并保持关键决策可追溯。",{"id":361,"data":362,"type":291},"responsibility-table",{"content":363,"stretched":43,"withHeadings":14},[364,368,372,376,380,384,388,392,396,400,404],[365,366,367],"架构领域","AI解决方案架构师必须解决的问题","典型输出",[369,370,371],"成果与范围","用户是谁？范围内包含什么任务？系统不得做什么？什么构成成功？","解决方案上下文、能力边界、验收标准",[373,374,375],"需求与非功能需求","适用哪些质量、安全、可用性、延迟、成本、驻留和合规约束？","需求映射、非功能需求、约束、验证标准",[377,378,379],"应用与编排","确定性应用逻辑在哪里结束，AI行为从哪里开始？工作流如何协调？","组件模型、API、编排边界、故障路径",[381,382,383],"权威数据与检索","什么是事实来源？数据如何被摄取、授权、检索、过滤、排序和引用？","数据流、检索架构、元数据和授权规则",[385,386,387],"模型与提供商层","需要哪些能力？哪些提供商\u002F运行时约束重要？应抽象什么？","模型\u002F提供商决策、路由\u002F回退策略、抽象边界",[389,390,391],"工具与代理","系统可以采取哪些行动？哪些行动需要批准？工具身份和权限如何强制执行？","工具契约、代理边界、批准和最小权限规则",[393,394,395],"身份与安全","存在哪些人类和机器身份？密钥保存在哪里？跨越了哪些信任边界？","威胁\u002F信任边界模型、身份传播、密钥和授权设计",[397,398,399],"运行时与部署","组件在哪里执行？什么是本地、云、边缘或混合？存在哪些网络和可用性假设？","部署视图、运行时拓扑、环境和连接决策",[401,402,403],"评估与可观测性","发布前后如何衡量质量？需要哪些追踪、指标、日志和证据？","评估计划、遥测、审计追踪、发布门禁",[405,406,407],"运营与变更","模型\u002F提示\u002F配置\u002F数据版本如何变更、回滚和支持？","运营模型、生命周期控制、ADR、运行手册、变更规则",{"id":409,"data":410,"type":42},"h-requirements",{"text":411,"level":241},"1. 将产品需求转化为架构需求",{"id":413,"data":414,"type":218},"p-requirements-1",{"text":415},"AI架构在模型选择之前就开始了。架构师首先确定解决方案预期实现什么以及在哪些约束下实现。这包括功能行为，也包括缩小设计空间的非功能需求和策略：安全、可靠性、延迟、隐私、驻留、可维护性、成本和运营支持。",{"id":417,"data":418,"type":218},"p-requirements-2",{"text":419},"这就是A02区分重要的地方：诸如“未经授权的用户不得检索受限文档”这样的需求不是架构决策。它是一个驱动因素。关于身份传播、索引分区、元数据过滤、API边界和授权执行的决策是架构响应，之后必须进行验证。",{"id":421,"data":422,"type":42},"h-data",{"text":423,"level":241},"2. 设计权威数据、检索和上下文",{"id":425,"data":426,"type":218},"p-data-1",{"text":427},"AI系统常常在模型行为与企业事实之间的边界上失败。架构师必须定义哪些来源是权威的，新鲜度和来源意味着什么，访问控制如何到达检索，以及检索到的证据如何成为模型上下文。向量数据库、嵌入模型或RAG库本身并不是架构。",{"id":429,"data":430,"type":218},"p-data-2",{"text":431},"微软当前的AI工作负载指南明确表达了同样的分离：应用代码不应绕过数据访问边界；用户或租户上下文应传播到检索和过滤中；基础数据必须为可搜索性而设计，同时仍满足安全和合规要求。",{"id":433,"data":434,"type":42},"h-model",{"text":435,"level":241},"3. 将模型和提供商视为依赖项，而非整个系统",{"id":437,"data":438,"type":218},"p-model-1",{"text":439},"模型选择很重要，但它应由所需能力和约束驱动。架构师考虑推理或生成质量、模态、上下文限制、延迟、数据处理、部署位置、提供商可用性、成本、可观测性和替换风险。",{"id":441,"data":442,"type":218},"p-model-2",{"text":443},"提供商抽象并不自动意味着“更好的架构”。它增加工程成本，并可能隐藏提供商特定的能力。当可移植性、回退、策略分离或多提供商路由是明确需求时，它才是合理的。否则，直接集成可能是更好的决策。关键在于让权衡是有意为之的。",{"id":445,"data":446,"type":42},"h-tools",{"text":447,"level":241},"4. 架构工具、行动和代理边界",{"id":449,"data":450,"type":218},"p-tools-1",{"text":451},"当AI系统可以调用工具、修改数据、发送消息、运行代码或操作业务系统时，架构风险就发生了变化。工具访问需要自己的身份和授权模型。模型请求行动的能力与执行该行动的权限并不相同。",{"id":453,"data":454,"type":218},"p-tools-2",{"text":455},"对于代理工作负载，当前AWS指南强调了额外维度，如代理身份、工具访问、编排、人工监督、追踪、故障处理和迭代推理循环的成本。即使框架隐藏了一些实现机制，这些也是解决方案关注点。",{"id":457,"data":458,"type":42},"h-security",{"text":459,"level":241},"5. 明确信任边界和权限",{"id":461,"data":462,"type":218},"p-security-1",{"text":463},"生产AI解决方案有多个信任边界：浏览器或客户端、应用后端、AI编排、检索\u002F数据服务、模型提供商、工具API、本地运行时和外部系统。每个边界都应回答：谁在调用，代表谁，使用什么凭证，针对哪个资源，有什么审计追踪，以及有什么故障遏制？",{"id":465,"data":466,"type":218},"p-security-2",{"text":467},"安全不能推迟到模型周围的“护栏”。微软的AI工作负载指南明确将安全置于所有架构层，并要求身份\u002F访问管理、数据保护、内容控制和生命周期安全。NIST同样将治理和风险管理视为贯穿AI生命周期的持续活动。",{"id":469,"data":470,"type":42},"h-runtime",{"text":471,"level":241},"6. 决定系统实际运行在哪里",{"id":473,"data":474,"type":218},"p-runtime-1",{"text":475},"“本地AI”、“云AI”和“混合AI”只有在执行和数据路径精确时才是架构陈述。本地桌面进程仍然可以调用云模型。云托管应用可以从本地数据源检索。气隙解决方案具有完全不同的更新、模型分发和可观测性约束。",{"id":477,"data":478,"type":218},"p-runtime-2",{"text":479},"因此，架构师将\u003Cstrong>运行时位置\u003C\u002Fstrong>、\u003Cstrong>推理位置\u003C\u002Fstrong>、\u003Cstrong>数据位置\u003C\u002Fstrong>和\u003Cstrong>控制平面\u003C\u002Fstrong>分开。将它们混为一谈会产生错误的安全性和部署假设。",{"id":481,"data":482,"type":42},"h-eval",{"text":483,"level":241},"7. 定义评估、可观测性和运营验收",{"id":485,"data":486,"type":218},"p-eval-1",{"text":487},"AI 行为具有部分非确定性，因此发布定义不能仅依赖传统的单元测试。架构需要可衡量的验收标准：任务成功率、在相关情况下的有据可依性或引用正确性、拒绝行为、工具安全性、延迟、成本、可靠性和安全测试。具体指标取决于用例。",{"id":489,"data":490,"type":218},"p-eval-2",{"text":491},"微软当前的 Well-Architected AI 指南将监控视为持续性的，并将其应用于模型行为、提示\u002F补全、异常、安全和生产质量门禁。AWS 同样将可观测性、生命周期管理以及模型\u002F提示可追溯性视为运营架构关注点。",{"id":493,"data":494,"type":42},"h-artifacts",{"text":495,"level":242},"该角色应产出什么？",{"id":497,"data":498,"type":218},"p-artifacts-1",{"text":499},"架构不是幻灯片。有用的产出是那些能让工程、安全、产品和运营做出一致决策，并在之后理解系统为何以当前形式存在的工件。",{"id":501,"data":502,"type":291},"artifacts-table",{"content":503,"stretched":43,"withHeadings":14},[504,507,510,513,516,519,522,525,528],[505,506],"工件","目的",[508,509],"解决方案上下文和边界","展示用户、外部系统、主要职责以及范围之外的内容",[511,512],"需求\u002F非功能需求映射","将产品需求和约束与架构工作及验证联系起来",[514,515],"组件和数据流视图","展示应用、数据\u002F检索、模型、工具、身份和运行时交互",[517,518],"信任和权限模型","明确身份、密钥、授权、敏感数据和高风险操作",[520,521],"架构决策记录","保留重要选择、替代方案、权衡、状态和后果",[523,524],"评估和验收计划","定义声称解决方案满足质量和安全期望所需的证据",[526,527],"部署和运营视图","定义环境、运行时位置、可观测性、回滚、事件和生命周期职责",[529,530],"可追溯性链接","连接需求、决策、实现工作、测试和运营证据",{"id":532,"data":533,"type":42},"h-tradeoffs",{"text":534,"level":242},"工作主要是权衡，而不是“最佳实践”选择",{"id":536,"data":537,"type":218},"p-tradeoffs-1",{"text":538},"架构之所以存在，是因为理想的品质会相互冲突。成本较低的模型可能会降低质量。能力更强的模型可能会增加延迟或数据治理约束。激进的缓存可以提高成本和速度，但会使新鲜度复杂化。更自主的代理可以减少人力投入，但会增加影响范围和审计要求。",{"id":540,"data":541,"type":291},"tradeoff-table",{"content":542,"stretched":43,"withHeadings":14},[543,548,553,558,563,568,573,578],[544,545,546,547],"决策","潜在收益","潜在成本\u002F风险","架构问题",[549,550,551,552],"托管云模型","快速采用、强大的托管能力","外部依赖、数据和成本约束","工作负载是否允许该提供商\u002F数据路径并满足弹性需求？",[554,555,556,557],"本地\u002F自托管推理","控制、离线\u002F私有选项","硬件、运营、模型生命周期负担","控制收益是否值得承担运营责任？",[559,560,561,562],"单一提供商集成","实现更简单、完整的提供商功能","更高的切换\u002F故障集中度","是否确实需要可移植性或回退？",[564,565,566,567],"提供商抽象","可移植性、路由和策略分离","最低公分母风险、更多代码\u002F测试","哪些差异必须保持可见而不是被抽象掉？",[569,570,571,572],"大上下文","每次请求更多信息","延迟、成本、注意力稀释、泄露面","是否应检索\u002F过滤数据，而不是始终注入？",[574,575,576,577],"强大的工具\u002F自主性","更多端到端自动化","更高权限和故障影响范围","哪些操作需要最小权限、确认或人工批准？",[579,580,581,582],"严格验证和日志记录","更好的证据和运营","延迟、存储、隐私和复杂性成本","此风险级别需要哪些证据？",{"id":584,"data":585,"type":42},"h-adjacent",{"text":586,"level":242},"这与相邻角色有何不同？",{"id":588,"data":589,"type":218},"p-adjacent-intro",{"text":590},"各公司的职位名称高度重叠。有用的区别在于\u003Cstrong>架构职责范围\u003C\u002Fstrong>，而不是人力资源标签。",{"id":592,"data":593,"type":299},"role-comparison",{"rows":594,"title":631,"layout":291,"columns":632},[595,601,607,613,619,625],{"id":596,"label":597,"values":598},"r1","AI 解决方案架构师",{"role":599,"focus":600},"One concrete AI-enabled solution\u002Fworkload","How requirements, data, models, tools, security, runtime and operations fit together to deliver the target outcome",{"id":602,"label":603,"values":604},"r2","AI 平台架构师",{"role":605,"focus":606},"Reusable AI platform capabilities across many solutions","Shared provider gateways, model access, identity, evaluation, retrieval services, observability, deployment patterns and developer experience",{"id":608,"label":609,"values":610},"r3","企业 AI 架构师",{"role":611,"focus":612},"Organization\u002Fportfolio-level target architecture","Capability landscape, governance, integration principles, shared platforms, standards, sourcing and strategic constraints across domains",{"id":614,"label":615,"values":616},"r4","AI \u002F ML 工程师",{"role":617,"focus":618},"Implementation of AI\u002FML behavior and pipelines","Models, data, inference, evaluation, application logic and engineering tasks within the architecture",{"id":620,"label":621,"values":622},"r5","安全架构师",{"role":623,"focus":624},"Security architecture across systems","Threats, identity, authorization, data protection, controls, assurance and compliance boundaries",{"id":626,"label":627,"values":628},"r6","产品\u002F交付负责人",{"role":629,"focus":630},"Outcome, scope, prioritization and delivery system","Why\u002Fwhat to build, sequencing, stakeholders, milestones, acceptance and value realization","相邻角色回答不同的主要问题",[633,636],{"id":634,"label":635},"role","角色",{"id":637,"label":638},"focus","主要架构关注点",{"id":640,"data":641,"type":218},"p-adjacent-2",{"text":642},"在小型产品团队中，一个人可能覆盖其中多个范围。在大型企业中，它们可能是具有正式评审委员会的独立角色。当职位名称改变时，架构职责并不会消失。",{"id":644,"data":645,"type":42},"h-implementation",{"text":646,"level":242},"实现证据：这些边界如何出现在我自己的工作中",{"id":648,"data":649,"type":225},"implementation-boundary",{"body":650,"title":651,"variant":652},"以下示例是\u003Cstrong>原始实现\u002F项目证据\u003C\u002Fstrong>。它们展示了我在真实项目工作中如何分离产品需求、需求、架构、运行时、模型\u002F提供商、权限和验证。它们并不声称每个组织都必须使用相同结构，也不意味着客户采用或企业级部署。","实现证据，而非通用规则","success",{"id":654,"data":655,"type":42},"h-senseflow",{"text":656,"level":241},"SenseFlow：需求 → 要求 → 架构 → 验证",{"id":658,"data":659,"type":218},"p-senseflow-1",{"text":660},"在 SenseFlow 项目 Source of Truth 中，技术明确从属于产品愿景。开发结构从问题和产品愿景出发，经过用户需求、价值、范围、史诗、故事和验收标准，进入架构、实现、验证和迭代。",{"id":662,"data":663,"type":218},"p-senseflow-2",{"text":664},"需求被设计为可从产品目标 → 能力 → 史诗 → 用户故事 → 验收标准 → 技术任务进行追溯。在可行的情况下，它们包括功能需求、非功能需求、依赖关系、风险、假设、验收标准和验证方法。重大决策保留决策、原因、备选方案、权衡、状态和日期\u002F版本。",{"id":666,"data":667,"type":218},"p-senseflow-3",{"text":668},"这是在选择特定 AI 框架或模型之前进行的架构工作：它保护产品意图与技术决策之间的联系，并使后续变更可审查而非隐式。",{"id":670,"data":671,"type":42},"h-client",{"text":672,"level":241},"Aaasaasa AI 客户端：在集成之前分离概念",{"id":674,"data":675,"type":218},"p-client-1",{"text":676},"Aaasaasa AI 客户端提供了一个更接近实现层面的示例。其 AI Hub 有意分离了\u003Cstrong>代理\u002F客户端\u003C\u002Fstrong>、\u003Cstrong>提供者\u003C\u002Fstrong>、\u003Cstrong>模型\u003C\u002Fstrong>、\u003Cstrong>连接\u002F运行时位置\u003C\u002Fstrong>、\u003Cstrong>权限\u003C\u002Fstrong>和\u003Cstrong>Web 客户端\u003C\u002Fstrong>。本地运行时并不假定意味着本地推理，权限被视为运行时\u002F工具策略，而非模型的属性。",{"id":678,"data":679,"type":218},"p-client-2",{"text":680},"桌面架构还定义了信任边界：Nuxt 渲染器相对于 Electron 主进程是不可信的。一个狭窄的预加载脚本和经过验证的 IPC 调解对 AI 服务、设置、加密密钥、工作区\u002F数据服务和运行时的访问。云凭证保留在特权主进程中；渲染器代码接收规范化状态，而非原始密钥或无限制的操作系统访问。",{"id":682,"data":683,"type":218},"p-client-3",{"text":684},"路由决策同样是架构性的。实现不会从本地路由静默回退到付费云推理；云路由需要明确确认。直接聊天默认没有文件系统或 shell 工具，而代理执行应用选定的工作区和权限配置文件。这些是关于信任、成本、执行和用户期望的解决方案级决策——而非模型特性。",{"id":686,"data":687,"type":42},"h-current-frameworks",{"text":688,"level":242},"当前架构框架如何支持这一更广泛的范围",{"id":690,"data":691,"type":218},"p-frameworks-1",{"text":692},"ISO\u002FIEC\u002FIEEE 42010:2022 为跨软件、系统和企业的架构描述提供了一般性规范。它有意比 AI 更广泛，并且不规定单一的架构方法或职位名称。这使得它在此处作为边界很有用：AI 解决方案架构仍然是架构，具有利益相关者关注点、多个视图和必须清晰表达的重要关系。",{"id":694,"data":695,"type":218},"p-frameworks-2",{"text":696},"NIST AI RMF 1.0 通过\u003Cstrong>治理、映射、测量和管理\u003C\u002Fstrong>来构建 AI 风险管理框架，并强调风险管理应在 AI 系统生命周期中持续进行。生成式 AI 配置文件（NIST AI 600-1）将该框架适配到 GAI 风险和组织优先级。这强化了架构不能止步于功能模型性能。",{"id":698,"data":699,"type":218},"p-frameworks-3",{"text":700},"微软当前的 Azure Well-Architected AI 指南分离了应用程序设计、应用平台、训练数据、基础数据和数据平台关注点，并反复将它们与可靠性、安全性、卓越运营、性能和成本联系起来。AWS 的生成式 AI 和代理式 AI 透镜同样将可观测性、安全性、可靠性、模型\u002F工具生命周期、成本和人工监督视为架构关注点。",{"id":702,"data":703,"type":42},"h-misconceptions",{"text":704,"level":242},"常见误解",{"id":706,"data":707,"type":291},"misconceptions-table",{"content":708,"stretched":43,"withHeadings":14},[709,712,715,718,721,724,727,730],[710,711],"误解","纠正",[713,714],"“架构师选择 LLM。”","模型选择是更大解决方案架构中的一个决策。",[716,717],"“提示工程就是架构。”","提示影响行为，但它们不定义身份、数据访问、信任边界、部署、工具权限或运营。",[719,720],"“RAG 解决企业知识。”","检索只是一个子系统；授权、来源、新鲜度、证据、索引、评估和源治理仍然需要设计。",[722,723],"“本地运行时意味着私有\u002F本地 AI。”","运行时、推理、数据和控制平面位置是独立的架构属性。",[725,726],"“如果供应商提供护栏，安全性就覆盖了。”","安全性涵盖身份、授权、密钥、数据流、工具、日志记录、部署、人工审批和提供者边界。",[728,729],"“架构师必须编写每个组件。”","动手实现可以提高架构质量，但角色由集成决策责任定义，而非亲自编写每一层代码。",[731,732],"“架构图证明生产就绪。”","就绪需要跨质量、安全、运营和业务验收的实施控制和验证证据。",{"id":734,"data":735,"type":42},"h-failures",{"text":736,"level":242},"AI 解决方案架构师应预防的失败模式",{"id":738,"data":739,"type":291},"failures-table",{"content":740,"stretched":43,"withHeadings":14},[741,745,749,753,757,761,765,769,773],[742,743,744],"失败模式","发生原因","架构纠正",[746,747,748],"模型优先设计","一个有前景的模型演示成为系统蓝图","从结果、约束和验证开始；在该框架内选择模型",[750,751,752],"生产环境中的原型权限","共享凭证和广泛访问在 PoC 后仍然存在","尽早定义身份传播、最小权限、工具范围和审批边界",[754,755,756],"未经授权的检索","搜索质量在设计数据访问规则之前就被设计","将用户\u002F租户上下文带入检索，并在数据访问边界强制执行授权",[758,759,760],"静默的提供者\u002F运行时假设","“本地”、“云”和“离线”使用不精确","分别记录运行时、推理、数据和控制平面位置",[762,763,764],"无失败契约","设计了快乐路径，但未设计拒绝\u002F回退\u002F错误行为","指定检索为空、模型不可用、工具失败和策略拒绝的行为",[766,767,768],"实现后评估","在发布前手动判断质量","在架构冻结之前定义可衡量的验收和代表性评估集",[770,771,772],"不可追溯的变更","模型、提示、检索或权限变更没有架构历史","对关键配置进行版本控制，并记录重大决策\u002F验证证据",[774,775,776],"运营仅视为基础设施","部署后 AI 行为不可观测","将跟踪、质量指标、安全事件、成本遥测和回滚一起设计",{"id":778,"data":779,"type":42},"h-decision-framework",{"text":780,"level":242},"实用的决策顺序",{"id":782,"data":783,"type":339},"decision-flow",{"steps":784,"title":809,"orientation":338},[785,788,791,794,797,800,803,806],{"label":786,"description":787},"结果","定义用户\u002F业务结果和明确的非目标。",{"label":789,"description":790},"证据和约束","识别权威数据、策略、非功能需求、风险和验收条件。",{"label":792,"description":793},"系统边界","映射用户、身份、应用程序、数据、模型\u002F提供者、工具和外部系统。",{"label":795,"description":796},"架构选项","比较检索、模型访问、编排、部署、权限、评估和可观测性的模式。",{"label":798,"description":799},"权衡决策","选择重要选项并保留理由、备选方案和后果。",{"label":801,"description":802},"实现契约","将决策转化为 API、模式、权限规则、部署定义和工程任务。",{"label":804,"description":805},"验证","根据原始功能和非功能需求测试已实现的系统。",{"label":807,"description":808},"运营反馈","使用生产证据、事件、质量指标和成本\u002F安全信号触发受控变更。","AI 解决方案架构决策顺序",{"id":811,"data":812,"type":42},"h-edge",{"text":813,"level":242},"边缘情况和角色限制",{"id":815,"data":816,"type":218},"p-edge-1",{"text":817},"一些 AI 产品以模型训练、科学实验或专用硬件为主。在这些情况下，模型\u002F数据科学和 ML 系统架构可能比此处显示的解决方案级映射深入得多。AI 解决方案架构师仍然需要集成和运营边界，但专家架构可能拥有训练平台本身。",{"id":819,"data":820,"type":218},"p-edge-2",{"text":821},"在另一个极端，简单的 SaaS 集成可能并不需要专职架构师。一位高级工程师或技术产品负责人也能承担同样的架构职责。有用的检验标准不是头衔，而是是否在有意地做出并验证重大的跨层决策。",{"id":823,"data":824,"type":218},"p-edge-3",{"text":825},"受监管、主权、气隙隔离、安全关键、高度自治或多租户系统也会改变重心。身份、隔离、驻留、保证、更新机制、人工监督和可审计性在架构中可能比模型质量更为重要。",{"id":827,"data":828,"type":42},"h-change-answer",{"text":829,"level":242},"什么会改变这个答案？",{"id":831,"data":832,"type":218},"p-change-1",{"text":833},"当架构从一个应用转向可复用平台或企业级目标架构时，确切的职责边界会发生变化。这就是为什么 \u003Cstrong>AI 平台架构师\u003C\u002Fstrong> 和 \u003Cstrong>企业 AI 架构\u003C\u002Fstrong> 值得单独进行规范处理，而不是并入这个角色。",{"id":835,"data":836,"type":218},"p-change-2",{"text":837},"技术变化也很重要。新的模型能力、协议、本地运行时和托管服务可以消除一些实现工作，同时创造新的信任或运营边界。稳定的职责是将这些变化理解为系统变化——而不是把新框架当作架构的替代品。",{"id":839,"data":840,"type":42},"h-checklist",{"text":841,"level":242},"AI 解决方案架构师检查清单",{"id":843,"data":844,"type":291},"checklist-table",{"content":845,"stretched":43,"withHeadings":14},[846,849,852,855,857,860,863,866,869,872,875,878,880],[847,848],"检查项","问题",[850,851],"成果","用户\u002F业务结果和非目标边界是否明确？",[853,854],"需求","功能需求、非功能需求、约束和验收标准是否可追溯？",[268,856],"权威来源、来源追溯、时效性、保留和访问规则是否已定义？",[858,859],"检索\u002F上下文","授权是否延伸到检索和上下文构建？",[861,862],"模型\u002F提供商","模型\u002F提供商选择是否基于能力和约束，而非偏好？",[864,865],"工具\u002F智能体","操作边界、权限、审批和失败行为是否明确？",[867,868],"身份\u002F安全","人类\u002F机器身份、密钥和信任边界是否已定义？",[870,871],"运行时","运行时、推理、数据和控制平面的位置是否已区分？",[873,874],"评估","是否有可衡量的证据证明质量、安全性和验收？",[876,877],"可观测性","能否调查生产行为、故障、成本和安全事件？",[286,879],"重大架构决策和替换是否可追溯？",[280,881],"部署、回滚、事件和生命周期的归属是否清晰？",{"id":883,"data":884,"type":42},"h-conclusion",{"text":885,"level":242},"结论",{"id":887,"data":888,"type":218},"p-conclusion-1",{"text":889},"AI 解决方案架构师是将 AI 机会转化为连贯技术系统的人或架构职能。关键技能不是知道最多的模型名称，而是连接产品需求、需求规格、数据、应用架构、AI 能力、安全、运行时、交付和验证，同时不丢失它们之间的边界。",{"id":891,"data":892,"type":218},"p-conclusion-2",{"text":893},"因此，一个强大的 AI 解决方案架构可以总结为：\u003Cstrong>定义目标 → 建立需求和约束 → 设计系统边界 → 明确重大权衡 → 通过清晰的契约实现 → 依据证据进行验证 → 有意识地运营和演进。\u003C\u002Fstrong> 模型很重要。解决方案才是产品。",{"id":895,"data":896,"type":895},"faq",{"items":897,"title":930},[898,902,906,910,914,918,922,926],{"id":899,"answer":900,"question":901},"faq1","AI 解决方案架构师将业务或产品需求转化为具体 AI 赋能解决方案的架构，定义应用逻辑、数据\u002F检索、模型、工具、身份、安全、运行时、评估和运营如何协同工作。","什么是 AI 解决方案架构师？",{"id":903,"answer":904,"question":905},"faq2","不是。这些角色可能重叠，尤其是在小团队中，但 AI 工程师主要是实现角色，而解决方案架构师负责或协调完整工作负载的跨层架构决策和权衡。","AI 解决方案架构师和 AI 工程师是同一个角色吗？",{"id":907,"answer":908,"question":909},"faq3","从定义上讲不需要，但动手实现知识非常有价值，因为 AI 架构跨越 API、数据、检索、安全、运行时和运营行为。这个角色由架构职责定义，而不是由亲自编写每个组件定义。","AI 解决方案架构师需要编码吗？",{"id":911,"answer":912,"question":913},"faq4","不是。模型选择只是一个决策。生产架构还需要数据和检索边界、权限、工具、提供商\u002F运行时选择、可观测性、评估、可靠性、成本和生命周期设计。","选择 LLM 是主要工作吗？",{"id":915,"answer":916,"question":917},"faq5","AI 解决方案架构师专注于一个具体的解决方案或工作负载。AI 平台架构师专注于支持多个解决方案的可复用 AI 能力和护栏。","AI 解决方案架构师和 AI 平台架构师有什么区别？",{"id":919,"answer":920,"question":921},"faq6","解决方案架构师在应用\u002F工作负载范围内工作。企业 AI 架构跨组织组合、目标架构、治理、共享能力、集成原则和战略约束工作。","AI 解决方案架构师和企业 AI 架构师有什么区别？",{"id":923,"answer":924,"question":925},"faq7","当需求证明其合理性时，它们是解决方案内部的架构模式或子系统。RAG 处理基于检索的上下文；智能体增加了规划\u002F工具执行，因此带来额外的身份、权限、编排和运营关注点。","RAG 和智能体适合放在哪里？",{"id":927,"answer":928,"question":929},"faq8","实现加上验证证据：功能测试、评估结果、安全\u002F授权测试、性能和可靠性测量、可观测性、运营演练以及针对原始需求的验收。","什么能证明架构有效？","AI 解决方案架构师 — 常见问题",{"id":932,"data":933,"type":932},"glossary",{"title":934,"entries":935},"核心术语",[936,939,942,946,950,953,956,959],{"term":597,"anchor":937,"definition":938},"ai-solution-architect","对一个具体 AI 赋能解决方案或工作负载的架构职责，将产品需求与应用、数据、模型、工具、安全、运行时和运营设计相结合。",{"term":792,"anchor":940,"definition":941},"system-boundary","解决方案所属范围与其交互的用户、系统、提供商、数据源和环境之间的明确分隔。",{"term":943,"anchor":944,"definition":945},"信任边界","trust-boundary","数据、身份或控制在不同信任假设的组件之间跨越的点，因此需要明确的安全控制。",{"term":947,"anchor":948,"definition":949},"接地","grounding","向 AI 模型提供相关外部信息或证据，使其响应可以基于模型参数之外的来源。",{"term":564,"anchor":951,"definition":952},"provider-abstraction","将解决方案的部分与某个模型\u002F提供商接口解耦的应用边界。