How DataBuddy Uses Ontology and a Semantic Layer to Answer Enterprise AI Semantic Architecture Questions

{"type":"doc","content":[{"type":"heading","attrs":{"id":"b4130bc4-2418-4ff4-afe6-d619bbaa9310","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"一句话里，藏着两个层级的问题"}]},{"type":"paragraph","attrs":{"id":"9d113061-5170-4cce-80e7-0a0b7a3bad51","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"某公司的周会上，运营总监问了 AI 助手一句话："}]},{"type":"paragraph","attrs":{"id":"b0408585-29d1-417b-aacc-484e5498511a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"「上个月华东区客户的 DAU 是多少？」"}]},{"type":"paragraph","attrs":{"id":"9b5a69b1-0689-4e8e-9aff-65556d16ab03","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"一句话，十二个字，听上去没什么刁难的意思，每个词你都觉得不会有歧义。"}]},{"type":"paragraph","attrs":{"id":"44a193b7-77d5-4a07-b70b-1e09fc87c11e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"可它其实踩中了两类完全不同的问题，而这两类问题分属两个不同的层。"}]},{"type":"heading","attrs":{"id":"3d5f3cdb-f9c9-4beb-9652-04c631864833","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"第一个问题：DAU怎么算"}]},{"type":"paragraph","attrs":{"id":"2f9bce20-d701-4d00-8689-91b2e3ca4983","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"AI 答了：12.7 万。"}]},{"type":"paragraph","attrs":{"id":"934ba3ba-6a04-4bc7-91f0-eeacf2289cd4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"数据负责人皱眉：「你这个 DAU，去重到什么粒度？」"}]},{"type":"paragraph","attrs":{"id":"a299f137-a742-43f5-b1e4-477ae1aeda13","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这家公司的 DAU 至少有三种算法同时在用："}]},{"type":"bulletList","attrs":{"id":"f3780214-bb72-44fa-9373-019d9fe55725","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"4566d036-9aad-448a-b303-bef5e6c13645"},"content":[{"type":"paragraph","attrs":{"id":"70bdfd32-364e-4eeb-9139-8afae84c9807","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"按"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"账号"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"去重——一个人两个账号，算两个"}]}]},{"type":"listItem","attrs":{"id":"60b09ba6-c238-4167-b8bf-696dd9ecbace"},"content":[{"type":"paragraph","attrs":{"id":"61d272b3-8167-4c4a-acca-a325336b315d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"按"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"自然人"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"去重——同一身份证下的多个账号合并算一个"}]}]},{"type":"listItem","attrs":{"id":"f7b19fe2-afb1-4722-b292-3dc55a2ac766"},"content":[{"type":"paragraph","attrs":{"id":"17bbdccc-3a18-4c7b-a017-bbd5145cec19","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"按"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"设备"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"去重——一人多端就算多个"}]}]}]},{"type":"paragraph","attrs":{"id":"067c7cc7-4fbd-4614-8a81-934fbeb2a279","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"再往下还有更细的分歧：算「登录即活跃」，还是「有一次有效操作才算活跃」？跨零点的会话归哪一天？「上个月」是自然月还是最近 30 天？"}]},{"type":"paragraph","attrs":{"id":"955d059f-1257-4726-9079-ad0c8074d2b1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"每换一种，数字都不一样。而三种都有正在跑的报表在用。"}]},{"type":"paragraph","attrs":{"id":"a233ce18-4655-49dd-8a7a-8f657f862f3b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"口径就是答案本身。同一个问题，换个算法就是另一个结论。"}]},{"type":"paragraph","attrs":{"id":"bc42b0c3-e731-43dd-bf1a-eac653b2d734","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这类问题问的其实是「这个数到底怎么算出来的」。这是语义模型该管的事：把「DAU」这个词钉成唯一一段计算逻辑，可预览、可验算、能追到血缘。"}]},{"type":"heading","attrs":{"id":"76244981-6ea1-481a-9063-8a73326fef12","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"第二个问题：「客户」指的是谁"}]},{"type":"paragraph","attrs":{"id":"5680954f-0933-4766-ad42-bb922aa9248e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这个问题比上一个更隐蔽，因为它连「算」都还没开始。"}]},{"type":"paragraph","attrs":{"id":"c79128c1-07cd-49c3-bd9e-bb0301e8f95c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"注意那句话里最不起眼的两个字："},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"客户"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"。"}]},{"type":"paragraph","attrs":{"id":"4a8d5ed5-e6b5-49c3-b691-5728b318d54b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"听起来这是个不需要解释的词。但在这家公司内部，它从来没被定清过。"}]},{"type":"paragraph","attrs":{"id":"0b1c0d71-7e92-45d1-b2ed-ef136610c6c3","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"销售部的看法：注册了、下过单、有转化的，才叫客户。"}]},{"type":"paragraph","attrs":{"id":"41a86da5-fc85-401f-acdb-aaa3c3c790a1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"市场部的看法：注册了、天天在用、明显有需求但还没下单的，那也是我们的人。"}]},{"type":"paragraph","attrs":{"id":"f46b22d9-01f3-4df1-90e1-a51f049fcfa2","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这两批人在系统里长得几乎一样，都有账号、都有行为日志、都能算出 DAU。但在业务上，它们是"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"两个不同的对象"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"："}]},{"type":"image","attrs":{"id":"d926a412-19e2-468e-863b-8f2bdd3f4856","src":"https://developer.qcloudimg.com/http-save/audit-1920198/f4d314c70860c4c7672f969242509976.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"0f518f9f-13ab-4c5b-b7f6-81d56e0ba320","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"同一批账号，在业务上其实是客户与商机两个不同的对象。"}]},{"type":"paragraph","attrs":{"id":"0c28798e-0882-4c93-8f5c-8fe7abe98fb4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"现在回头看那个 12.7 万。它里面混着这两类人。"}]},{"type":"paragraph","attrs":{"id":"22c3e658-2698-44b3-aa70-8210f12d51dc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"如果这个数是拿来判断「我们的客户还活跃吗、要不要启动挽留」，那把大量尚未转化的商机算进来，会让留存看起来比实际好得多——一个正在流失的客户群，被一批活跃的潜在用户掩盖掉了。"}]},{"type":"paragraph","attrs":{"id":"8c781951-00a8-4551-b746-1ef0eeba9362","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"反过来，如果这个数是拿来评估「市场投放带来的潜在需求有多少」，那把已成交客户算进来，又会让转化潜力被严重高估。"}]},{"type":"paragraph","attrs":{"id":"a26c3b46-2d83-4589-91bd-b934115c17d5","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"同一个 12.7 万，可以同时支撑两个相反的经营判断。"}]},{"type":"paragraph","attrs":{"id":"d6f91562-cfeb-44e7-87dc-48172cb30314","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"而这并不是算错了。DAU 的计算逻辑可以完全正确，每一行日志也都真实。真正错的是它统计的那群人：他们在业务上本来就该被分成两个对象。"}]},{"type":"paragraph","attrs":{"id":"70558e08-3ab1-46de-824c-e8cffe0dc067","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"口径错了，你得到一个错的数字；对象错了，你得到一个错的世界。"}]},{"type":"heading","attrs":{"id":"a644f450-f220-41b8-b682-6dc672f56033","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"两个问题，两个层级"}]},{"type":"image","attrs":{"id":"20f2796c-b56a-45d5-92aa-7cd20481c84e","src":"https://developer.qcloudimg.com/http-save/audit-1920198/aaf8fcbb2dc8251829a20fd4863eaa2a.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"20630d26-20c6-45b0-b4c7-2a4c8086afb6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"一个错在算法、归语义模型，一个错在分类、只有本体模型能解决。"}]},{"type":"paragraph","attrs":{"id":"f0b0502d-1891-4d65-ad53-781980475716","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"放回那句提问里看，它们的位置是这样的："}]},{"type":"image","attrs":{"id":"dd457590-b3f2-42d0-a7d2-cfd5fa76a739","src":"https://developer.qcloudimg.com/http-save/audit-1920198/dc4c9495d5453f12d834a175cf96747b.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"83f61823-8121-4b0b-9c06-cc7ceab01d1f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"一句话里相邻的两个词，「DAU」归语义模型、「客户」归本体模型，分属两层。"}]},{"type":"paragraph","attrs":{"id":"dda3f23c-291f-4ccd-b7a5-66cf238c95db","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"「DAU 怎么算」，公司里一定有人能说清三种算法的差别。