当路由、可移植性或策略需求证明其合理性时有用，但并非没有权衡。",{"term":873,"anchor":954,"definition":955},"evaluation","根据定义的验收标准对 AI 工作负载行为进行结构化测量，包括任务质量以及相关的安全、安保、性能和运营属性。",{"term":603,"anchor":957,"definition":958},"ai-platform-architect","专注于多个解决方案使用的可复用 AI 平台能力而非单个工作负载架构的架构角色。",{"term":960,"anchor":961,"definition":962},"企业 AI 架构","enterprise-ai-architecture","组织级架构，协调整个组合中的 AI 能力、平台、治理、集成和战略约束。",{"id":964,"data":965,"type":42},"h-related",{"text":966,"level":242},"相关规范知识",{"id":968,"data":969,"type":218},"p-related-1",{"text":970},"本文属于 AI 架构基础集群。其直接基础是 \u003Cstrong>生成式 AI 解释：模型、检索、工具和应用不是同一回事\u003C\u002Fstrong> 和 \u003Cstrong>ADR 与 NFR：架构决策和系统质量不是同一回事\u003C\u002Fstrong>。相邻的规范节点包括 \u003Cstrong>智能体 AI 解释\u003C\u002Fstrong>、\u003Cstrong>AI 系统中的真相来源\u003C\u002Fstrong>、\u003Cstrong>向量数据库、嵌入和重排序\u003C\u002Fstrong>、\u003Cstrong>什么是上下文工程？\u003C\u002Fstrong>、\u003Cstrong>RBAC 与租户隔离\u003C\u002Fstrong>、\u003Cstrong>AI 平台架构师\u003C\u002Fstrong>、\u003Cstrong>企业 AI 架构\u003C\u002Fstrong> 和 \u003Cstrong>AI 治理\u003C\u002Fstrong>。在这些节点尚未发布的地方，URL 有意不进行虚构。",{"id":972,"data":973,"type":980},"related-rag",{"link":974,"meta":975},"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works",{"image":976,"title":978,"description":979},{"url":977},"","什么是 RAG？对其工作原理的最简单解释","stajic.de 现有的关于检索增强生成的规范解释，对 AI 解决方案架构的检索\u002F接地部分有用。","linkTool",{"id":982,"data":983,"type":42},"h-sources",{"text":984,"level":242},"主要来源和当前架构指南",{"id":986,"data":987,"type":218},"p-sources-note",{"text":988},"以下外部来源支持一般架构主张；SenseFlow 和 Aaasaasa AI Client 部分是明确的原创项目\u002F实现证据。当前状态参考已于 2026 年 10 月 8 日核对。NIST 指出 AI RMF 1.0 正在修订中，因此当后续版本发布时，对版本敏感的治理参考应重新核对。",{"id":990,"data":991,"type":980},"src-iso-42010",{"link":992,"meta":993},"https:\u002F\u002Fwww.iso.org\u002Fstandard\u002F74393.html",{"image":994,"title":995,"description":996},{"url":977},"ISO\u002FIEC\u002FIEEE 42010:2022 — 架构描述","关于架构描述的结构和表达的现行国际标准。它将架构与其描述区分开来，并且不规定单一的架构方法、工具或记录格式。",{"id":998,"data":999,"type":980},"src-nist-rmf",{"link":1000,"meta":1001},"https:\u002F\u002Fwww.nist.gov\u002Fitl\u002Fai-risk-management-framework",{"image":1002,"title":1003,"description":1004},{"url":977},"NIST AI 风险管理框架","NIST 的 AI RMF 资源页面。截至 2026 年 10 月，该页面指出 AI RMF 1.0 正在修订中，并链接了生成式 AI 配置文件及相关资源。",{"id":1006,"data":1007,"type":980},"src-nist-core",{"link":1008,"meta":1009},"https:\u002F\u002Fairc.nist.gov\u002Fairmf-resources\u002Fairmf\u002F5-sec-core\u002F",{"image":1010,"title":1011,"description":1012},{"url":977},"NIST AI RMF 核心 — 治理、映射、测量、管理","NIST AIRC 对 AI RMF 1.0 核心的官方介绍，包括四大功能以及面向生命周期的风险管理框架。",{"id":1014,"data":1015,"type":980},"src-nist-gai",{"link":1016,"meta":1017},"https:\u002F\u002Fwww.nist.gov\u002Fpublications\u002Fartificial-intelligence-risk-management-framework-generative-artificial-intelligence",{"image":1018,"title":1019,"description":1020},{"url":977},"NIST AI 600-1 — 生成式 AI 配置文件","面向 AI RMF 1.0 的跨行业生成式 AI 配置文件，于 2024 年 7 月 26 日发布，并由 NIST 于 2026 年更新。",{"id":1022,"data":1023,"type":980},"src-ms-start",{"link":1024,"meta":1025},"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fget-started",{"image":1026,"title":1027,"description":1028},{"url":977},"Microsoft Azure 架构完善框架 — AI 工作负载","当前的工作负载级架构指南，涵盖 AI 应用程序设计、应用程序平台、训练数据、基础数据、数据平台以及生产就绪相关事项。",{"id":1030,"data":1031,"type":980},"src-ms-app",{"link":1032,"meta":1033},"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fapplication-design",{"image":1034,"title":1035,"description":1036},{"url":977},"Microsoft — AI 工作负载的应用程序设计","关于模型\u002F工具抽象、数据访问边界、身份传播、授权以及客户端、智能、知识和工具层分离的指南。",{"id":1038,"data":1039,"type":980},"src-ms-security",{"link":1040,"meta":1041},"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fdesign-principles",{"image":1042,"title":1043,"description":1044},{"url":977},"Microsoft — AI 工作负载的设计原则","当前 AI 工作负载设计原则，涵盖可靠性、安全性、成本、运营卓越和性能，包括身份和数据保护责任。",{"id":1046,"data":1047,"type":980},"src-ms-ops",{"link":1048,"meta":1049},"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fmlops-genaiops",{"image":1050,"title":1051,"description":1052},{"url":977},"Microsoft — 面向 AI 工作负载的 MLOps 和 GenAIOps","生产生命周期指南，涵盖监控、质量门禁、模型\u002F提示行为、安全性和运营度量。",{"id":1054,"data":1055,"type":980},"src-aws-genai",{"link":1056,"meta":1057},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fgenerative-ai-lens\u002F",{"image":1058,"title":1059,"description":1060},{"url":977},"AWS 架构完善框架生成式 AI 透镜","面向生成式 AI 工作负载的 AWS 架构指南，涵盖运营卓越、安全性、可靠性、性能效率、成本优化和可持续性。",{"id":1062,"data":1063,"type":980},"src-aws-agentic",{"link":1064,"meta":1065},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fagentic-ai-lens\u002F",{"image":1066,"title":1067,"description":1068},{"url":977},"AWS 架构完善框架代理式 AI 透镜","发布于 2026 年，涵盖代理式 AI 特有的架构关注点，包括身份、工具、编排、人工监督、可靠性、追踪和推理循环成本。","2.31","AI解决方案架构师将业务需求转化为生产就绪的AI系统，涵盖数据、模型、工具、安全、运行时、评估和运维。","\u002Fuploads\u002F2026\u002F10\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs-1791476643267-1st5xz.webp","what-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs-1791476643267-1st5xz","PUBLISHED","2026-10-08T12:23:00.000Z","2026-10-08T16:23:08.916Z","2026-10-08T16:31:54.093Z",{"en":1078,"de":1079,"sr":1080,"es":1081,"fr":1082,"it":1083,"ru":1084,"zh":1085},"\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fde\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fsr\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fes\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Ffr\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fit\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fru\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","\u002Fzh\u002Fblog\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs",[1087,1091,1095,1099],{"id":1088,"name":1089,"slug":1090},57,"数据边界","data-boundaries",{"id":1092,"name":1093,"slug":1094},84,"策略与数据边界","policy-and-data",{"id":1096,"name":1097,"slug":1098},80,"访问与身份","access-and-identity",{"id":1100,"name":1101,"slug":1102},54,"威胁模型","threat-model",{"id":1104,"login":1105,"email":1106,"displayName":1107},"20","rooth8233","aleksandar@stajic.de","Aleksandar Stajić",[1109,1796],{"lang":1110,"title":1111,"content":1112,"contentJson":1113,"excerpt":1795},"en","What Is an AI Solution Architect? System Boundaries, Responsibilities and Trade-offs","{\"time\":1791476244367,\"blocks\":[{\"id\":\"intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"An \u003Cstrong>AI Solution Architect\u003C\u002Fstrong> translates a business or product need into the architecture of a concrete AI-enabled solution. The role defines system boundaries and the significant choices across application logic, authoritative data, retrieval and context, models and providers, tools or agents, identity and permissions, security, runtime and deployment, observability, evaluation, cost and operational behavior. It is not simply model selection or prompt engineering: the architectural responsibility is to make the whole solution implementable, governable, testable and operable.\"}},{\"id\":\"direct\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"\u003Cstrong>An AI Solution Architect designs the complete AI-enabled solution, not just the AI model.\u003C\u002Fstrong> The role connects requirements and non-functional requirements to architecture decisions, composes the necessary application\u002Fdata\u002Fmodel\u002Ftool\u002Fruntime layers, makes trust and failure boundaries explicit, and defines how the implemented system will be validated and operated.\"}},{\"id\":\"role-note\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Terminology note\",\"body\":\"\u003Cstrong>AI Solution Architect is a practical role label, not a universally standardized job title.\u003C\u002Fstrong> ISO\u002FIEC\u002FIEEE 42010:2022 standardizes concepts for architecture descriptions; it does not define this job role. Organizations can distribute the responsibilities across several people. In this article, the term means the architecture responsibility for one concrete AI-enabled solution or workload.\"}},{\"id\":\"version-note\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Current-source note — 8 October 2026\",\"body\":\"The architecture principles here are intentionally vendor-neutral, while current vendor guidance is used as implementation evidence. NIST AI RMF 1.0 is currently under revision; NIST AI 600-1 remains the published Generative AI Profile. Microsoft and AWS guidance cited below reflects current production concerns such as identity, data boundaries, model abstraction, security, observability, evaluation, reliability and cost.\"}},{\"id\":\"toc\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3}},{\"id\":\"h-meaning\",\"type\":\"header\",\"data\":{\"text\":\"What does an AI Solution Architect actually architect?\",\"level\":2}},{\"id\":\"p-meaning-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The object of the work is the \u003Cstrong>solution\u003C\u002Fstrong>: the complete socio-technical system that turns a need into useful, controlled behavior. A model may be central to that system, but it is still only one dependency. The same model can participate in a safe internal search assistant, an unsafe over-privileged agent, a low-latency customer feature, or a high-cost prototype that cannot be operated economically. Architecture determines those differences.\"}},{\"id\":\"p-meaning-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"A useful boundary is therefore: \u003Cstrong>business outcome → requirements → system responsibilities → architecture decisions → implementation → validation → operation\u003C\u002Fstrong>. The AI Solution Architect works across this chain while collaborating with product, engineering, data, security, infrastructure, governance and domain specialists.\"}},{\"id\":\"solution-vs-model\",\"type\":\"comparison\",\"data\":{\"title\":\"The solution is wider than the model\",\"layout\":\"table\",\"columns\":[{\"id\":\"model\",\"label\":\"Model-centric question\"},{\"id\":\"solution\",\"label\":\"Solution-architecture question\"}],\"rows\":[{\"id\":\"m1\",\"label\":\"Capability\",\"values\":{\"model\":\"Which model can generate or reason well enough?\",\"solution\":\"Which combination of model, data, application logic, retrieval, tools and controls produces the required behavior?\"}},{\"id\":\"m2\",\"label\":\"Data\",\"values\":{\"model\":\"What context can fit in the prompt?\",\"solution\":\"What is authoritative, who may access it, how is it retrieved, versioned, filtered and cited?\"}},{\"id\":\"m3\",\"label\":\"Security\",\"values\":{\"model\":\"Does the provider offer security features?\",\"solution\":\"What are the trust boundaries, identities, permissions, secrets, data flows and failure containment mechanisms?\"}},{\"id\":\"m4\",\"label\":\"Operations\",\"values\":{\"model\":\"What is the token latency?\",\"solution\":\"How is the complete workload deployed, observed, evaluated, recovered, versioned and cost-controlled?\"}},{\"id\":\"m5\",\"label\":\"Change\",\"values\":{\"model\":\"Can we switch models?\",\"solution\":\"Which dependencies are abstracted, what changes require an ADR, and how do we validate that a replacement still meets requirements?\"}}]}},{\"id\":\"h-simple\",\"type\":\"header\",\"data\":{\"text\":\"The simplest example\",\"level\":2}},{\"id\":\"p-simple-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Imagine a company wants an internal assistant that answers technicians’ questions from maintenance manuals and operating procedures. The visible feature sounds simple: type a question and receive an answer with sources.\"}},{\"id\":\"p-simple-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The architecture question is much larger. Which documents are authoritative? How are users authenticated? Must retrieval respect department or site permissions? Is the answer allowed to use only retrieved evidence? Which model is acceptable for the data classification? Can a cloud provider receive the content? What happens when retrieval finds nothing? How are citations produced? How is answer quality evaluated? What latency and cost are acceptable? Who can see logs, and what may be stored in them?\"}},{\"id\":\"simple-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"From need to an operable AI solution\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Define the outcome\",\"description\":\"Clarify the user, business value, task boundary and what a successful answer or action means.\"},{\"label\":\"2. Capture requirements\",\"description\":\"Make functional requirements, NFRs, constraints, data rules, risk tolerance and acceptance criteria explicit.\"},{\"label\":\"3. Establish boundaries\",\"description\":\"Identify users, identities, applications, authoritative data, model\u002Fprovider dependencies, tools, external systems and trust zones.\"},{\"label\":\"4. Design the architecture\",\"description\":\"Choose data\u002Fretrieval, model, orchestration, tool, permission, runtime, deployment, fallback and observability patterns.\"},{\"label\":\"5. Record significant decisions\",\"description\":\"Preserve architectural choices, alternatives, trade-offs and consequences so later changes remain understandable.\"},{\"label\":\"6. Implement and integrate\",\"description\":\"Turn the architecture into application code, APIs, policies, infrastructure, workflows and operational controls.\"},{\"label\":\"7. Validate and operate\",\"description\":\"Test quality, security, reliability, cost and user outcomes; monitor the real workload and feed evidence back into decisions.\"}]}},{\"id\":\"h-where-simple-stops\",\"type\":\"header\",\"data\":{\"text\":\"Where the simple example stops\",\"level\":2}},{\"id\":\"p-stop-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A proof of concept can often skip architecture that production cannot. A developer may hard-code one provider, use a shared API key, place all documents in one index, run retrieval without user-context filtering, log prompts verbatim and judge quality manually. That can demonstrate feasibility, but it does not establish a production architecture.\"}},{\"id\":\"p-stop-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Production introduces constraints that interact: tenant or user isolation, privacy, data residency, throughput, latency, cost, provider quotas, fallback behavior, auditability, model version changes, retrieval quality, tool permissions, incident response and deployment lifecycle. The architect’s job is not to maximize every quality at once; it is to make the trade-offs explicit and design a solution that satisfies the actual priority set.\"}},{\"id\":\"h-responsibility-map\",\"type\":\"header\",\"data\":{\"text\":\"Architecture responsibility map\",\"level\":2}},{\"id\":\"p-resp-intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"The exact split varies by organization, but the following map captures the recurring responsibilities of solution-level AI architecture. The architect may not personally implement every layer; the responsibility is to make the layers fit together coherently and to keep the critical decisions traceable.\"}},{\"id\":\"responsibility-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Architecture area\",\"Questions the AI Solution Architect must resolve\",\"Typical outputs\"],[\"Outcome and scope\",\"Who is the user? What task is in scope? What must the system not do? What constitutes success?\",\"Solution context, capability boundary, acceptance criteria\"],[\"Requirements and NFRs\",\"What quality, security, availability, latency, cost, residency and compliance constraints apply?\",\"Requirement map, NFRs, constraints, validation criteria\"],[\"Application and orchestration\",\"Where does deterministic application logic end and AI behavior begin? How are workflows coordinated?\",\"Component model, APIs, orchestration boundaries, failure paths\"],[\"Authoritative data and retrieval\",\"What is the Source of Truth? How is data ingested, authorized, retrieved, filtered, ranked and cited?\",\"Data flows, retrieval architecture, metadata and authorization rules\"],[\"Model and provider layer\",\"Which capabilities are required? Which provider\u002Fruntime constraints matter? What should be abstracted?\",\"Model\u002Fprovider decision, routing\u002Ffallback policy, abstraction boundary\"],[\"Tools and agents\",\"What actions can the system take? Which actions require approval? How are tool identities and permissions enforced?\",\"Tool contracts, agent boundaries, approval and least-privilege rules\"],[\"Identity and security\",\"Which human and machine identities exist? Where are secrets held? Which trust boundaries are crossed?\",\"Threat\u002Ftrust boundary model, identity propagation, secrets and authorization design\"],[\"Runtime and deployment\",\"Where do components execute? What is local, cloud, edge or hybrid? What network and availability assumptions exist?