它是那种有人懂、只是没统一的问题，缺的只是把哪一种定为准。"}]},{"type":"paragraph","attrs":{"id":"4bf9201e-8461-4ab6-909f-6d49e5ae67bf","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"「客户指的是谁」不一样，它从来没被正式回答过。销售总监心里有一套答案，市场总监心里有另一套，两套都对，两套都没写进系统。于是 AI 拿到的是一个悬空的词。"}]},{"type":"paragraph","attrs":{"id":"99a2224e-2e25-4ad6-bf2a-521792f87edc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"大多数团队只做了第一件事，把口径统一就以为完事了，结果养出一个「算得越来越准、却始终不知道自己在算谁」的助手。"}]},{"type":"paragraph","attrs":{"id":"02751ded-bd91-480d-901a-b365f38b6b27","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"接下来一个一个解决，先从看起来简单的那个说起。"}]},{"type":"heading","attrs":{"id":"7ce40bfd-f6b3-4147-9c3e-f05d8dfb5ecb","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"先解决第一个：让「DAU」只有一个数"}]},{"type":"paragraph","attrs":{"id":"c7efda2f-63f6-47f7-8072-854ff7aff3a1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"既然有人懂，为什么还是有三个答案？因为这个共识从来没被写在同一个地方。它散落在十几张报表的 SQL 里、几个数据同学各自的习惯里、某次会议的口头约定里，谁要用谁就近拿一个版本。"}]},{"type":"paragraph","attrs":{"id":"d706c89d-63da-4191-9304-1232a05877c1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"语义模型做的，就是把它收拢成唯一一段可执行的定义："}]},{"type":"bulletList","attrs":{"id":"0dd9033f-5d18-4b95-a896-6a775dd067d8","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"f1cbc2b5-70ac-4e10-9ef0-c74c68aadf02"},"content":[{"type":"paragraph","attrs":{"id":"a21a924d-0bb6-4bbf-a5f7-8f6c07c6a27b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"「日活跃客户数」= 按自然人去重 · 有一次有效操作即算活跃 · 自然月口径"}]}]},{"type":"listItem","attrs":{"id":"a9075dda-545e-4d17-845b-e0d5cff875ee"},"content":[{"type":"paragraph","attrs":{"id":"e368c9c3-be32-432e-9bc8-70b1d0dded2d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这段定义只存在一份，谁问都走它"}]}]},{"type":"listItem","attrs":{"id":"dc07c196-fa1c-4542-82e7-349387dbee89"},"content":[{"type":"paragraph","attrs":{"id":"6128c698-dbcf-4451-9eb8-e8f7aa77a963","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"它可以「预览 SQL」——业务方能亲眼看到这个数是怎么算出来的"}]}]},{"type":"listItem","attrs":{"id":"0d424371-d7de-4831-a963-058c3be55bf6"},"content":[{"type":"paragraph","attrs":{"id":"4ec27e23-2095-4934-b0b5-d9428296177f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"它能一路追到底表字段——出了问题知道去哪查"}]}]}]},{"type":"heading","attrs":{"id":"c762c338-28c1-4907-9e96-5f35ba95f12c","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"语义建模具体在做什么"}]},{"type":"paragraph","attrs":{"id":"5ac50f1c-e2fe-4b17-8d61-30b5e86a145b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"「收拢成一段定义」听着还是抽象。落到操作上，语义建模是在建好的 DWD / DIM 明细层之上做两件事。"}]},{"type":"paragraph","attrs":{"id":"076f791c-644f-4bd0-beac-565e367abb94","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"第一步是搭模型。以事实表为中心，把相关维度表按业务关系连成星型或雪花模型，维度自己有层级的就让它分层。这一步固定下两件事：JOIN 路径（哪张表用哪个键连哪张表，写死，不再让 Agent 每次猜），以及可用维度（每张维度表能提供哪些维度，在这里圈定好）。"}]},{"type":"paragraph","attrs":{"id":"9ddec449-a29e-4dd3-ba3a-1150c7ca4b6e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"第二步是定指标。先定不可再拆的原子指标，再往上层层派生衍生指标，加筛选、算同比、做累积、多指标混算都行。指标不是孤立的，它挂在某个模型上，所以天然知道自己能按哪些维度拆、数据从哪张底表来，血缘一路可查。"}]},{"type":"paragraph","attrs":{"id":"9099bcaf-a87e-4cec-b612-b8de70976d30","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这两步走完，「DAU」就不再是十几段各写各的 SQL，也不再是某张 ADS 表里的一列，而是一个挂在明确模型上、口径唯一、可层层追溯的指标定义。它能按任意已声明的维度去拆，不必为每种拆法预先建一张表。"}]},{"type":"paragraph","attrs":{"id":"6f2895fd-251c-4a55-b26d-f43ad13aebd5","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"语义建模不是给字段起中文名，是把「表怎么连、维度有哪些、指标怎么算」一次性定下来。"}]},{"type":"image","attrs":{"id":"ee812130-938e-4bc4-9790-cab86446db70","src":"https://developer.qcloudimg.com/http-save/audit-1920198/4d0c733709c7415cdb73ebc9448c9bd9.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"0ba5d467-717a-4b49-9705-e8c357094bda","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"指标口径页——口径、计算逻辑、维度、血缘与权限一屏可查，并支持「预览 SQL」。"}]},{"type":"paragraph","attrs":{"id":"957c22b7-6157-49df-9a75-d32ab4cd3ad4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"到这里，第一个问题解决得很干净。「DAU」有了唯一去重粒度、唯一活跃判定、唯一时间窗，可预览 SQL、可追到底表字段。任何人、任何 Agent 问 DAU，拿到的都是同一段逻辑算出的同一个数。"}]},{"type":"heading","attrs":{"id":"58d0586b-e3e0-4acd-a998-f3cd2292d353","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"顺带省掉的事：不用再堆 ADS 表"}]},{"type":"paragraph","attrs":{"id":"6cae1626-894a-4d9c-8cd8-0d3c2f110587","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"上面这套建法除了让「DAU」只有一个数，还顺带改变了数仓要建到哪一层。这件事平时藏在语义建模背后，但它对数据团队的影响，可能比「口径统一」还大。"}]},{"type":"paragraph","attrs":{"id":"f1aa2ae4-410d-4a40-b7d6-c0eb7e769321","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"传统链路是明细层（DWD / DIM）、轻度汇总（DWS）、应用层（ADS）一层层往上堆。每来一个新口径、每多一个看数角度，数据团队就得写一条新 ETL、加一张 ADS 表。时间一长，ADS 层膨胀成几百张互相重叠、口径又微妙不同的表，而这些表里大部分只是同一批明细数据的不同切法而已。"}]},{"type":"paragraph","attrs":{"id":"1cd40b62-68ec-462e-bf2a-ad933af094bc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"语义建模换了个思路：只要把 DWD / DIM 这两层建标准、建干净，往上的 DWS 和 ADS 就可以不再用 ETL 去堆。"}]},{"type":"paragraph","attrs":{"id":"9f05ff52-e17c-4ce4-abca-f0f709d882ce","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"因为「怎么切」这件事被挪到了语义层，用声明来表达。模型声明表怎么连、指标声明数怎么算，查询进来时引擎按声明实时编译成 SQL，直接在明细层上算。换个维度看数，不是加一张表，而是换一个 groupby。"}]},{"type":"image","attrs":{"id":"c7b67964-446f-4584-b707-c24f63336463","src":"https://developer.qcloudimg.com/http-save/audit-1920198/c35e91a1af604a57e2714572fe4c7f27.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"d87547cc-559a-4ab1-a830-a407ec3ce13e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"只需建好 DWD / DIM 两层，往上的语义模型与指标全靠声明，不再逐场景堆 DWS / ADS 表。"}]},{"type":"paragraph","attrs":{"id":"1c9df575-93e5-4427-90b2-5693f813d82a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"口径变化不该靠新增一张表来承接。DWD / DIM 建标准，往上的切法交给语义层声明。"}]},{"type":"paragraph","attrs":{"id":"49ea2b71-5713-488e-b298-8182cb9fdfcb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"数据团队的工作重心也跟着变了：以前是不断响应取数需求、不断加 ADS 表，永远在救火；现在是把明细层的标准建好、把语义声明维护准，一次建好就能长期复用。"}]},{"type":"heading","attrs":{"id":"81d85106-39bb-4272-8fc3-b08abc82aa17","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"再解决第二个：先说清「客户」是什么"}]},{"type":"paragraph","attrs":{"id":"62fcc061-f2d0-44e1-b075-26f27c6ea078","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"可是回到最开始那个数，那个 12.7 万，还是错的。 因为它统计的那群人里，依然混着客户和商机。口径修得再干净，也回答不了「这个数算的是谁」。"}]},{"type":"paragraph","attrs":{"id":"f2622c48-4149-4bdc-8835-6ebac85b253f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"行业里大部分语义层方案做到「指标口径统一」就收工了。但那个 12.7 万说明，这只是中途站。"}]},{"type":"paragraph","attrs":{"id":"3485733e-59cc-4b85-b2e2-c1976d1561b8","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"问题出在一个更根本的地方：AI 手里攥着一堆读数，却不知道这些读数属于谁。"}]},{"type":"paragraph","attrs":{"id":"f0af8b88-0670-4e5c-8e7a-03161c09cc4d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"它能把 DAU、留存率、客单价都算得分毫不差，可它不知道业务世界里有「客户」和「商机」两类对象，不知道它们各自该被怎么对待、能对它们做什么。这不是再加几个指标能解决的，读数再多，也说不清读的是谁。"}]},{"type":"paragraph","attrs":{"id":"eb2a1ba8-ff34-461d-b305-2cef031f879a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"要补的不是语义模型的「加强版」，而是一副业务骨架。客户、商机、订单、销售、合同，以及它们之间负责、持有、转化的关系，构成了业务的真实结构；「DAU」「留存率」「转化率」这些指标和维度，"},{"type":"text","marks":[{"type":"bold"}],"text":"只是挂在这副骨架上的元素"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"。"}]},{"type":"paragraph","attrs":{"id":"8ffe614b-5ed3-4013-b572-2cd950064918","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"指标和维度是骨架上的读数，本体是那副骨架本身。