\",\"Deployment view, runtime topology, environment and connectivity decisions\"],[\"Evaluation and observability\",\"How is quality measured before and after release? What traces, metrics, logs and evidence are needed?\",\"Evaluation plan, telemetry, audit trail, release gates\"],[\"Operations and change\",\"How are models\u002Fprompts\u002Fconfiguration\u002Fdata versions changed, rolled back and supported?\",\"Operational model, lifecycle controls, ADRs, runbooks, change rules\"]]}},{\"id\":\"h-requirements\",\"type\":\"header\",\"data\":{\"text\":\"1. Turn product need into architectural requirements\",\"level\":3}},{\"id\":\"p-requirements-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"AI architecture begins before model selection. The architect first determines what the solution is expected to achieve and under which constraints. This includes functional behavior, but also the NFRs and policies that narrow the design space: security, reliability, latency, privacy, residency, maintainability, cost and operational support.\"}},{\"id\":\"p-requirements-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is where A02’s distinction matters: a requirement such as “unauthorized users must not retrieve restricted documents” is not an architecture decision. It is a driver. Decisions about identity propagation, index partitioning, metadata filtering, API boundaries and authorization enforcement are architectural responses that must later be validated.\"}},{\"id\":\"h-data\",\"type\":\"header\",\"data\":{\"text\":\"2. Design authoritative data, retrieval and context\",\"level\":3}},{\"id\":\"p-data-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"AI systems often fail at the boundary between model behavior and enterprise truth. An architect must define which sources are authoritative, what freshness and provenance mean, how access control reaches retrieval, and how retrieved evidence becomes model context. A vector database, embedding model or RAG library is not the architecture by itself.\"}},{\"id\":\"p-data-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft’s current AI workload guidance makes the same separation explicit: application code should not bypass data-access boundaries; user or tenant context should propagate into retrieval and filtering; grounding data must be designed for searchability while still meeting security and compliance requirements.\"}},{\"id\":\"h-model\",\"type\":\"header\",\"data\":{\"text\":\"3. Treat models and providers as dependencies, not the whole system\",\"level\":3}},{\"id\":\"p-model-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Model selection matters, but it should be driven by required capability and constraints. The architect considers reasoning or generation quality, modality, context limits, latency, data handling, deployment location, provider availability, cost, observability and replacement risk.\"}},{\"id\":\"p-model-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Provider abstraction is not automatically “better architecture.” It adds engineering cost and can hide provider-specific capabilities. It is justified when portability, fallback, policy separation or multi-provider routing is an explicit requirement. Otherwise a direct integration can be the better decision. The point is to make the trade-off intentional.\"}},{\"id\":\"h-tools\",\"type\":\"header\",\"data\":{\"text\":\"4. Architect tools, actions and agent boundaries\",\"level\":3}},{\"id\":\"p-tools-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"When an AI system can call tools, modify data, send messages, run code or operate business systems, the architectural risk changes. Tool access needs its own identity and authorization model. The model’s ability to request an action is not the same as permission to execute it.\"}},{\"id\":\"p-tools-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"For agentic workloads, current AWS guidance emphasizes additional dimensions such as agent identities, tool access, orchestration, human oversight, tracing, failure handling and cost of iterative reasoning loops. These are solution concerns even when a framework hides some of the implementation mechanics.\"}},{\"id\":\"h-security\",\"type\":\"header\",\"data\":{\"text\":\"5. Make trust boundaries and permissions explicit\",\"level\":3}},{\"id\":\"p-security-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A production AI solution has multiple trust boundaries: browser or client, application backend, AI orchestration, retrieval\u002Fdata services, model providers, tool APIs, local runtimes and external systems. Each boundary should answer: who is calling, on whose behalf, with what credential, for which resource, with what audit trail, and with what failure containment?\"}},{\"id\":\"p-security-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Security cannot be deferred to a “guardrail” around the model. Microsoft’s AI workload guidance explicitly places security across all architecture layers and calls for identity\u002Faccess management, data protection, content controls and lifecycle security. NIST likewise treats governance and risk management as continuous across the AI lifecycle.\"}},{\"id\":\"h-runtime\",\"type\":\"header\",\"data\":{\"text\":\"6. Decide where the system actually runs\",\"level\":3}},{\"id\":\"p-runtime-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"“Local AI,” “cloud AI,” and “hybrid AI” are architectural statements only when the execution and data paths are precise. A local desktop process can still call a cloud model. A cloud-hosted application can retrieve from an on-premises data source. An air-gapped solution has entirely different update, model-distribution and observability constraints.\"}},{\"id\":\"p-runtime-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The architect therefore separates \u003Cstrong>runtime location\u003C\u002Fstrong>, \u003Cstrong>inference location\u003C\u002Fstrong>, \u003Cstrong>data location\u003C\u002Fstrong> and \u003Cstrong>control plane\u003C\u002Fstrong>. Conflating them creates false security and deployment assumptions.\"}},{\"id\":\"h-eval\",\"type\":\"header\",\"data\":{\"text\":\"7. Define evaluation, observability and operational acceptance\",\"level\":3}},{\"id\":\"p-eval-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"AI behavior is partly nondeterministic, so the release definition cannot rely only on conventional unit tests. The architecture needs measurable acceptance: task success, groundedness or citation correctness where relevant, refusal behavior, tool safety, latency, cost, reliability and security tests. The exact metrics depend on the use case.\"}},{\"id\":\"p-eval-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft’s current Well-Architected AI guidance treats monitoring as continuous and applies it across model behavior, prompts\u002Fcompletions, anomalies, security and production quality gates. AWS similarly treats observability, lifecycle management and model\u002Fprompt traceability as operational architecture concerns.\"}},{\"id\":\"h-artifacts\",\"type\":\"header\",\"data\":{\"text\":\"What should the role produce?\",\"level\":2}},{\"id\":\"p-artifacts-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Architecture is not the slide deck. The useful outputs are the artifacts that let engineering, security, product and operations make consistent decisions and later understand why the system exists in its current form.\"}},{\"id\":\"artifacts-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Artifact\",\"Purpose\"],[\"Solution context and boundary\",\"Shows users, external systems, major responsibilities and what is outside scope\"],[\"Requirement\u002FNFR map\",\"Connects product need and constraints to architecture work and validation\"],[\"Component and data-flow views\",\"Shows application, data\u002Fretrieval, model, tools, identity and runtime interactions\"],[\"Trust and permission model\",\"Makes identities, secrets, authorization, sensitive data and high-risk actions explicit\"],[\"Architecture Decision Records\",\"Preserves significant choices, alternatives, trade-offs, status and consequences\"],[\"Evaluation and acceptance plan\",\"Defines evidence required to claim that the solution meets quality and safety expectations\"],[\"Deployment and operational view\",\"Defines environments, runtime locations, observability, rollback, incident and lifecycle responsibilities\"],[\"Traceability links\",\"Connects requirements, decisions, implementation work, tests and operational evidence\"]]}},{\"id\":\"h-tradeoffs\",\"type\":\"header\",\"data\":{\"text\":\"The work is mostly trade-offs, not “best practice” selection\",\"level\":2}},{\"id\":\"p-tradeoffs-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Architecture exists because desirable qualities conflict. A lower-cost model may reduce quality. A more capable model may increase latency or data-governance constraints. Aggressive caching can improve cost and speed while complicating freshness. More autonomous agents can reduce human effort while increasing blast radius and audit requirements.\"}},{\"id\":\"tradeoff-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Decision\",\"Potential benefit\",\"Potential cost \u002F risk\",\"Architectural question\"],[\"Managed cloud model\",\"Fast adoption, strong managed capabilities\",\"External dependency, data and cost constraints\",\"Does the workload permit the provider\u002Fdata path and meet resilience needs?\"],[\"Local\u002Fself-hosted inference\",\"Control, offline\u002Fprivate options\",\"Hardware, operations, model lifecycle burden\",\"Is the control benefit worth the operational responsibility?\"],[\"Single provider integration\",\"Simpler implementation, full provider features\",\"Higher switching\u002Ffailure concentration\",\"Is portability or fallback actually required?\"],[\"Provider abstraction\",\"Portability, routing and policy separation\",\"Lowest-common-denominator risk, more code\u002Ftests\",\"Which differences must remain visible rather than abstracted?\"],[\"Large context\",\"More information per request\",\"Latency, cost, attention dilution, leakage surface\",\"Should data be retrieved\u002Ffiltered instead of always injected?\"],[\"Powerful tools \u002F autonomy\",\"More end-to-end automation\",\"Higher privilege and failure blast radius\",\"Which actions require least privilege, confirmation or human approval?\"],[\"Strict validation and logging\",\"Better evidence and operations\",\"Latency, storage, privacy and complexity cost\",\"What evidence is required for this risk level?\"]]}},{\"id\":\"h-adjacent\",\"type\":\"header\",\"data\":{\"text\":\"How is this different from adjacent roles?\",\"level\":2}},{\"id\":\"p-adjacent-intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"Titles overlap heavily across companies. The useful distinction is the \u003Cstrong>scope of architecture responsibility\u003C\u002Fstrong>, not the HR label.\"}},{\"id\":\"role-comparison\",\"type\":\"comparison\",\"data\":{\"title\":\"Adjacent roles answer different primary questions\",\"layout\":\"table\",\"columns\":[{\"id\":\"role\",\"label\":\"Role\"},{\"id\":\"focus\",\"label\":\"Primary architecture focus\"}],\"rows\":[{\"id\":\"r1\",\"label\":\"AI Solution Architect\",\"values\":{\"role\":\"One concrete AI-enabled solution\u002Fworkload\",\"focus\":\"How requirements, data, models, tools, security, runtime and operations fit together to deliver the target outcome\"}},{\"id\":\"r2\",\"label\":\"AI Platform Architect\",\"values\":{\"role\":\"Reusable AI platform capabilities across many solutions\",\"focus\":\"Shared provider gateways, model access, identity, evaluation, retrieval services, observability, deployment patterns and developer experience\"}},{\"id\":\"r3\",\"label\":\"Enterprise AI Architect\",\"values\":{\"role\":\"Organization\u002Fportfolio-level target architecture\",\"focus\":\"Capability landscape, governance, integration principles, shared platforms, standards, sourcing and strategic constraints across domains\"}},{\"id\":\"r4\",\"label\":\"AI \u002F ML Engineer\",\"values\":{\"role\":\"Implementation of AI\u002FML behavior and pipelines\",\"focus\":\"Models, data, inference, evaluation, application logic and engineering tasks within the architecture\"}},{\"id\":\"r5\",\"label\":\"Security Architect\",\"values\":{\"role\":\"Security architecture across systems\",\"focus\":\"Threats, identity, authorization, data protection, controls, assurance and compliance boundaries\"}},{\"id\":\"r6\",\"label\":\"Product \u002F Delivery Lead\",\"values\":{\"role\":\"Outcome, scope, prioritization and delivery system\",\"focus\":\"Why\u002Fwhat to build, sequencing, stakeholders, milestones, acceptance and value realization\"}}]}},{\"id\":\"p-adjacent-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"In a small product team, one person may cover several of these scopes. In a large enterprise, they may be separate roles with formal review boards. The architecture responsibility does not disappear when the title changes.\"}},{\"id\":\"h-implementation\",\"type\":\"header\",\"data\":{\"text\":\"Implementation evidence: how these boundaries appear in my own work\",\"level\":2}},{\"id\":\"implementation-boundary\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"Implementation evidence, not a universal rule\",\"body\":\"The examples below are \u003Cstrong>original implementation\u002Fproject evidence\u003C\u002Fstrong>. They show how I have separated product need, requirements, architecture, runtime, model\u002Fprovider, permissions and validation in real project work. They are not claims that every organization must use the same structure, and they do not imply customer adoption or enterprise-scale deployment.\"}},{\"id\":\"h-senseflow\",\"type\":\"header\",\"data\":{\"text\":\"SenseFlow: need → requirements → architecture → validation\",\"level\":3}},{\"id\":\"p-senseflow-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"In the SenseFlow project Source of Truth, technology is explicitly subordinate to Product Vision. The development structure moves from problem and product vision through user needs, value, scope, epics, stories and acceptance criteria into architecture, implementation, validation and iteration.\"}},{\"id\":\"p-senseflow-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Requirements are designed to be traceable from Product Goal → Capability → Epic → User Story → Acceptance Criteria → Technical Tasks. Where practical, they include functional requirements, NFRs, dependencies, risks, assumptions, acceptance criteria and validation methods. Significant decisions preserve the decision, reason, alternatives, trade-offs, status and date\u002Fversion.\"}},{\"id\":\"p-senseflow-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That is architectural work before a specific AI framework or model is chosen: it protects the connection between product intent and technical decisions and makes later change reviewable rather than implicit.\"}},{\"id\":\"h-client\",\"type\":\"header\",\"data\":{\"text\":\"Aaasaasa AI Client: separate concepts before integrating them\",\"level\":3}},{\"id\":\"p-client-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Aaasaasa AI Client provides a more implementation-level example. Its AI Hub deliberately separates \u003Cstrong>agent\u002Fclient\u003C\u002Fstrong>, \u003Cstrong>provider\u003C\u002Fstrong>, \u003Cstrong>model\u003C\u002Fstrong>, \u003Cstrong>connection\u002Fruntime location\u003C\u002Fstrong>, \u003Cstrong>permissions\u003C\u002Fstrong> and \u003Cstrong>web client\u003C\u002Fstrong>. A local runtime is not assumed to mean local inference, and permissions are treated as runtime\u002Ftool policy rather than as a property of the model.\"}},{\"id\":\"p-client-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The desktop architecture also defines a trust boundary: the Nuxt renderer is untrusted relative to Electron main. A narrow preload and validated IPC mediate access to AI services, settings, encrypted secrets, workspace\u002Fdata services and runtimes. Cloud credentials remain in the privileged main process; renderer code receives normalized state instead of raw secrets or unrestricted operating-system access.\"}},{\"id\":\"p-client-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Routing decisions are similarly architectural. The implementation does not silently fall back from a local route to paid cloud inference; a cloud route requires explicit confirmation. Direct Chat has no filesystem or shell tools by default, while agent execution applies a selected workspace and permission profile. These are solution-level decisions about trust, cost, execution and user expectation—not model features.