读数再准，也不构成一副骨架。"}]},{"type":"paragraph","attrs":{"id":"14a7a688-e1f4-4f85-8160-692d48e5d28b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"还是那个 DAU。一旦「客户」和「商机」被声明成两个对象，事情立刻不一样了。它们不只是分开统计，而是挂着完全不同的骨架："}]},{"type":"image","attrs":{"id":"edab9f6e-49b0-4dd4-8243-d25a2380de65","src":"https://developer.qcloudimg.com/http-save/audit-1920198/604f2df0e83276aea2bc3a1ea1de3e7c.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"09f7b5c6-3214-44ac-84d7-2908267e09fd","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"客户与商机挂着的属性、关系、规则、动作，是两副完全不同的骨架。"}]},{"type":"paragraph","attrs":{"id":"1fdde400-105b-41a4-ae7b-3030832f33f3","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"看最下面那行：同样叫 DAU，挂在客户上问的是「还活跃吗」，挂在商机上问的是「有多少可能成单」。同一个读数，挂在不同对象上，业务含义完全变了。"}]},{"type":"paragraph","attrs":{"id":"7fa735aa-4462-4a16-b057-fd8a11b074db","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"更关键的是它们之间那条「转化」关系。有了它，AI 才能回答一类以前根本答不上来的问题：上个月有多少商机转成了客户？转化后的客户留存如何？哪个渠道来的商机转化最快？这些问题不需要任何新指标，只需要那条关系被声明出来。"}]},{"type":"paragraph","attrs":{"id":"1f855e2d-211c-4d31-abe2-8f43bdb3c3c2","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这也正是本体和「知识图谱」最容易被混淆、又必须分清的地方。图数据库能存关系，但如果所有连线都叫「关联」，Agent 仍然只知道两个对象相连，不知道这种关系会不会影响下一步动作。组成、来源、触发、责任、约束，画出来都是一条线，业务含义却天差地别。本体要的不是「有关系」，而是「关系到底是什么语义」。"}]},{"type":"heading","attrs":{"id":"f131a308-f8b2-4282-9d15-028cb1a316a1","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"骨架由什么构成"}]},{"type":"paragraph","attrs":{"id":"ae8b9256-b5d8-482e-a2b8-ad46e5a48169","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"本体真正做的事，是"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"把散落的业务元素串联成一个完整的业务模型"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"。这副骨架由五个要素搭起来："}]},{"type":"bulletList","attrs":{"id":"cff9abaf-b18b-4008-a0fb-53843c438491","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"d4c8d4c3-9d49-4f33-83c9-5a5a4763e962"},"content":[{"type":"paragraph","attrs":{"id":"5db24a93-2049-44db-8852-46bb2ae3545d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"对象"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"是骨架的节点，比如客户、商机、订单、渠道，每一个都是业务里真实存在、需要被单独辨认的一类东西。"}]}]},{"type":"listItem","attrs":{"id":"6077ba30-0630-4036-9ef2-1540977ce174"},"content":[{"type":"paragraph","attrs":{"id":"ba188410-8114-4b3e-a46d-0c93afd3a63b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"属性"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"是节点的画像，比如客户的累计消费额、商机的意向评分、订单的成交金额。有了它们，一个对象才从「一个名字」变成可被度量的实体。"}]}]},{"type":"listItem","attrs":{"id":"a06c19a5-93a6-487e-a6dc-2840263d1650"},"content":[{"type":"paragraph","attrs":{"id":"be7cf51a-5493-40da-9d14-4ccef302ad50","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"关系"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"是节点之间的连线，比如「商机可以转化为客户」「客户由客户成功负责」「订单归属某个渠道」。连线一旦被声明出语义，Agent 才能沿着它走，而不是在一张宽表里靠相关性猜。"}]}]},{"type":"listItem","attrs":{"id":"d7880226-d4ed-419b-9b23-3204ad53f3d1"},"content":[{"type":"paragraph","attrs":{"id":"9d920c59-0ebc-487b-9bb2-0475a168de47","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"规则"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"是骨架上的约束，比如「累计消费额 > 50 万的客户，流失预警必须同步客户成功团队」，它规定了这副骨架在什么条件下该触发什么。"}]}]},{"type":"listItem","attrs":{"id":"4ae0ac09-6090-4c57-a626-1e1b5291d987"},"content":[{"type":"paragraph","attrs":{"id":"8bbd011f-c07e-4657-9d42-d60cd7262b72","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"动作"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"是骨架上可执行的操作，比如派跟进任务、发起挽留、升级预警，它让骨架不只被「看」，还能被「用」。"