\"}},{\"id\":\"h-current-frameworks\",\"type\":\"header\",\"data\":{\"text\":\"How current architecture frameworks support this broader scope\",\"level\":2}},{\"id\":\"p-frameworks-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"ISO\u002FIEC\u002FIEEE 42010:2022 provides a general discipline for architecture descriptions across software, systems and enterprises. It is deliberately broader than AI and does not prescribe one architecting method or job title. That makes it useful here as a boundary: AI solution architecture is still architecture, with stakeholder concerns, multiple views and significant relationships that must be expressed clearly.\"}},{\"id\":\"p-frameworks-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"NIST AI RMF 1.0 frames AI risk management through \u003Cstrong>Govern, Map, Measure and Manage\u003C\u002Fstrong> and emphasizes that risk management should be continuous across the AI system lifecycle. The Generative AI Profile (NIST AI 600-1) adapts that framework to GAI risks and organizational priorities. This reinforces that architecture cannot stop at functional model performance.\"}},{\"id\":\"p-frameworks-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft’s current Azure Well-Architected AI guidance separates application design, application platform, training data, grounding data and data platform concerns and repeatedly connects them to reliability, security, operational excellence, performance and cost. AWS’s Generative AI and Agentic AI lenses similarly treat observability, security, reliability, model\u002Ftool lifecycle, cost and human oversight as architecture concerns.\"}},{\"id\":\"h-misconceptions\",\"type\":\"header\",\"data\":{\"text\":\"Common misconceptions\",\"level\":2}},{\"id\":\"misconceptions-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Misconception\",\"Correction\"],[\"“The architect chooses the LLM.”\",\"Model choice is one decision inside a larger solution architecture.\"],[\"“Prompt engineering is the architecture.”\",\"Prompts affect behavior, but they do not define identity, data access, trust boundaries, deployment, tool permissions or operations.\"],[\"“RAG solves enterprise knowledge.”\",\"Retrieval is only one subsystem; authorization, provenance, freshness, evidence, indexing, evaluation and source governance still need design.\"],[\"“Local runtime means private\u002Flocal AI.”\",\"Runtime, inference, data and control-plane locations are separate architectural properties.\"],[\"“If a vendor offers guardrails, security is covered.”\",\"Security spans identity, authorization, secrets, data flows, tools, logging, deployment, human approval and provider boundaries.\"],[\"“The architect must write every component.”\",\"Hands-on implementation can improve architectural quality, but the role is defined by integrated decision responsibility, not by personally coding every layer.\"],[\"“An architecture diagram proves production readiness.”\",\"Readiness requires implemented controls and validation evidence across quality, security, operations and business acceptance.\"]]}},{\"id\":\"h-failures\",\"type\":\"header\",\"data\":{\"text\":\"Failure modes an AI Solution Architect should prevent\",\"level\":2}},{\"id\":\"failures-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Failure mode\",\"Why it happens\",\"Architectural correction\"],[\"Model-first design\",\"A promising model demo becomes the system blueprint\",\"Start from outcome, constraints and validation; select the model inside that frame\"],[\"Prototype permissions in production\",\"Shared credentials and broad access survive the PoC\",\"Define identity propagation, least privilege, tool scopes and approval boundaries early\"],[\"Retrieval without authorization\",\"Search quality is designed before data-access rules\",\"Carry user\u002Ftenant context into retrieval and enforce authorization at data-access boundaries\"],[\"Silent provider\u002Fruntime assumptions\",\"“Local”, “cloud” and “offline” are used imprecisely\",\"Document runtime, inference, data and control-plane location separately\"],[\"No failure contract\",\"The happy path is designed but refusal\u002Ffallback\u002Ferror behavior is not\",\"Specify retrieval-empty, model-unavailable, tool-failure and policy-denied behavior\"],[\"Evaluation after implementation\",\"Quality is judged manually near launch\",\"Define measurable acceptance and representative evaluation sets before architecture freezes\"],[\"Untraceable change\",\"Models, prompts, retrieval or permissions change without architectural history\",\"Version critical configuration and record significant decisions\u002Fvalidation evidence\"],[\"Operations treated as infrastructure only\",\"AI behavior is not observable after deployment\",\"Design traces, quality metrics, security events, cost telemetry and rollback together\"]]}},{\"id\":\"h-decision-framework\",\"type\":\"header\",\"data\":{\"text\":\"A practical decision sequence\",\"level\":2}},{\"id\":\"decision-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"AI solution architecture decision sequence\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"Outcome\",\"description\":\"Define the user\u002Fbusiness result and explicit non-goals.\"},{\"label\":\"Evidence and constraints\",\"description\":\"Identify authoritative data, policies, NFRs, risks and acceptance conditions.\"},{\"label\":\"System boundary\",\"description\":\"Map users, identities, applications, data, models\u002Fproviders, tools and external systems.\"},{\"label\":\"Architecture options\",\"description\":\"Compare patterns for retrieval, model access, orchestration, deployment, permissions, evaluation and observability.\"},{\"label\":\"Trade-off decisions\",\"description\":\"Select significant options and preserve the rationale, alternatives and consequences.\"},{\"label\":\"Implementation contracts\",\"description\":\"Turn decisions into APIs, schemas, permission rules, deployment definitions and engineering tasks.\"},{\"label\":\"Validation\",\"description\":\"Test the implemented system against the original functional and non-functional requirements.\"},{\"label\":\"Operational feedback\",\"description\":\"Use production evidence, incidents, quality metrics and cost\u002Fsecurity signals to trigger controlled change.\"}]}},{\"id\":\"h-edge\",\"type\":\"header\",\"data\":{\"text\":\"Edge cases and limits of the role\",\"level\":2}},{\"id\":\"p-edge-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Some AI products are dominated by model training, scientific experimentation or specialized hardware. In those cases, model\u002Fdata science and ML systems architecture can become much deeper than the solution-level map shown here. The AI Solution Architect still needs integration and operational boundaries, but specialist architecture may own the training platform itself.\"}},{\"id\":\"p-edge-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"At the other extreme, a simple SaaS integration may not justify a dedicated architect. A senior engineer or technical product lead can carry the same architecture responsibility. The useful test is not the title but whether significant cross-layer decisions are being made deliberately and validated.\"}},{\"id\":\"p-edge-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Regulated, sovereign, air-gapped, safety-critical, highly autonomous or multi-tenant systems also shift the center of gravity. Identity, isolation, residency, assurance, update mechanisms, human oversight and auditability may dominate model quality in the architecture.\"}},{\"id\":\"h-change-answer\",\"type\":\"header\",\"data\":{\"text\":\"What would change this answer?\",\"level\":2}},{\"id\":\"p-change-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The exact responsibility boundary changes when architecture moves from one application to a reusable platform or to enterprise-wide target architecture. That is why \u003Cstrong>AI Platform Architect\u003C\u002Fstrong> and \u003Cstrong>Enterprise AI Architecture\u003C\u002Fstrong> deserve separate canonical treatment rather than being merged into this role.\"}},{\"id\":\"p-change-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Technology changes also matter. New model capabilities, protocols, local runtimes and managed services can remove some implementation work while creating new trust or operational boundaries. The stable responsibility is to understand those changes as system changes—not to treat a new framework as a replacement for architecture.\"}},{\"id\":\"h-checklist\",\"type\":\"header\",\"data\":{\"text\":\"AI Solution Architect checklist\",\"level\":2}},{\"id\":\"checklist-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Check\",\"Question\"],[\"Outcome\",\"Is the user\u002Fbusiness result and non-goal boundary explicit?\"],[\"Requirements\",\"Are functional requirements, NFRs, constraints and acceptance criteria traceable?\"],[\"Data\",\"Are authoritative sources, provenance, freshness, retention and access rules defined?\"],[\"Retrieval\u002Fcontext\",\"Does authorization reach retrieval and context construction?\"],[\"Model\u002Fprovider\",\"Is model\u002Fprovider selection tied to capabilities and constraints rather than preference?\"],[\"Tools\u002Fagents\",\"Are action boundaries, permissions, approvals and failure behavior explicit?\"],[\"Identity\u002Fsecurity\",\"Are human\u002Fmachine identities, secrets and trust boundaries defined?\"],[\"Runtime\",\"Are runtime, inference, data and control-plane locations distinguished?\"],[\"Evaluation\",\"Is there measurable evidence for quality, security and acceptance?\"],[\"Observability\",\"Can production behavior, failures, cost and security events be investigated?\"],[\"Change\",\"Are significant architecture decisions and replacements traceable?\"],[\"Operations\",\"Is ownership for deployment, rollback, incidents and lifecycle clear?\"]]}},{\"id\":\"h-conclusion\",\"type\":\"header\",\"data\":{\"text\":\"Conclusion\",\"level\":2}},{\"id\":\"p-conclusion-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"An AI Solution Architect is the person or architecture function that turns an AI opportunity into a coherent technical system. The key skill is not knowing the most model names; it is connecting product need, requirements, data, application architecture, AI capabilities, security, runtime, delivery and validation without losing the boundaries between them.\"}},{\"id\":\"p-conclusion-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"A strong AI solution architecture can therefore be summarized as: \u003Cstrong>define the target → establish requirements and constraints → design the system boundaries → make significant trade-offs explicit → implement through clear contracts → validate against evidence → operate and evolve deliberately.\u003C\u002Fstrong> The model is important. The solution is the product.\"}},{\"id\":\"faq\",\"type\":\"faq\",\"data\":{\"title\":\"AI Solution Architect — FAQ\",\"items\":[{\"id\":\"faq1\",\"question\":\"What is an AI Solution Architect?\",\"answer\":\"An AI Solution Architect translates a business or product need into the architecture of a concrete AI-enabled solution, defining how application logic, data\u002Fretrieval, models, tools, identity, security, runtime, evaluation and operations work together.\"},{\"id\":\"faq2\",\"question\":\"Is an AI Solution Architect the same as an AI engineer?\",\"answer\":\"No. The roles can overlap, especially in small teams, but an AI engineer is primarily an implementation role while the solution architect owns or coordinates cross-layer architecture decisions and trade-offs for the complete workload.\"},{\"id\":\"faq3\",\"question\":\"Does an AI Solution Architect need to code?\",\"answer\":\"Not by definition, but hands-on implementation knowledge is highly valuable because AI architecture crosses APIs, data, retrieval, security, runtimes and operational behavior. The role is defined by architecture responsibility, not by writing every component personally.\"},{\"id\":\"faq4\",\"question\":\"Is choosing an LLM the main job?\",\"answer\":\"No. Model selection is one decision. Production architecture also needs data and retrieval boundaries, permissions, tools, provider\u002Fruntime choices, observability, evaluation, reliability, cost and lifecycle design.\"},{\"id\":\"faq5\",\"question\":\"What is the difference between an AI Solution Architect and an AI Platform Architect?\",\"answer\":\"An AI Solution Architect focuses on one concrete solution or workload. An AI Platform Architect focuses on reusable AI capabilities and guardrails that support multiple solutions.\"},{\"id\":\"faq6\",\"question\":\"What is the difference between an AI Solution Architect and an Enterprise AI Architect?\",\"answer\":\"The solution architect works at application\u002Fworkload scope. Enterprise AI architecture works across the organizational portfolio, target architecture, governance, shared capabilities, integration principles and strategic constraints.\"},{\"id\":\"faq7\",\"question\":\"Where do RAG and agents fit?\",\"answer\":\"They are architectural patterns or subsystems inside a solution when the requirements justify them. RAG addresses retrieval-grounded context; agents add planning\u002Ftool execution and therefore additional identity, permission, orchestration and operational concerns.\"},{\"id\":\"faq8\",\"question\":\"What proves that the architecture works?\",\"answer\":\"Implementation plus validation evidence: functional tests, evaluation results, security\u002Fauthorization tests, performance and reliability measurements, observability, operational rehearsal and acceptance against the original requirements.\"}]}},{\"id\":\"glossary\",\"type\":\"glossary\",\"data\":{\"title\":\"Core terms\",\"entries\":[{\"term\":\"AI Solution Architect\",\"definition\":\"Architecture responsibility for one concrete AI-enabled solution or workload, integrating product requirements with application, data, model, tool, security, runtime and operational design.\",\"anchor\":\"ai-solution-architect\"},{\"term\":\"System boundary\",\"definition\":\"The explicit separation between what belongs to the solution and the users, systems, providers, data sources and environments it interacts with.\",\"anchor\":\"system-boundary\"},{\"term\":\"Trust boundary\",\"definition\":\"A point where data, identities or control cross between components with different trust assumptions and therefore require explicit security controls.\",\"anchor\":\"trust-boundary\"},{\"term\":\"Grounding\",\"definition\":\"Supplying an AI model with relevant external information or evidence so its response can be based on sources beyond model parameters.\",\"anchor\":\"grounding\"},{\"term\":\"Provider abstraction\",\"definition\":\"An application boundary that decouples parts of the solution from one model\u002Fprovider interface. Useful when justified by routing, portability or policy needs, but not free of trade-offs.\",\"anchor\":\"provider-abstraction\"},{\"term\":\"Evaluation\",\"definition\":\"Structured measurement of AI workload behavior against defined acceptance criteria, including task quality and relevant safety, security, performance and operational properties.\",\"anchor\":\"evaluation\"},{\"term\":\"AI Platform Architect\",\"definition\":\"Architectural role focused on reusable AI platform capabilities used by multiple solutions rather than the architecture of one workload.\",\"anchor\":\"ai-platform-architect\"},{\"term\":\"Enterprise AI Architecture\",\"definition\":\"Organization-level architecture that coordinates AI capabilities, platforms, governance, integration and strategic constraints across a portfolio.\",\"anchor\":\"enterprise-ai-architecture\"}]}},{\"id\":\"h-related\",\"type\":\"header\",\"data\":{\"text\":\"Related canonical knowledge\",\"level\":2}},{\"id\":\"p-related-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"This article sits in the AI Architecture Foundations cluster. Its direct foundations are \u003Cstrong>Generative AI Explained: Models, Retrieval, Tools and Applications Are Not the Same Thing\u003C\u002Fstrong> and \u003Cstrong>ADR vs NFR: Architecture Decisions and System Quality Are Not the Same Thing\u003C\u002Fstrong>. Adjacent canonical nodes include \u003Cstrong>Agentic AI Explained\u003C\u002Fstrong>, \u003Cstrong>Source of Truth in AI Systems\u003C\u002Fstrong>, \u003Cstrong>Vector Databases, Embeddings and Reranking\u003C\u002Fstrong>, \u003Cstrong>What Is Context Engineering?