}]}]}]},{"type":"paragraph","attrs":{"id":"1b37b084-8dde-4c6e-b71d-2f09a608e8de","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"前三个（对象、属性、关系）划定了 AI 面对的是一个什么样的世界，决定它能不能做对决策；后两个（规则、动作）规定了它能做什么、该怎么做，决定它能不能按决策把事办对。这五个要素缺一个，骨架就塌一块：只有对象没有关系，是一堆孤立的点；只有读数没有规则和动作，AI 永远停在「会算」，走不到「会办」。"}]},{"type":"image","attrs":{"id":"99ac79a4-81d9-4441-bede-dcc7d0c333cf","src":"https://developer.qcloudimg.com/http-save/audit-1920198/27067b62b6eb59b3fa21f585a49aa268.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"80e8b36d-dd24-405b-9ae6-9a39e7de3eec","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"对象、属性、关系、规则、动作五者，共同构成一副完整的业务骨架。"}]},{"type":"heading","attrs":{"id":"7443f082-e7ad-417d-9eb2-7552e1447965","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"关键在于：这一层不需要补任何数据"}]},{"type":"paragraph","attrs":{"id":"50623ee5-95c9-40c7-a0e6-4a9181229394","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"有人会问：这不就是加个字段、洗一遍数据的事吗？"}]},{"type":"paragraph","attrs":{"id":"fff65f3b-fbd1-49f1-8085-59d5084dae8f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"不是。订单表里有没有成交记录，系统从第一天就知道，不需要补采、清洗或人工梳理。"}]},{"type":"paragraph","attrs":{"id":"fd61646b-1aa1-4366-9009-7fc3a84347ed","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"缺的是一句声明："},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"「有转化的叫客户，没转化但活跃的叫商机，这是两个对象；客户归客户成功、商机归销售；商机可以转化为客户。」"}]},{"type":"paragraph","attrs":{"id":"55ffec49-0d62-4b7e-b9b4-2176a3827070","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这句话不产生任何新数据。它做的，是把一个一直存在、却只活在人脑子里的业务结构，变成机器可读的声明。这样一来，每个 Agent、每张报表、每次提问，都按同一副骨架理解业务。"}]},{"type":"paragraph","attrs":{"id":"25b2a071-dadd-4b71-aa41-e1efd4edf935","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"数据齐不等于结构在。本体补的不是数据，是那句从没被写下来的业务声明。"}]},{"type":"paragraph","attrs":{"id":"50c040e9-07ca-4e63-85a5-8a26d799c0a6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这也是本体建模往往比想象中好落地的原因：大多数时候你不缺数据，缺的是把业务共识写下来。"}]},{"type":"image","attrs":{"id":"82b46a39-a48b-4653-a3ae-91759f95be10","src":"https://developer.qcloudimg.com/http-save/audit-1920198/cee2b070c65d91233de5d3f3da1e13c8.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"8b908754-356e-41d9-865d-f8f03f774c8c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"实体图谱按业务域组织，实体与关系一目了然。"}]},{"type":"paragraph","attrs":{"id":"630adbcc-f7ad-44eb-9de4-a7f11e98789c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"骨架一旦立起来，Agent 的工作方式就变了。问「这批商机为什么迟迟不转化」，它可以沿着「商机 → 跟进记录 → 销售 → 所属渠道」这条链走，而不是在一张宽表里做相关性猜测；问「该怎么办」，它能读到「高意向商机 7 天未跟进须升级提醒」这条规则，给出合规的处置。沿着图走因果，跟用提示词猜因果，是两种完全不同的可靠性。这些实体、属性、关系、规则、动作，在 DataBuddy 里都是界面可配、可发布、可下线的资产，不再是散落在文档和口头约定里的经验。"}]},{"type":"heading","attrs":{"id":"3526eb35-e426-4467-94e7-84ae1226a626","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"但骨架上的规则，离不开语义给的刻度"}]},{"type":"paragraph","attrs":{"id":"10cf8029-5990-4ebf-98ae-5e6c09b76b20","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"不过这里还有关键的一环：骨架能立起来，前提是它上面挂着的每一个读数都由语义模型保证唯一。本体划定「算的是谁」，语义模型回答「怎么算」，正是开场那两个问题，现在它们咬在了一起。"}]},{"type":"image","attrs":{"id":"44dc34fd-40f3-4704-8e52-fff6ee4dbbb4","src":"https://developer.qcloudimg.com/http-save/audit-1920198/7696d4810af6d9f854047283a98ad2ff.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"35ffa5c1-eb75-4b84-9fa4-92f78265b57b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"本体负责「算的是谁」，语义负责「怎么算」，缺一边另一边都立不住。"}]},{"type":"paragraph","attrs":{"id":"561b838c-8e11-4e48-b625-c20e8099d4ce","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"最能说明这种依赖的，是规则。本体上有这样一条："}]},{"type":"paragraph","attrs":{"id":"723fe831-c640-4d9d-ab7c-b014c819c4f6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"「累计消费额 > 50 万的客户，流失预警必须同步客户成功团队。」"