\u003C\u002Fstrong>, \u003Cstrong>RBAC vs Tenant Isolation\u003C\u002Fstrong>, \u003Cstrong>AI Platform Architect\u003C\u002Fstrong>, \u003Cstrong>Enterprise AI Architecture\u003C\u002Fstrong> and \u003Cstrong>AI Governance\u003C\u002Fstrong>. URLs are intentionally not fabricated where those nodes are not yet published.\"}},{\"id\":\"related-rag\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works\",\"meta\":{\"title\":\"What Is RAG? The Simplest Explanation of How It Works\",\"description\":\"Existing stajic.de canonical explanation of retrieval-augmented generation, useful for the retrieval\u002Fgrounding part of AI solution architecture.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"h-sources\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources and current architecture guidance\",\"level\":2}},{\"id\":\"p-sources-note\",\"type\":\"paragraph\",\"data\":{\"text\":\"External sources below support the general architecture claims; the SenseFlow and Aaasaasa AI Client sections are explicitly original project\u002Fimplementation evidence. Current-state references were checked on 8 October 2026. NIST notes that AI RMF 1.0 is being revised, so version-sensitive governance references should be rechecked when a successor is published.\"}},{\"id\":\"src-iso-42010\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.iso.org\u002Fstandard\u002F74393.html\",\"meta\":{\"title\":\"ISO\u002FIEC\u002FIEEE 42010:2022 — Architecture Description\",\"description\":\"Current international standard for the structure and expression of architecture descriptions. It distinguishes architecture from its description and does not prescribe one architecting method, tool or recording format.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-nist-rmf\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nist.gov\u002Fitl\u002Fai-risk-management-framework\",\"meta\":{\"title\":\"NIST AI Risk Management Framework\",\"description\":\"NIST’s AI RMF resource page. As of October 2026 it states that AI RMF 1.0 is being revised and links the Generative AI Profile and related resources.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-nist-core\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fairc.nist.gov\u002Fairmf-resources\u002Fairmf\u002F5-sec-core\u002F\",\"meta\":{\"title\":\"NIST AI RMF Core — Govern, Map, Measure, Manage\",\"description\":\"Official NIST AIRC presentation of the AI RMF 1.0 Core, including the four functions and lifecycle-oriented risk-management framing.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-nist-gai\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nist.gov\u002Fpublications\u002Fartificial-intelligence-risk-management-framework-generative-artificial-intelligence\",\"meta\":{\"title\":\"NIST AI 600-1 — Generative AI Profile\",\"description\":\"Cross-sectoral Generative AI profile for AI RMF 1.0, published 26 July 2024 and updated by NIST in 2026.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-ms-start\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fget-started\",\"meta\":{\"title\":\"Microsoft Azure Well-Architected — AI Workloads\",\"description\":\"Current workload-level architecture guidance covering AI application design, application platform, training data, grounding data, data platform and production-readiness concerns.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-ms-app\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fapplication-design\",\"meta\":{\"title\":\"Microsoft — Application Design for AI Workloads\",\"description\":\"Guidance on model\u002Ftool abstraction, data-access boundaries, identity propagation, authorization and separation of client, intelligence, knowledge and tool layers.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-ms-security\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fdesign-principles\",\"meta\":{\"title\":\"Microsoft — Design Principles for AI Workloads\",\"description\":\"Current AI workload design principles across reliability, security, cost, operational excellence and performance, including identity and data-protection responsibilities.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-ms-ops\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fwell-architected\u002Fai\u002Fmlops-genaiops\",\"meta\":{\"title\":\"Microsoft — MLOps and GenAIOps for AI Workloads\",\"description\":\"Production lifecycle guidance covering monitoring, quality gates, model\u002Fprompt behavior, security and operational measurement.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-aws-genai\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fgenerative-ai-lens\u002F\",\"meta\":{\"title\":\"AWS Well-Architected Generative AI Lens\",\"description\":\"AWS architectural guidance for generative AI workloads across operational excellence, security, reliability, performance efficiency, cost optimization and sustainability.\",\"image\":{\"url\":\"\"}}}},{\"id\":\"src-aws-agentic\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fagentic-ai-lens\u002F\",\"meta\":{\"title\":\"AWS Well-Architected Agentic AI Lens\",\"description\":\"Published in 2026, covering agentic-specific architecture concerns including identities, tools, orchestration, human oversight, reliability, tracing and reasoning-loop cost.\",\"image\":{\"url\":\"\"}}}}],\"version\":\"2.31.0\"}",{"time":1114,"blocks":1115,"version":1794},1791476244367,[1116,1119,1123,1127,1131,1134,1137,1140,1143,1167,1170,1173,1176,1201,1204,1207,1210,1213,1216,1263,1266,1269,1272,1275,1278,1281,1284,1287,1290,1293,1296,1299,1302,1305,1308,1311,1314,1317,1320,1323,1326,1329,1332,1362,1365,1368,1411,1414,1417,1444,1447,1450,1454,1457,1460,1463,1466,1469,1472,1475,1478,1481,1484,1487,1490,1493,1520,1523,1562,1565,1593,1596,1599,1602,1605,1608,1611,1614,1617,1655,1658,1661,1664,1692,1715,1718,1721,1728,1731,1734,1740,1746,1752,1758,1764,1770,1776,1782,1788],{"id":215,"data":1117,"type":218},{"text":1118},"An \u003Cstrong>AI Solution Architect\u003C\u002Fstrong> translates a business or product need into the architecture of a concrete AI-enabled solution. The role defines system boundaries and the significant choices across application logic, authoritative data, retrieval and context, models and providers, tools or agents, identity and permissions, security, runtime and deployment, observability, evaluation, cost and operational behavior. It is not simply model selection or prompt engineering: the architectural responsibility is to make the whole solution implementable, governable, testable and operable.",{"id":220,"data":1120,"type":225},{"body":1121,"title":1122,"variant":224},"\u003Cstrong>An AI Solution Architect designs the complete AI-enabled solution, not just the AI model.\u003C\u002Fstrong> The role connects requirements and non-functional requirements to architecture decisions, composes the necessary application\u002Fdata\u002Fmodel\u002Ftool\u002Fruntime layers, makes trust and failure boundaries explicit, and defines how the implemented system will be validated and operated.","Direct answer",{"id":227,"data":1124,"type":225},{"body":1125,"title":1126,"variant":231},"\u003Cstrong>AI Solution Architect is a practical role label, not a universally standardized job title.\u003C\u002Fstrong> ISO\u002FIEC\u002FIEEE 42010:2022 standardizes concepts for architecture descriptions; it does not define this job role. Organizations can distribute the responsibilities across several people. In this article, the term means the architecture responsibility for one concrete AI-enabled solution or workload.","Terminology note",{"id":233,"data":1128,"type":225},{"body":1129,"title":1130,"variant":231},"The architecture principles here are intentionally vendor-neutral, while current vendor guidance is used as implementation evidence. NIST AI RMF 1.0 is currently under revision; NIST AI 600-1 remains the published Generative AI Profile. Microsoft and AWS guidance cited below reflects current production concerns such as identity, data boundaries, model abstraction, security, observability, evaluation, reliability and cost.","Current-source note — 8 October 2026",{"id":238,"data":1132,"type":243},{"title":1133,"maxLevel":241,"minLevel":242},"Contents",{"id":245,"data":1135,"type":42},{"text":1136,"level":242},"What does an AI Solution Architect actually architect?",{"id":249,"data":1138,"type":218},{"text":1139},"The object of the work is the \u003Cstrong>solution\u003C\u002Fstrong>: the complete socio-technical system that turns a need into useful, controlled behavior. A model may be central to that system, but it is still only one dependency. The same model can participate in a safe internal search assistant, an unsafe over-privileged agent, a low-latency customer feature, or a high-cost prototype that cannot be operated economically. Architecture determines those differences.",{"id":253,"data":1141,"type":218},{"text":1142},"A useful boundary is therefore: \u003Cstrong>business outcome → requirements → system responsibilities → architecture decisions → implementation → validation → operation\u003C\u002Fstrong>. The AI Solution Architect works across this chain while collaborating with product, engineering, data, security, infrastructure, governance and domain specialists.",{"id":257,"data":1144,"type":299},{"rows":1145,"title":1161,"layout":291,"columns":1162},[1146,1149,1152,1155,1158],{"id":261,"label":1147,"values":1148},"Capability",{"model":264,"solution":265},{"id":267,"label":1150,"values":1151},"Data",{"model":270,"solution":271},{"id":273,"label":1153,"values":1154},"Security",{"model":276,"solution":277},{"id":279,"label":1156,"values":1157},"Operations",{"model":282,"solution":283},{"id":285,"label":1159,"values":1160},"Change",{"model":288,"solution":289},"The solution is wider than the model",[1163,1165],{"id":294,"label":1164},"Model-centric question",{"id":297,"label":1166},"Solution-architecture question",{"id":301,"data":1168,"type":42},{"text":1169,"level":242},"The simplest example",{"id":305,"data":1171,"type":218},{"text":1172},"Imagine a company wants an internal assistant that answers technicians’ questions from maintenance manuals and operating procedures. The visible feature sounds simple: type a question and receive an answer with sources.",{"id":309,"data":1174,"type":218},{"text":1175},"The architecture question is much larger. Which documents are authoritative? How are users authenticated? Must retrieval respect department or site permissions? Is the answer allowed to use only retrieved evidence? Which model is acceptable for the data classification? Can a cloud provider receive the content? What happens when retrieval finds nothing? How are citations produced? How is answer quality evaluated? What latency and cost are acceptable? Who can see logs, and what may be stored in them?",{"id":313,"data":1177,"type":339},{"steps":1178,"title":1200,"orientation":338},[1179,1182,1185,1188,1191,1194,1197],{"label":1180,"description":1181},"1. Define the outcome","Clarify the user, business value, task boundary and what a successful answer or action means.",{"label":1183,"description":1184},"2. Capture requirements","Make functional requirements, NFRs, constraints, data rules, risk tolerance and acceptance criteria explicit.",{"label":1186,"description":1187},"3. Establish boundaries","Identify users, identities, applications, authoritative data, model\u002Fprovider dependencies, tools, external systems and trust zones.",{"label":1189,"description":1190},"4. Design the architecture","Choose data\u002Fretrieval, model, orchestration, tool, permission, runtime, deployment, fallback and observability patterns.",{"label":1192,"description":1193},"5. Record significant decisions","Preserve architectural choices, alternatives, trade-offs and consequences so later changes remain understandable.",{"label":1195,"description":1196},"6. Implement and integrate","Turn the architecture into application code, APIs, policies, infrastructure, workflows and operational controls.",{"label":1198,"description":1199},"7. Validate and operate","Test quality, security, reliability, cost and user outcomes; monitor the real workload and feed evidence back into decisions.","From need to an operable AI solution",{"id":341,"data":1202,"type":42},{"text":1203,"level":242},"Where the simple example stops",{"id":345,"data":1205,"type":218},{"text":1206},"A proof of concept can often skip architecture that production cannot. A developer may hard-code one provider, use a shared API key, place all documents in one index, run retrieval without user-context filtering, log prompts verbatim and judge quality manually. That can demonstrate feasibility, but it does not establish a production architecture.",{"id":349,"data":1208,"type":218},{"text":1209},"Production introduces constraints that interact: tenant or user isolation, privacy, data residency, throughput, latency, cost, provider quotas, fallback behavior, auditability, model version changes, retrieval quality, tool permissions, incident response and deployment lifecycle. The architect’s job is not to maximize every quality at once; it is to make the trade-offs explicit and design a solution that satisfies the actual priority set.",{"id":353,"data":1211,"type":42},{"text":1212,"level":242},"Architecture responsibility map",{"id":357,"data":1214,"type":218},{"text":1215},"The exact split varies by organization, but the following map captures the recurring responsibilities of solution-level AI architecture. The architect may not personally implement every layer; the responsibility is to make the layers fit together coherently and to keep the critical decisions traceable.",{"id":361,"data":1217,"type":291},{"content":1218,"stretched":43,"withHeadings":14},[1219,1223,1227,1231,1235,1239,1243,1247,1251,1255,1259],[1220,1221,1222],"Architecture area","Questions the AI Solution Architect must resolve","Typical outputs",[1224,1225,1226],"Outcome and scope","Who is the user? What task is in scope? What must the system not do? What constitutes success?","Solution context, capability boundary, acceptance criteria",[1228,1229,1230],"Requirements and NFRs","What quality, security, availability, latency, cost, residency and compliance constraints apply?","Requirement map, NFRs, constraints, validation criteria",[1232,1233,1234],"Application and orchestration","Where does deterministic application logic end and AI behavior begin? How are workflows coordinated?","Component model, APIs, orchestration boundaries, failure paths",[1236,1237,1238],"Authoritative data and retrieval","What is the Source of Truth? How is data ingested, authorized, retrieved, filtered, ranked and cited?","Data flows, retrieval architecture, metadata and authorization rules",[1240,1241,1242],"Model and provider layer","Which capabilities are required? Which provider\u002Fruntime constraints matter? What should be abstracted?","Model\u002Fprovider decision, routing\u002Ffallback policy, abstraction boundary",[1244,1245,1246],"Tools and agents","What actions can the system take? Which actions require approval? How are tool identities and permissions enforced?","Tool contracts, agent boundaries, approval and least-privilege rules",[1248,1249,1250],"Identity and security","Which human and machine identities exist? Where are secrets held? Which trust boundaries are crossed?","Threat\u002Ftrust boundary model, identity propagation, secrets and authorization design",[1252,1253,1254],"Runtime and deployment","Where do components execute? What is local, cloud, edge or hybrid? What network and availability assumptions exist?","Deployment view, runtime topology, environment and connectivity decisions",[1256,1257,1258],"Evaluation and observability","How is quality measured before and after release? What traces, metrics, logs and evidence are needed?","Evaluation plan, telemetry, audit trail, release gates",[1260,1261,1262],"Operations and change","How are models\u002Fprompts\u002Fconfiguration\u002Fdata versions changed, rolled back and supported?","Operational model, lifecycle controls, ADRs, runbooks, change rules",{"id":409,"data":1264,"type":42},{"text":1265,"level":241},"1. Turn product need into architectural requirements",{"id":413,"data":1267,"type":218},{"text":1268},"AI architecture begins before model selection. The architect first determines what the solution is expected to achieve and under which constraints. This includes functional behavior, but also the NFRs and policies that narrow the design space: security, reliability, latency, privacy, residency, maintainability, cost and operational support.",{"id":417,"data":1270,"type":218},{"text":1271},"This is where A02’s distinction matters: a requirement such as “unauthorized users must not retrieve restricted documents” is not an architecture