}]},{"type":"paragraph","attrs":{"id":"cce213fd-f3af-481b-adc6-a2a9fb7c251c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这条规则能不能生效，完全取决于「累计消费额」这个口径是否唯一。含不含税？退货要不要扣减？跨子公司要不要合并？如果这个数有三个版本，规则就有三种触发结果，等于失效。一条会随口径漂移的规则，不是规则，是掷骰子。"}]},{"type":"paragraph","attrs":{"id":"f18d5edd-977b-442c-b61a-2700caa3a9dc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"反过来，正因为语义模型保证了「累计消费额」只有一个版本、可预览 SQL、可追到底表字段，本体上这条规则才是可执行、可审计、能被业务背书的。这也是「统计口径可控、可信」的真正含义：本体上挂的每一个属性、每一条规则的阈值，背后都对应一个口径明确、可验算的语义模型指标。"}]},{"type":"paragraph","attrs":{"id":"ff968014-9509-4504-be37-95972a925dad","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"本体决定 Agent 看到什么样的世界，语义决定这个世界里的每个数字是否可信。"}]},{"type":"paragraph","attrs":{"id":"d4cb1951-0072-4bb7-b702-7f86d77cffd9","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"没有语义的本体，是一副没有刻度的骨架；没有本体的语义，是一堆没有归属的读数。"}]},{"type":"heading","attrs":{"id":"cc0184ed-76ca-4cc0-a453-af744071634d","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"完整形态：三层 + 双翼"}]},{"type":"paragraph","attrs":{"id":"92026088-59af-4b33-9476-b9034c7a2660","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"真正落地时，这两层并不是凭空立着的。"}]},{"type":"paragraph","attrs":{"id":"e14ccf2d-f395-4483-94cc-6a0e27aca84b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"DataBuddy Unity Semantics 的完整形态是这样：纵向三层语义主线，横向两翼治理，最上面是消费它的 Agent。"}]},{"type":"image","attrs":{"id":"d36bb4a0-e915-49ce-b0a4-0ef0926400ea","src":"https://developer.qcloudimg.com/http-save/audit-1920198/17ceb142b3c356f2f70e31c8a147fc0b.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"6c549b89-9b8c-4cbe-8ac9-38e9ff5b1c27","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Unity Semantics 整体架构——三层语义主线纵向递进，质量与安全治理双翼横向贯穿，顶部由 Agent 消费。"}]},{"type":"paragraph","attrs":{"id":"63130d66-b4f7-4b22-ac0a-b47f631a9844","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"纵向三层各自回答前面提过的一个问题：统一 Catalog 让物理表变得可读，语义模型让每个读数唯一（回答「怎么算」），本体模型把这些读数装回业务骨架（回答「算的是谁」）。"}]},{"type":"paragraph","attrs":{"id":"3dfff2df-b423-497e-80bc-5d211b4e6b4f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"从下往上走，每一层回答的问题都在变深："}]},{"type":"bulletList","attrs":{"id":"f15c6810-6b04-4c49-9a18-0439a5bf7e7e","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"aa2430a9-20b4-4e3a-850b-21f760a2e489"},"content":[{"type":"paragraph","attrs":{"id":"6c751667-00c6-47ec-b6c4-fdbf4fb6b730","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"统一 Catalog"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"：元数据、全域血缘。它是另外两层的地基，得先把「这张表这个字段是什么、从哪来」摸清楚，字段来源说不清，上面就没法定口径，更没法挂进骨架。"}]}]},{"type":"listItem","attrs":{"id":"6dc2e547-eeab-4663-83c3-ae0bf68fb67c"},"content":[{"type":"paragraph","attrs":{"id":"337e3b20-653e-4aeb-8efa-55af4491a9da","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"语义模型"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"：指标、维度、Join 关系。它回答「过去有多少」。"}]}]},{"type":"listItem","attrs":{"id":"8472dae6-b81e-42ee-b853-859288f07e11"},"content":[{"type":"paragraph","attrs":{"id":"69550496-2300-45c8-bb44-2430cca235eb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"本体模型"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"：对象、属性、关系、规则、动作。它回答「是什么、为什么、会怎样、该怎么办」。"}]}]}]},{"type":"heading","attrs":{"id":"1885a4e6-90b4-4840-8adf-0e70465b1ea1","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"双翼：质量治理 + 安全治理"}]},{"type":"paragraph","attrs":{"id":"554bb2f7-7285-42b8-8754-9bafbb52616b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"骨架立起来之后，新的风险也跟着来了：语义会腐化，也会失控。定义过时了没人改，就成了「更漂亮的歧义」；而 Agent 一旦从「只会说」变成「能动手」，就可能读到不该读的数据、触发不该触发的动作。能力越往行动端走，治理就越不能事后再补。"}]},{"type":"paragraph","attrs":{"id":"7151a512-a248-4fb8-b0d8-7f14646cdd85","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"所以在 Unity Semantics 里，质量治理防腐化，安全治理防失控，两者跟语义一起生长，横向贯穿全部三层。"