decision. It is a driver. Decisions about identity propagation, index partitioning, metadata filtering, API boundaries and authorization enforcement are architectural responses that must later be validated.",{"id":421,"data":1273,"type":42},{"text":1274,"level":241},"2. Design authoritative data, retrieval and context",{"id":425,"data":1276,"type":218},{"text":1277},"AI systems often fail at the boundary between model behavior and enterprise truth. An architect must define which sources are authoritative, what freshness and provenance mean, how access control reaches retrieval, and how retrieved evidence becomes model context. A vector database, embedding model or RAG library is not the architecture by itself.",{"id":429,"data":1279,"type":218},{"text":1280},"Microsoft’s current AI workload guidance makes the same separation explicit: application code should not bypass data-access boundaries; user or tenant context should propagate into retrieval and filtering; grounding data must be designed for searchability while still meeting security and compliance requirements.",{"id":433,"data":1282,"type":42},{"text":1283,"level":241},"3. Treat models and providers as dependencies, not the whole system",{"id":437,"data":1285,"type":218},{"text":1286},"Model selection matters, but it should be driven by required capability and constraints. The architect considers reasoning or generation quality, modality, context limits, latency, data handling, deployment location, provider availability, cost, observability and replacement risk.",{"id":441,"data":1288,"type":218},{"text":1289},"Provider abstraction is not automatically “better architecture.” It adds engineering cost and can hide provider-specific capabilities. It is justified when portability, fallback, policy separation or multi-provider routing is an explicit requirement. Otherwise a direct integration can be the better decision. The point is to make the trade-off intentional.",{"id":445,"data":1291,"type":42},{"text":1292,"level":241},"4. Architect tools, actions and agent boundaries",{"id":449,"data":1294,"type":218},{"text":1295},"When an AI system can call tools, modify data, send messages, run code or operate business systems, the architectural risk changes. Tool access needs its own identity and authorization model. The model’s ability to request an action is not the same as permission to execute it.",{"id":453,"data":1297,"type":218},{"text":1298},"For agentic workloads, current AWS guidance emphasizes additional dimensions such as agent identities, tool access, orchestration, human oversight, tracing, failure handling and cost of iterative reasoning loops. These are solution concerns even when a framework hides some of the implementation mechanics.",{"id":457,"data":1300,"type":42},{"text":1301,"level":241},"5. Make trust boundaries and permissions explicit",{"id":461,"data":1303,"type":218},{"text":1304},"A production AI solution has multiple trust boundaries: browser or client, application backend, AI orchestration, retrieval\u002Fdata services, model providers, tool APIs, local runtimes and external systems. Each boundary should answer: who is calling, on whose behalf, with what credential, for which resource, with what audit trail, and with what failure containment?",{"id":465,"data":1306,"type":218},{"text":1307},"Security cannot be deferred to a “guardrail” around the model. Microsoft’s AI workload guidance explicitly places security across all architecture layers and calls for identity\u002Faccess management, data protection, content controls and lifecycle security. NIST likewise treats governance and risk management as continuous across the AI lifecycle.",{"id":469,"data":1309,"type":42},{"text":1310,"level":241},"6. Decide where the system actually runs",{"id":473,"data":1312,"type":218},{"text":1313},"“Local AI,” “cloud AI,” and “hybrid AI” are architectural statements only when the execution and data paths are precise. A local desktop process can still call a cloud model. A cloud-hosted application can retrieve from an on-premises data source. An air-gapped solution has entirely different update, model-distribution and observability constraints.",{"id":477,"data":1315,"type":218},{"text":1316},"The architect therefore separates \u003Cstrong>runtime location\u003C\u002Fstrong>, \u003Cstrong>inference location\u003C\u002Fstrong>, \u003Cstrong>data location\u003C\u002Fstrong> and \u003Cstrong>control plane\u003C\u002Fstrong>. Conflating them creates false security and deployment assumptions.",{"id":481,"data":1318,"type":42},{"text":1319,"level":241},"7. Define evaluation, observability and operational acceptance",{"id":485,"data":1321,"type":218},{"text":1322},"AI behavior is partly nondeterministic, so the release definition cannot rely only on conventional unit tests. The architecture needs measurable acceptance: task success, groundedness or citation correctness where relevant, refusal behavior, tool safety, latency, cost, reliability and security tests. The exact metrics depend on the use case.",{"id":489,"data":1324,"type":218},{"text":1325},"Microsoft’s current Well-Architected AI guidance treats monitoring as continuous and applies it across model behavior, prompts\u002Fcompletions, anomalies, security and production quality gates. AWS similarly treats observability, lifecycle management and model\u002Fprompt traceability as operational architecture concerns.",{"id":493,"data":1327,"type":42},{"text":1328,"level":242},"What should the role produce?",{"id":497,"data":1330,"type":218},{"text":1331},"Architecture is not the slide deck. The useful outputs are the artifacts that let engineering, security, product and operations make consistent decisions and later understand why the system exists in its current form.",{"id":501,"data":1333,"type":291},{"content":1334,"stretched":43,"withHeadings":14},[1335,1338,1341,1344,1347,1350,1353,1356,1359],[1336,1337],"Artifact","Purpose",[1339,1340],"Solution context and boundary","Shows users, external systems, major responsibilities and what is outside scope",[1342,1343],"Requirement\u002FNFR map","Connects product need and constraints to architecture work and validation",[1345,1346],"Component and data-flow views","Shows application, data\u002Fretrieval, model, tools, identity and runtime interactions",[1348,1349],"Trust and permission model","Makes identities, secrets, authorization, sensitive data and high-risk actions explicit",[1351,1352],"Architecture Decision Records","Preserves significant choices, alternatives, trade-offs, status and consequences",[1354,1355],"Evaluation and acceptance plan","Defines evidence required to claim that the solution meets quality and safety expectations",[1357,1358],"Deployment and operational view","Defines environments, runtime locations, observability, rollback, incident and lifecycle responsibilities",[1360,1361],"Traceability links","Connects requirements, decisions, implementation work, tests and operational evidence",{"id":532,"data":1363,"type":42},{"text":1364,"level":242},"The work is mostly trade-offs, not “best practice” selection",{"id":536,"data":1366,"type":218},{"text":1367},"Architecture exists because desirable qualities conflict. A lower-cost model may reduce quality. A more capable model may increase latency or data-governance constraints. Aggressive caching can improve cost and speed while complicating freshness. More autonomous agents can reduce human effort while increasing blast radius and audit requirements.",{"id":540,"data":1369,"type":291},{"content":1370,"stretched":43,"withHeadings":14},[1371,1376,1381,1386,1391,1396,1401,1406],[1372,1373,1374,1375],"Decision","Potential benefit","Potential cost \u002F risk","Architectural question",[1377,1378,1379,1380],"Managed cloud model","Fast adoption, strong managed capabilities","External dependency, data and cost constraints","Does the workload permit the provider\u002Fdata path and meet resilience needs?",[1382,1383,1384,1385],"Local\u002Fself-hosted inference","Control, offline\u002Fprivate options","Hardware, operations, model lifecycle burden","Is the control benefit worth the operational responsibility?",[1387,1388,1389,1390],"Single provider integration","Simpler implementation, full provider features","Higher switching\u002Ffailure concentration","Is portability or fallback actually required?",[1392,1393,1394,1395],"Provider abstraction","Portability, routing and policy separation","Lowest-common-denominator risk, more code\u002Ftests","Which differences must remain visible rather than abstracted?",[1397,1398,1399,1400],"Large context","More information per request","Latency, cost, attention dilution, leakage surface","Should data be retrieved\u002Ffiltered instead of always injected?",[1402,1403,1404,1405],"Powerful tools \u002F autonomy","More end-to-end automation","Higher privilege and failure blast radius","Which actions require least privilege, confirmation or human approval?",[1407,1408,1409,1410],"Strict validation and logging","Better evidence and operations","Latency, storage, privacy and complexity cost","What evidence is required for this risk level?",{"id":584,"data":1412,"type":42},{"text":1413,"level":242},"How is this different from adjacent roles?",{"id":588,"data":1415,"type":218},{"text":1416},"Titles overlap heavily across companies. The useful distinction is the \u003Cstrong>scope of architecture responsibility\u003C\u002Fstrong>, not the HR label.",{"id":592,"data":1418,"type":299},{"rows":1419,"title":1438,"layout":291,"columns":1439},[1420,1423,1426,1429,1432,1435],{"id":596,"label":1421,"values":1422},"AI Solution Architect",{"role":599,"focus":600},{"id":602,"label":1424,"values":1425},"AI Platform Architect",{"role":605,"focus":606},{"id":608,"label":1427,"values":1428},"Enterprise AI Architect",{"role":611,"focus":612},{"id":614,"label":1430,"values":1431},"AI \u002F ML Engineer",{"role":617,"focus":618},{"id":620,"label":1433,"values":1434},"Security Architect",{"role":623,"focus":624},{"id":626,"label":1436,"values":1437},"Product \u002F Delivery Lead",{"role":629,"focus":630},"Adjacent roles answer different primary questions",[1440,1442],{"id":634,"label":1441},"Role",{"id":637,"label":1443},"Primary architecture focus",{"id":640,"data":1445,"type":218},{"text":1446},"In a small product team, one person may cover several of these scopes. In a large enterprise, they may be separate roles with formal review boards. The architecture responsibility does not disappear when the title changes.",{"id":644,"data":1448,"type":42},{"text":1449,"level":242},"Implementation evidence: how these boundaries appear in my own work",{"id":648,"data":1451,"type":225},{"body":1452,"title":1453,"variant":652},"The examples below are \u003Cstrong>original implementation\u002Fproject evidence\u003C\u002Fstrong>. They show how I have separated product need, requirements, architecture, runtime, model\u002Fprovider, permissions and validation in real project work. They are not claims that every organization must use the same structure, and they do not imply customer adoption or enterprise-scale deployment.","Implementation evidence, not a universal rule",{"id":654,"data":1455,"type":42},{"text":1456,"level":241},"SenseFlow: need → requirements → architecture → validation",{"id":658,"data":1458,"type":218},{"text":1459},"In the SenseFlow project Source of Truth, technology is explicitly subordinate to Product Vision. The development structure moves from problem and product vision through user needs, value, scope, epics, stories and acceptance criteria into architecture, implementation, validation and iteration.",{"id":662,"data":1461,"type":218},{"text":1462},"Requirements are designed to be traceable from Product Goal → Capability → Epic → User Story → Acceptance Criteria → Technical Tasks. Where practical, they include functional requirements, NFRs, dependencies, risks, assumptions, acceptance criteria and validation methods. Significant decisions preserve the decision, reason, alternatives, trade-offs, status and date\u002Fversion.",{"id":666,"data":1464,"type":218},{"text":1465},"That is architectural work before a specific AI framework or model is chosen: it protects the connection between product intent and technical decisions and makes later change reviewable rather than implicit.",{"id":670,"data":1467,"type":42},{"text":1468,"level":241},"Aaasaasa AI Client: separate concepts before integrating them",{"id":674,"data":1470,"type":218},{"text":1471},"Aaasaasa AI Client provides a more implementation-level example. Its AI Hub deliberately separates \u003Cstrong>agent\u002Fclient\u003C\u002Fstrong>, \u003Cstrong>provider\u003C\u002Fstrong>, \u003Cstrong>model\u003C\u002Fstrong>, \u003Cstrong>connection\u002Fruntime location\u003C\u002Fstrong>, \u003Cstrong>permissions\u003C\u002Fstrong> and \u003Cstrong>web client\u003C\u002Fstrong>. A local runtime is not assumed to mean local inference, and permissions are treated as runtime\u002Ftool policy rather than as a property of the model.",{"id":678,"data":1473,"type":218},{"text":1474},"The desktop architecture also defines a trust boundary: the Nuxt renderer is untrusted relative to Electron main. A narrow preload and validated IPC mediate access to AI services, settings, encrypted secrets, workspace\u002Fdata services and runtimes. Cloud credentials remain in the privileged main process; renderer code receives normalized state instead of raw secrets or unrestricted operating-system access.",{"id":682,"data":1476,"type":218},{"text":1477},"Routing decisions are similarly architectural. The implementation does not silently fall back from a local route to paid cloud inference; a cloud route requires explicit confirmation. Direct Chat has no filesystem or shell tools by default, while agent execution applies a selected workspace and permission profile. These are solution-level decisions about trust, cost, execution and user expectation—not model features.",{"id":686,"data":1479,"type":42},{"text":1480,"level":242},"How current architecture frameworks support this broader scope",{"id":690,"data":1482,"type":218},{"text":1483},"ISO\u002FIEC\u002FIEEE 42010:2022 provides a general discipline for architecture descriptions across software, systems and enterprises. It is deliberately broader than AI and does not prescribe one architecting method or job title. That makes it useful here as a boundary: AI solution architecture is still architecture, with stakeholder concerns, multiple views and significant relationships that must be expressed clearly.",{"id":694,"data":1485,"type":218},{"text":1486},"NIST AI RMF 1.0 frames AI risk management through \u003Cstrong>Govern, Map, Measure and Manage\u003C\u002Fstrong> and emphasizes that risk management should be continuous across the AI system lifecycle. The Generative AI Profile (NIST AI 600-1) adapts that framework to GAI risks and organizational priorities. This reinforces that architecture cannot stop at functional model performance.",{"id":698,"data":1488,"type":218},{"text":1489},"Microsoft’s current Azure Well-Architected AI guidance separates application design, application platform, training data, grounding data and data platform concerns and repeatedly connects them to reliability, security, operational excellence, performance and cost. AWS’s Generative AI and Agentic AI lenses similarly treat observability, security, reliability, model\u002Ftool lifecycle, cost and human oversight as architecture concerns.",{"id":702,"data":1491,"type":42},{"text":1492,"level":242},"Common misconceptions",{"id":706,"data":1494,"type":291},{"content":1495,"stretched":43,"withHeadings":14},[1496,1499,1502,1505,1508,1511,1514,1517],[1497,1498],"Misconception","Correction",[1500,1501],"“The architect chooses the LLM.”","Model choice is one decision inside a larger solution architecture.",[1503,1504],"“Prompt engineering is the architecture.”","Prompts affect behavior, but they do not define identity, data access, trust boundaries, deployment, tool permissions or operations.",[1506,1507],"“RAG solves enterprise knowledge.”","Retrieval is only one subsystem; authorization, provenance, freshness, evidence, indexing, evaluation and source governance still need design.",[1509,1510],"“Local runtime means private\u002Flocal AI.”","Runtime, inference, data and control-plane locations are separate architectural properties.",[1512,1513],"“If a vendor offers guardrails, security is covered.”","Security spans identity, authorization, secrets, data flows, tools, logging, deployment, human approval and provider boundaries.",[1515,1516],"“The architect must write every component.”","Hands-on implementation can improve architectural quality, but the role is defined by integrated decision responsibility, not by personally coding every layer.",[1518,1519],"“An architecture diagram proves production readiness.”","Readiness requires implemented controls and validation evidence across quality, security, operations and business acceptance.",{"id":734,"data":1521,"type":42},{"text":1522,"level":242},"Failure modes an AI Solution Architect should prevent",{"id":738,"data":1524,"type":291},{"content":1525,"stretched":43,"withHeadings":14},[1526,1530,1534,1538,1542,1546,1550,1554,1558],[1527,1528,1529],"Failure mode","Why it happens","Architectural correction",[1531,1532,1533],"Model-first design","A promising model demo becomes the system blueprint","Start from outcome, constraints and validation; select the model inside that frame",[1535,1536,1537],"Prototype permissions in production","Shared credentials and broad access survive the PoC","Define identity propagation, least privilege, tool scopes and approval boundaries early",[1539,1540,1541],"Retrieval without authorization","Search quality is designed before data-access rules","Carry user\u002Ftenant context into retrieval and enforce authorization at data-access boundaries",[1543,1544,1545],"Silent provider\u002Fruntime assumptions","“Local”, “cloud” and “offline” are used imprecisely","Document runtime, inference, data and control-plane location separately",[1547,1548,1549],"No failure contract","The happy path is designed but refusal\u002Ffallback\u002Ferror behavior is not","Specify retrieval-empty, model-unavailable, tool-failure and policy-denied behavior",[1551,1552,1553],"Evaluation after implementation","Quality is judged manually near launch","Define measurable acceptance and representative evaluation sets before architecture freezes",[1555,1556,1557],"Untraceable change","Models, prompts, retrieval or permissions change without architectural history","Version critical configuration and record significant decisions\u002Fvalidation evidence",[1559,1560,1561],"Operations treated as infrastructure only","AI behavior is not observable after deployment","Design traces, quality metrics, security events, cost telemetry and rollback together",{"id":778,"data":1563,"type":42},{"text":1564,"level":242},"A practical decision sequence",{"id":782,"data":1566,"type":339},{"steps":1567,"title":1592,"orientation":338},[1568,1571,1574,1577,1580,1583,1586,1589],{"label":1569,"description":1570},"Outcome","Define the user\u002Fbusiness result and explicit non-goals.",{"label":1572,"description":1573},"Evidence and constraints","Identify authoritative data, policies, NFRs, risks and acceptance conditions.",{"label":1575,"description":1576},"System boundary","Map users, identities, applications, data, models\u002Fproviders, tools and external systems.",{"label":1578,"description":1579},"Architecture options","Compare patterns for retrieval, model access, orchestration, deployment, permissions, evaluation and observability.",{"label":1581,"description":1582},"Trade-off decisions","Select significant options and preserve the rationale, alternatives and consequences.",{"label":1584,"description":1585},"Implementation contracts","Turn decisions into APIs, schemas, permission rules, deployment definitions and engineering tasks.",{"label":1587,"description":1588},"Validation","Test the implemented system against the original functional and non-functional requirements.",{"label":1590,"description":1591},"Operational feedback","Use production evidence, incidents, quality metrics and cost\u002Fsecurity signals to trigger controlled change.","AI solution architecture decision sequence",{"id":811,"data":1594,"type":42},{"text":1595,"level":242},"Edge cases and limits of the role",{"id":815,"data":1597,"type":218},{"text":1598},"Some AI products are dominated by model training, scientific experimentation or specialized hardware. In those cases, model\u002Fdata science and ML systems architecture can become much deeper than the solution-level map shown here. The AI Solution Architect still needs integration and operational boundaries, but specialist architecture may own the training platform itself.",{"id":819,"data":1600,"type":218},{"text":1601},"At the other extreme, a simple SaaS integration may not justify a dedicated architect. A senior engineer or technical product lead can carry the same architecture responsibility. The useful test is not the title but whether significant cross-layer decisions are being made deliberately and validated.",{"id":823,"data":1603,"type":218},{"text":1604},"Regulated, sovereign, air-gapped, safety-critical, highly autonomous or multi-tenant systems also shift the center of gravity. Identity, isolation, residency, assurance, update mechanisms, human oversight and auditability may dominate model quality in the architecture.",{"id":827,"data":1606,"type":42},{"text":1607,"level":242},"What would change this answer?",{"id":831,"data":1609,"type":218},{"text":1610},"The exact responsibility boundary changes when architecture moves from one application to a reusable platform or to enterprise-wide target architecture. That is why \u003Cstrong>AI Platform Architect\u003C\u002Fstrong> and \u003Cstrong>Enterprise AI Architecture\u003C\u002Fstrong> deserve separate canonical treatment rather than being merged into this role.",{"id":835,"data":1612,"type":218},{"text":1613},"Technology changes also matter. New model capabilities, protocols, local runtimes and managed services can remove some implementation work while creating new trust or operational boundaries. The stable responsibility is to understand those changes as system changes—not to treat a new framework as a replacement for architecture.",{"id":839,"data":1615,"type":42},{"text":1616,"level":242},"AI Solution Architect checklist",{"id":843,"data":1618,"type":291},{"content":1619,"stretched":43,"withHeadings":14},[1620,1623,1625,1628,1630,1633,1636,1639,1642,1645,1648,1651,1653],[1621,1622],"Check","Question",[1569,1624],"Is the user\u002Fbusiness result and non-goal boundary explicit?",[1626,1627],"Requirements","Are functional requirements, NFRs, constraints and acceptance criteria traceable?",[1150,1629],"Are authoritative sources, provenance, freshness, retention and access rules defined?",[1631,1632],"Retrieval\u002Fcontext","Does authorization reach retrieval and context construction?",[1634,1635],"Model\u002Fprovider","Is model\u002Fprovider selection tied to capabilities and constraints rather than preference?",[1637,1638],"Tools\u002Fagents","Are action boundaries, permissions, approvals and failure behavior explicit?",[1640,1641],"Identity\u002Fsecurity","Are human\u002Fmachine identities, secrets and trust boundaries defined?",[1643,1644],"Runtime","Are runtime, inference, data and control-plane locations distinguished?",[1646,1647],"Evaluation","Is there measurable evidence for quality, security and acceptance?",[1649,1650],"Observability","Can production behavior, failures, cost and security events be investigated?",[1159,1652],"Are significant architecture decisions and replacements traceable?",[1156,1654],"Is ownership for deployment, rollback, incidents and lifecycle clear?",{"id":883,"data":1656,"type":42},{"text":1657,"level":242},"Conclusion",{"id":887,"data":1659,"type":218},{"text":1660},"An AI Solution Architect is the person or architecture function that turns an AI opportunity into a coherent technical system. The key skill is not knowing the most model names; it is connecting product need, requirements, data, application architecture, AI capabilities, security, runtime, delivery and validation without losing the boundaries between them.",{"id":891,"data":1662,"type":218},{"text":1663},"A strong AI solution architecture can therefore be summarized as: \u003Cstrong>define the target → establish requirements and constraints → design the system boundaries → make significant trade-offs explicit → implement through clear contracts → validate against evidence → operate and evolve deliberately.\u003C\u002Fstrong> The model is important. The solution is the product.",{"id":895,"data":1665,"type":895},{"items":1666,"title":1691},[1667,1670,1673,1676,1679,1682,1685,1688],{"id":899,"answer":1668,"question":1669},"An AI Solution Architect translates a business or product need into the architecture of a concrete AI-enabled solution, defining how application logic, data\u002Fretrieval, models, tools, identity, security, runtime, evaluation and operations work together.","What is an AI Solution Architect?",{"id":903,"answer":1671,"question":1672},"No. The roles can overlap, especially in small teams, but an AI engineer is primarily an implementation role while the solution architect owns or coordinates cross-layer architecture decisions and trade-offs for the complete workload.","Is an AI Solution Architect the same as an AI engineer?",{"id":907,"answer":1674,"question":1675},"Not by definition, but hands-on implementation knowledge is highly valuable because AI architecture crosses APIs, data, retrieval, security, runtimes and operational behavior. The role is defined by architecture responsibility, not by writing every component personally.","Does an AI Solution Architect need to code?",{"id":911,"answer":1677,"question":1678},"No. Model selection is one decision. Production architecture also needs data and retrieval boundaries, permissions, tools, provider\u002Fruntime choices, observability, evaluation, reliability, cost and lifecycle design.","Is choosing an LLM the main job?",{"id":915,"answer":1680,"question":1681},"An AI Solution Architect focuses on one concrete solution or workload. An AI Platform Architect focuses on reusable AI capabilities and guardrails that support multiple solutions.","What is the difference between an AI Solution Architect and an AI Platform Architect?",{"id":919,"answer":1683,"question":1684},"The solution architect works at application\u002Fworkload scope. Enterprise AI architecture works across the organizational portfolio, target architecture, governance, shared capabilities, integration principles and strategic constraints.","What is the difference between an AI Solution Architect and an Enterprise AI Architect?",{"id":923,"answer":1686,"question":1687},"They are architectural patterns or subsystems inside a solution when the requirements justify them. RAG addresses retrieval-grounded context; agents add planning\u002Ftool execution and therefore additional identity, permission, orchestration and operational concerns.","Where do RAG and agents fit?",{"id":927,"answer":1689,"question":1690},"Implementation plus validation evidence: functional tests, evaluation results, security\u002Fauthorization tests, performance and reliability measurements, observability, operational rehearsal and acceptance against the original requirements.","What proves that the architecture works?","AI Solution Architect — FAQ",{"id":932,"data":1693,"type":932},{"title":1694,"entries":1695},"Core terms",[1696,1698,1700,1703,1706,1708,1710,1712],{"term":1421,"anchor":937,"definition":1697},"Architecture responsibility for one concrete AI-enabled solution or workload, integrating product requirements with application, data, model, tool, security, runtime and operational design.",{"term":1575,"anchor":940,"definition":1699},"The explicit separation between what belongs to the solution and the users, systems, providers, data sources and environments it interacts with.",{"term":1701,"anchor":944,"definition":1702},"Trust boundary","A point where data, identities or control cross between components with different trust assumptions and therefore require explicit security controls.",{"term":1704,"anchor":948,"definition":1705},"Grounding","Supplying an AI model with relevant external information or evidence so its response can be based on sources beyond model parameters.",{"term":1392,"anchor":951,"definition":1707},"An application boundary that decouples parts of the solution from one model\u002Fprovider interface. Useful when justified by routing, portability or policy needs, but not free of trade-offs.",{"term":1646,"anchor":954,"definition":1709},"Structured measurement of AI workload behavior against defined acceptance criteria, including task quality and relevant safety, security, performance and operational properties.",{"term":1424,"anchor":957,"definition":1711},"Architectural role focused on reusable AI platform capabilities used by multiple solutions rather than the architecture of one workload.",{"term":1713,"anchor":961,"definition":1714},"Enterprise AI Architecture","Organization-level architecture that coordinates AI capabilities, platforms, governance, integration and strategic constraints across a portfolio.",{"id":964,"data":1716,"type":42},{"text":1717,"level":242},"Related canonical knowledge",{"id":968,"data":1719,"type":218},{"text":1720},"This article sits in the AI Architecture Foundations cluster. Its direct foundations are \u003Cstrong>Generative AI Explained: Models, Retrieval, Tools and Applications Are Not the Same Thing\u003C\u002Fstrong> and \u003Cstrong>ADR vs NFR: Architecture Decisions and System Quality Are Not the Same Thing\u003C\u002Fstrong>. Adjacent canonical nodes include \u003Cstrong>Agentic AI Explained\u003C\u002Fstrong>, \u003Cstrong>Source of Truth in AI Systems\u003C\u002Fstrong>, \u003Cstrong>Vector Databases, Embeddings and Reranking\u003C\u002Fstrong>, \u003Cstrong>What Is Context Engineering?\u003C\u002Fstrong>, \u003Cstrong>RBAC vs Tenant Isolation\u003C\u002Fstrong>, \u003Cstrong>AI Platform Architect\u003C\u002Fstrong>, \u003Cstrong>Enterprise AI Architecture\u003C\u002Fstrong> and \u003Cstrong>AI Governance\u003C\u002Fstrong>. URLs are intentionally not fabricated where those nodes are not yet published.",{"id":972,"data":1722,"type":980},{"link":1723,"meta":1724},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works",{"image":1725,"title":1726,"description":1727},{"url":977},"What Is RAG? The Simplest Explanation of How It Works","Existing stajic.de canonical explanation of retrieval-augmented generation, useful for the retrieval\u002Fgrounding part of AI solution architecture.",{"id":982,"data":1729,"type":42},{"text":1730,"level":242},"Primary sources and current architecture guidance",{"id":986,"data":1732,"type":218},{"text":1733},"External sources below support the general architecture claims; the SenseFlow and Aaasaasa AI Client sections are explicitly original project\u002Fimplementation evidence. Current-state references were checked on 8 October 2026. NIST notes that AI RMF 1.0 is being revised, so version-sensitive governance references should be rechecked when a successor is published.",{"id":990,"data":1735,"type":980},{"link":992,"meta":1736},{"image":1737,"title":1738,"description":1739},{"url":977},"ISO\u002FIEC\u002FIEEE 42010:2022 — Architecture Description","Current international standard for the structure and expression of architecture descriptions. It distinguishes architecture from its description and does not prescribe one architecting method, tool or recording format.",{"id":998,"data":1741,"type":980},{"link":1000,"meta":1742},{"image":1743,"title":1744,"description":1745},{"url":977},"NIST AI Risk Management Framework","NIST’s AI RMF resource page. As of October 2026 it states that AI RMF 1.0 is being revised and links the Generative AI Profile and related resources.",{"id":1006,"data":1747,"type":980},{"link":1008,"meta":1748},{"image":1749,"title":1750,"description":1751},{"url":977},"NIST AI RMF Core — Govern, Map, Measure, Manage","Official NIST AIRC presentation of the AI RMF 1.0 Core, including the four functions and lifecycle-oriented risk-management framing.",{"id":1014,"data":1753,"type":980},{"link":1016,"meta":1754},{"image":1755,"title":1756,"description":1757},{"url":977},"NIST AI 600-1 — Generative AI Profile","Cross-sectoral Generative AI profile for AI RMF 1.0, published 26 July 2024 and updated by NIST in 2026.",{"id":1022,"data":1759,"type":980},{"link":1024,"meta":1760},{"image":1761,"title":1762,"description":1763},{"url":977},"Microsoft Azure Well-Architected — AI Workloads","Current workload-level architecture guidance covering AI application design, application platform, training data, grounding data, data platform and production-readiness concerns.",{"id":1030,"data":1765,"type":980},{"link":1032,"meta":1766},{"image":1767,"title":1768,"description":1769},{"url":977},"Microsoft — Application Design for AI Workloads","Guidance on model\u002Ftool abstraction, data-access boundaries, identity propagation, authorization and separation of client, intelligence, knowledge and tool layers.",{"id":1038,"data":1771,"type":980},{"link":1040,"meta":1772},{"image":1773,"title":1774,"description":1775},{"url":977},"Microsoft — Design Principles for AI Workloads","Current AI workload design principles across reliability, security, cost, operational excellence and performance, including identity and data-protection responsibilities.",{"id":1046,"data":1777,"type":980},{"link":1048,"meta":1778},{"image":1779,"title":1780,"description":1781},{"url":977},"Microsoft — MLOps and GenAIOps for AI Workloads","Production lifecycle guidance covering monitoring, quality gates, model\u002Fprompt behavior, security and operational measurement.",{"id":1054,"data":1783,"type":980},{"link":1056,"meta":1784},{"image":1785,"title":1786,"description":1787},{"url":977},"AWS Well-Architected Generative AI Lens","AWS architectural guidance for generative AI workloads across operational excellence, security, reliability, performance efficiency, cost optimization and sustainability.",{"id":1062,"data":1789,"type":980},{"link":1064,"meta":1790},{"image":1791,"title":1792,"description":1793},{"url":977},"AWS Well-Architected Agentic AI Lens","Published in 2026, covering agentic-specific architecture concerns including identities, tools, orchestration, human oversight, reliability, tracing and reasoning-loop cost.","2.31.0","An AI Solution Architect turns business requirements into a production-ready AI system across data, models, tools, security, runtime, evaluation and operations.",{"lang":7,"title":208,"content":210,"contentJson":1797,"excerpt":1070},{"time":212,"blocks":1798,"version":1069},[1799,1801,1803,1805,1807,1809,1811,1813,1815,1831,1833,1835,1837,1847,1849,1851,1853,1855,1857,1871,1873,1875,1877,1879,1881,1883,1885,1887,1889,1891,1893,1895,1897,1899,1901,1903,1905,1907,1909,1911,1913,1915,1917,1929,1931,1933,1944,1946,1948,1966,1968,1970,1972,1974,1976,1978,1980,1982,1984,1986,1988,1990,1992,1994,1996,1998,2009,2011,2023,2025,2036,2038,2040,2042,2044,2046,2048,2050,2052,2068,2070,2072,2074,2085,2096,2098,2100,2104,2106,2108,2112,2116,2120,2124,2128,2132,2136,2140,2144],{"id":215,"data":1800,"type":218},{"text":217},{"id":220,"data":1802,"type":225},{"body":222,"title":223,"variant":224},{"id":227,"data":1804,"type":225},{"body":229,"title":230,"variant":231},{"id":233,"data":1806,"type":225},{"body":235,"title":236,"variant":231},{"id":238,"data":1808,"type":243},{"title":240,"maxLevel":241,"minLevel":242},{"id":245,"data":1810,"type":42},{"text":247,"level":242},{"id":249,"data":1812,"type":218},{"text":251},{"id":253,"data":1814,"type":218},{"text":255},{"id":257,"data":1816,"type":299},{"rows":1817,"title":290,"layout":291,"columns":1828},[1818,1820,1822,1824,1826],{"id":261,"label":262,"values":1819},{"model":264,"solution":265},{"id":267,"label":268,"values":1821},{"model":270,"solution":271},{"id":273,"label":274,"values":1823},{"model":276,"solution":277},{"id":279,"label":280,"values":1825},{"model":282,"solution":283},{"id":285,"label":286,"values":1827},{"model":288,"solution":289},[1829,1830],{"id":294,"label":295},{"id":297,"label":298},{"id":301,"data":1832,"type":42},{"text":303,"level":242},{"id":305,"data":1834,"type":218},{"text":307},{"id":309,"data":1836,"type":218},{"text":311},{"id":313,"data":1838,"type":339},{"steps":1839,"title":337,"orientation":338},[1840,1841,1842,1843,1844,1845,1846],{"label":317,"description":318},{"label":320,"description":321},{"label":323,"description":324},{"label":326,"description":327},{"label":329,"description":330},{"label":332,"description":333},{"label":335,"description":336},{"id":341,"data":1848,"type":42},{"text":343,"level":242},{"id":345,"data":1850,"type":218},{"text":347},{"id":349,"data":1852,"type":218},{"text":351},{"id":353,"data":1854,"type":42},{"text":355,"level":242},{"id":357,"data":1856,"type":218},{"text":359},{"id":361,"data":1858,"type":291},{"content":1859,"stretched":43,"withHeadings":14},[1860,1861,1862,1863,1864,1865,1866,1867,1868,1869,1870],[365,366,367],[369,370,371],[373,374,375],[377,378,379],[381,382,383],[385,386,387],[389,390,391],[393,394,395],[397,398,399],[401,402,403],[405,406,407],{"id":409,"data":1872,"type":42},{"text":411,"level":241},{"id":413,"data":1874,"type":218},{"text":415},{"id":417,"data":1876,"type":218},{"text":419},{"id":421,"data":1878,"type":42},{"text":423,"level":241},{"id":425,"data":1880,"type":218},{"text":427},{"id":429,"data":1882,"type":218},{"text":431},{"id":433,"data":1884,"type":42},{"text":435,"level":241},{"id":437,"data":1886,"type":218},{"text":439},{"id":441,"data":1888,"type":218},{"text":443},{"id":445,"data":1890,"type":42},{"text":447,"level":241},{"id":449,"data":1892,"type":218},{"text":451},{"id":453,"data":1894,"type":218},{"text":455},{"id":457,"data":1896,"type":42},{"text":459,"level":241},{"id":461,"data":1898,"type":218},{"text":463},{"id":465,"data":1900,"type":218},{"text":467},{"id":469,"data":1902,"type":42},{"text":471,"level":241},{"id":473,"data":1904,"type":218},{"text":475},{"id":477,"data":1906,"type":218},{"text":479},{"id":481,"data":1908,"type":42},{"text":483,"level":241},{"id":485,"data":1910,"type":218},{"text":487},{"id":489,"data":1912,"type":218},{"text":491},{"id":493,"data":1914,"type":42},{"text":495,"level":242},{"id":497,"data":1916,"type":218},{"text":499},{"id":501,"data":1918,"type":291},{"content":1919,"stretched":43,"withHeadings":14},[1920,1921,1922,1923,1924,1925,1926,1927,1928],[505,506],[508,509],[511,512],[514,515],[517,518],[520,521],[523,524],[526,527],[529,530],{"id":532,"data":1930,"type":42},{"text":534,"level":242},{"id":536,"data":1932,"type":218},{"text":538},{"id":540,"data":1934,"type":291},{"content":1935,"stretched":43,"withHeadings":14},[1936,1937,1938,1939,1940,1941,1942,1943],[544,545,546,547],[549,550,551,552],[554,555,556,557],[559,560,561,562],[564,565,566,567],[569,570,571,572],[574,575,576,577],[579,580,581,582],{"id":584,"data":1945,"type":42},{"text":586,"level":242},{"id":588,"data":1947,"type":218},{"text":590},{"id":592,"data":1949,"type":299},{"rows":1950,"title":631,"layout":291,"columns":1963},[1951,1953,1955,1957,1959,1961],{"id":596,"label":597,"values":1952},{"role":599,"focus":600},{"id":602,"label":603,"values":1954},{"role":605,"focus":606},{"id":608,"label":609,"values":1956},{"role":611,"focus":612},{"id":614,"label":615,"values":1958},{"role":617,"focus":618},{"id":620,"label":621,"values":1960},{"role":623,"focus":624},{"id":626,"label":627,"values":1962},{"role":629,"focus":630},[1964,1965],{"id":634,"label":635},{"id":637,"label":638},{"id":640,"data":1967,"type":218},{"text":642},{"id":644,"data":1969,"type":42},{"text":646,"level":242},{"id":648,"data":1971,"type":225},{"body":650,"title":651,"variant":652},{"id":654,"data":1973,"type":42},{"text":656,"level":241},{"id":658,"data":1975,"type":218},{"text":660},{"id":662,"data":1977,"type":218},{"text":664},{"id":666,"data":1979,"type":218},{"text":668},{"id":670,"data":1981,"type":42},{"text":672,"level":241},{"id":674,"data":1983,"type":218},{"text":676},{"id":678,"data":1985,"type":218},{"text":680},{"id":682,"data":1987,"type":218},{"text":684},{"id":686,"data":1989,"type":42},{"text":688,"level":242},{"id":690,"data":1991,"type":218},{"text":692},{"id":694,"data":1993,"type":218},{"text":696},{"id":698,"data":1995,"type":218},{"text":700},{"id":702,"data":1997,"type":42},{"text":704,"level":242},{"id":706,"data":1999,"type":291},{"content":2000,"stretched":43,"withHeadings":14},[2001,2002,2003,2004,2005,2006,2007,2008],[710,711],[713,714],[716,717],[719,720],[722,723],[725,726],[728,729],[731,732],{"id":734,"data":2010,"type":42},{"text":736,"level":242},{"id":738,"data":2012,"type":291},{"content":2013,"stretched":43,"withHeadings":14},[2014,2015,2016,2017,2018,2019,2020,2021,2022],[742,743,744],[746,747,748],[750,751,752],[754,755,756],[758,759,760],[762,763,764],[766,767,768],[770,771,772],[774,775,776],{"id":778,"data":2024,"type":42},{"text":780,"level":242},{"id":782,"data":2026,"type":339},{"steps":2027,"title":809,"orientation":338},[2028,2029,2030,2031,2032,2033,2034,2035],{"label":786,"description":787},{"label":789,"description":790},{"label":792,"description":793},{"label":795,"description":796},{"label":798,"description":799},{"label":801,"description":802},{"label":804,"description":805},{"label":807,"description":808},{"id":811,"data":2037,"type":42},{"text":813,"level":242},{"id":815,"data":2039,"type":218},{"text":817},{"id":819,"data":2041,"type":218},{"text":821},{"id":823,"data":2043,"type":218},{"text":825},{"id":827,"data":2045,"type":42},{"text":829,"level":242},{"id":831,"data":2047,"type":218},{"text":833},{"id":835,"data":2049,"type":218},{"text":837},{"id":839,"data":2051,"type":42},{"text":841,"level":242},{"id":843,"data":2053,"type":291},{"content":2054,"stretched":43,"withHeadings":14},[2055,2056,2057,2058,2059,2060,2061,2062,2063,2064,2065,2066,2067],[847,848],[850,851],[853,854],[268,856],[858,859],[861,862],[864,865],[867,868],[870,871],[873,874],[876,877],[286,879],[280,881],{"id":883,"data":2069,"type":42},{"text":885,"level":242},{"id":887,"data":2071,"type":218},{"text":889},{"id":891,"data":2073,"type":218},{"text":893},{"id":895,"data":2075,"type":895},{"items":2076,"title":930},[2077,2078,2079,2080,2081,2082,2083,2084],{"id":899,"answer":900,"question":901},{"id":903,"answer":904,"question":905},{"id":907,"answer":908,"question":909},{"id":911,"answer":912,"question":913},{"id":915,"answer":916,"question":917},{"id":919,"answer":920,"question":921},{"id":923,"answer":924,"question":925},{"id":927,"answer":928,"question":929},{"id":932,"data":2086,"type":932},{"title":934,"entries":2087},[2088,2089,2090,2091,2092,2093,2094,2095],{"term":597,"anchor":937,"definition":938},{"term":792,"anchor":940,"definition":941},{"term":943,"anchor":944,"definition":945},{"term":947,"anchor":948,"definition":949},{"term":564,"anchor":951,"definition":952},{"term":873,"anchor":954,"definition":955},{"term":603,"anchor":957,"definition":958},{"term":960,"anchor":961,"definition":962},{"id":964,"data":2097,"type":42},{"text":966,"level":242},{"id":968,"data":2099,"type":218},{"text":970},{"id":972,"data":2101,"type":980},{"link":974,"meta":2102},{"image":2103,"title":978,"description":979},{"url":977},{"id":982,"data":2105,"type":42},{"text":984,"level":242},{"id":986,"data":2107,"type":218},{"text":988},{"id":990,"data":2109,"type":980},{"link":992,"meta":2110},{"image":2111,"title":995,"description":996},{"url":977},{"id":998,"data":2113,"type":980},{"link":1000,"meta":2114},{"image":2115,"title":1003,"description":1004},{"url":977},{"id":1006,"data":2117,"type":980},{"link":1008,"meta":2118},{"image":2119,"title":1011,"description":1012},{"url":977},{"id":1014,"data":2121,"type":980},{"link":1016,"meta":2122},{"image":2123,"title":1019,"description":1020},{"url":977},{"id":1022,"data":2125,"type":980},{"link":1024,"meta":2126},{"image":2127,"title":1027,"description":1028},{"url":977},{"id":1030,"data":2129,"type":980},{"link":1032,"meta":2130},{"image":2131,"title":1035,"description":1036},{"url":977},{"id":1038,"data":2133,"type":980},{"link":1040,"meta":2134},{"image":2135,"title":1043,"description":1044},{"url":977},{"id":1046,"data":2137,"type":980},{"link":1048,"meta":2138},{"image":2139,"title":1051,"description":1052},{"url":977},{"id":1054,"data":2141,"type":980},{"link":1056,"meta":2142},{"image":2143,"title":1059,"description":1060},{"url":977},{"id":1062,"data":2145,"type":980},{"link":1064,"meta":2146},{"image":2147,"title":1067,"description":1068},{"url":977},"Post erfolgreich abgerufen",{"items":2150,"source":2235,"manualIds":2236,"manualMatchedIds":2237},[2151,2158,2165,2172,2179,2186,2193,2200,2207,2214,2221,2228],{"id":2152,"slug":2153,"title":2154,"excerpt":2155,"featuredImage":2156,"publishedAt":2157},"478","what-is-rag-the-simplest-explanation-of-how-it-works","什么是RAG？对其工作原理的最简单解释","RAG听起来很复杂，但想法很简单：在AI回答之前，它先从知识源查找有用的信息，并将该信息提供给语言模型。本指南使用一个简单的思维模型来解释RAG、LLM、状态、记忆和工具。","\u002Fuploads\u002F2026\u002F09\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works-1790377492124-khjagt.webp","2026-09-25T19:03:00.000Z",{"id":2159,"slug":2160,"title":2161,"excerpt":2162,"featuredImage":2163,"publishedAt":2164},"468","ai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context","AI代理记忆不是RAG：如何区分记忆、检索、状态和上下文","代理记忆、RAG、状态和上下文经常被当作可以互换的概念来使用。它们并不是。这个实用的架构模型将这四个层次区分开来，展示了每一层各自应处的位置，并解释了当系统将它们合并为一层时会出现什么问题。","\u002Fuploads\u002F2026\u002F09\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context-1790350560308-np0xy6.webp","2026-09-25T11:34:00.000Z",{"id":2166,"slug":2167,"title":2168,"excerpt":2169,"featuredImage":2170,"publishedAt":2171},"481","generative-ai-explained-models-retrieval-tools-and-applications-are-not-the-same-thing","生成式人工智能解析：模型、检索、工具与应用并非同一回事","生成式AI不仅仅是一个模型。了解模型、检索、工具、上下文、运行时和应用程序如何在生产AI系统中协同工作。","\u002Fuploads\u002F2026\u002F10\u002Fgenerative-ai-explained-models-retrieval-tools-and-applications-are-not-the-same-thing-1791475411822-pp0dvz.webp","2026-10-08T12:00:00.000Z",{"id":2173,"slug":2174,"title":2175,"excerpt":2176,"featuredImage":2177,"publishedAt":2178},"476","mcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained","MCP vs A2A vs UCP vs AP2 vs A2UI：智能体协议栈详解","MCP、A2A、UCP、AP2 和 A2UI 常被描述为相互竞争的智能体标准。它们大多解决的是不同的互操作性问题。本指南将每个协议映射到其实际标准化的边界，并展示它们如何在同一个生产系统中协同工作。","\u002Fuploads\u002F2026\u002F09\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained-1790352625869-2ezle0.webp","2026-09-25T12:09:00.000Z",{"id":2180,"slug":2181,"title":2182,"excerpt":2183,"featuredImage":2184,"publishedAt":2185},"467","the-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers","答案有效性边界：相关性到可靠AI答案之间缺失的层级","一个来源可能相关、权威，但对于所提出的问题仍然是错误的。缺失的层次是适用性：答案成立的条件，以及迫使其被重新考虑的变化。本文介绍了“答案有效性边界”这一面向人类、AI搜索和RAG系统的来源设计模式。","\u002Fuploads\u002F2026\u002F09\u002Fthe-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers-1790272901306-1g5jly.webp","2026-09-24T11:59:00.000Z",{"id":2187,"slug":2188,"title":2189,"excerpt":2190,"featuredImage":2191,"publishedAt":2192},"381","enterprise-grade-multi-tenant-architecture-for-an-international-platform","企业级多租户架构，适用于国际平台","Loving Rocks 是一款企业级婚礼平台，采用真正的多租户架构设计，实现租户间数据库隔离，并内置国际化支持，以确保全球可扩展性、安全性及长期运营稳定性。","\u002Fuploads\u002F2026\u002F01\u002Fenterprise-grade-multi-tenant-architecture-for-an-international-platform-1769789121298-b6v7ak.webp","2026-01-30T12:04:00.000Z",{"id":2194,"slug":2195,"title":2196,"excerpt":2197,"featuredImage":2198,"publishedAt":2199},"494","air-gapped-ai-how-ai-systems-work-without-internet-or-cloud-access","气隙AI：AI系统如何在没有互联网或云访问的情况下工作","气隙AI在隔离的安全域内运行模型、RAG和AI应用，无需互联网或云依赖。了解模型、数据、更新和工具如何离线运行。","\u002Fuploads\u002F2026\u002F10\u002Fair-gapped-ai-how-ai-systems-work-without-internet-or-cloud-access-1791487983978-e6xqf0.webp","2026-10-08T11:32:00.000Z",{"id":2201,"slug":2202,"title":2203,"excerpt":2204,"featuredImage":2205,"publishedAt":2206},"470","what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","AI代理应该记住、遗忘、重新计算还是再次检索什么？","长时间运行的代理不应记住所有内容。本文提供了一个实用的生命周期模型，用于决定哪些内容应属于持久记忆、哪些内容应重新检索、哪些内容重新计算更安全，以及哪些内容应过期或被取代。","\u002Fuploads\u002F2026\u002F09\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again-1790351131087-iehz28.webp","2026-09-25T09:43:00.000Z",{"id":2208,"slug":2209,"title":2210,"excerpt":2211,"featuredImage":2212,"publishedAt":2213},"479","where-does-an-llm-get-its-data-rag-data-sources-in-python","LLM从哪里获取数据？Python中的RAG数据源","LLM 并不会神奇地知道你的文件、数据库或 API。这个 RAG 系列的实用续篇用简单的 Python 展示了外部数据如何变成可检索的证据：从文本文件和 SQL 到全文搜索、嵌入、上下文组装以及最终的 LLM 调用。","\u002Fuploads\u002F2026\u002F09\u002Fwhere-does-an-llm-get-its-data-rag-data-sources-in-python-1790517200521-nfsi5i.webp","2026-09-27T05:51:00.000Z",{"id":2215,"slug":2216,"title":2217,"excerpt":2218,"featuredImage":2219,"publishedAt":2220},"489","agentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","智能体AI解析：当AI系统能够规划、使用工具并采取行动","代理式AI在多步执行循环中使用模型，这些模型可以在明确的运行时和权限边界内选择工具、观察结果、更新状态并调整其下一步行动。","\u002Fuploads\u002F2026\u002F10\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act-1791481499084-wnji2a.webp","2026-10-08T11:43:00.000Z",{"id":2222,"slug":2223,"title":2224,"excerpt":2225,"featuredImage":2226,"publishedAt":2227},"480","when-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger","人工智能何时应停止信任自身知识？——检索触发机制","AI 模型并非每个问题都需要检索。重要的问题在于知道何时其内部知识已不再足够。检索触发器是一个实用的决策边界，它决定 AI 系统何时应停止仅依赖模型知识，并在回答前获取外部证据。","\u002Fuploads\u002F2026\u002F09\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger-1790574991244-f4rpyg.webp","2026-09-28T01:49:00.000Z",{"id":2229,"slug":2230,"title":2231,"excerpt":2232,"featuredImage":2233,"publishedAt":2234},"471","how-to-know-whether-an-ai-agent-actually-used-the-right-evidence","如何判断一个AI智能体是否真正使用了正确的证据","AI代理可以引用来源，却仍然使用错误的证据。本文介绍一种实用方法，用于核查主张支持、来源权威性、适用性、出处，以及证据是否实际影响了答案。","\u002Fuploads\u002F2026\u002F09\u002Fhow-to-know-whether-an-ai-agent-actually-used-the-right-evidence-1790351317188-o5z9ve.webp","2026-09-25T11:47:00.000Z","fallback",[],[]]