}]},{"type":"paragraph","attrs":{"id":"cf47e954-f9f5-4205-ab4e-4ee2f3b7638a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"它们被画成「翼」而不是某一层里的一个模块，是因为每一层都用得上：Catalog 层要给字段打安全分级、标出加密项，语义层要保证口径定义不过时，本体层要约束哪些动作能被 Agent 执行、由谁执行。权限在查询编译期就生效，模型根本看不到它无权看的数据，也就谈不上「不小心说漏」。"}]},{"type":"paragraph","attrs":{"id":"94c6c2d5-d513-4426-96bf-70e60181603b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"缺语义 → 幻觉；缺治理 → 腐化 / 失控。两者结合，才是护城河。"}]},{"type":"heading","attrs":{"id":"fbecf3de-8563-401b-a012-044b6f7f7eec","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"为什么这套东西可信"}]},{"type":"paragraph","attrs":{"id":"9ba27014-1213-4a51-a6e2-bbc1a94ae95c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"讲到这里，一个务实的追问会浮上来：当 Agent 开始派单、改状态、写回系统，凭什么相信它不会出错，又凭什么信它每一步都说得清依据？"}]},{"type":"paragraph","attrs":{"id":"0d2a5156-f186-4948-a6c9-18083f468466","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这个问题分三层来答：执行是确定的，结果可溯源，准确率经过实测。"}]},{"type":"heading","attrs":{"id":"519ab678-bb3b-473a-9e16-8b5683075269","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"第一层：语义查询到物理 SQL，全程无 LLM"}]},{"type":"paragraph","attrs":{"id":"0a15fe7a-d1a7-4c19-9c09-3f14023dd92a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这是最容易被忽略、却最决定可信度的一环。很多人以为「AI 取数」是模型一路生成 SQL，在 DataBuddy 里并不是。"}]},{"type":"paragraph","attrs":{"id":"06e82c25-565a-4f3b-a397-ffdc3a089bdb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"模型（Agent）只负责读懂问题、检索语义层、匹配出该用哪个指标、哪个语义模型、哪些维度，然后调用指标查询引擎。从语义查询翻译到物理 SQL 的全过程，都发生在这个引擎内部，没有一步经过 LLM。"}]},{"type":"image","attrs":{"id":"1d1a9088-02ed-45c8-bbd0-618e988adde7","src":"https://developer.qcloudimg.com/http-save/audit-1920198/765f6ff8c3aca19bbba72257a7f91cd9.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"92efaecb-b9a1-4346-9b64-8f3ff1677895","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Agent 只负责检索匹配语义，翻译成 SQL 交给引擎内的确定性流水线，六段无一步经过大模型。"}]},{"type":"paragraph","attrs":{"id":"5b595a86-d886-48fb-8fca-7148c27fcc3e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"不堆 ADS 表，性能靠什么撑？答案也在这里：物化由引擎按代价自动决定，而不是人按场景手工建表。同一份语义声明，引擎判断走聚合表更划算就走聚合表，否则直接算明细。口径只有一份，加速是引擎的事。"}]},{"type":"paragraph","attrs":{"id":"a47ebed1-7a04-4e7b-a34b-0e047d937680","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"降低幻觉靠的是这套结构，不是提示词技巧。同一个语义查询，任何时候执行都得到同一个结果。"}]},{"type":"heading","attrs":{"id":"3b6725c5-b500-4aef-95e0-086b41cd9f9f","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"第二层：每个数都能被追问「凭什么」"}]},{"type":"paragraph","attrs":{"id":"c116d5f3-8f72-4d16-925e-49f24c20d4ae","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"语义资产建好之后，它不属于某一个应用，而是所有消费者的共同底座。分析 Agent、AI 看板、业务系统，走的都是同一套定义。"}]},{"type":"image","attrs":{"id":"db8d554b-da0f-4364-93e6-07eabd11b560","src":"https://developer.qcloudimg.com/http-save/audit-1920198/87a69de267b17021ec1d638407303ae2.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"c0af6195-3221-47c3-828a-a204d91ab5d4","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Agent 直接向语义层取数，每次查询都带来源与 QueryId，可回放、可审计。"}]},{"type":"image","attrs":{"id":"69c0113a-b513-4ffd-a05f-4afd75e28ed7","src":"https://developer.qcloudimg.com/http-save/audit-1920198/0496d911a0c1c9439f8c9666f3ebf099.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"f8b6bb8d-3462-46c9-847a-83607b13070a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"AI 生成的看板，每个指标都能展开溯源、同样受治理。"}]},{"type":"paragraph","attrs":{"id":"ad249ef6-db4a-4250-8eda-be03db6741fd","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"不只是数。当 Agent 开始派单、改状态、写回系统，每一次动作也要留痕：命中了哪条规则、用的哪个口径、谁有权执行。业界主流语义层大多只做到指标口径这一层，规则语义只是局部覆盖，动作语义普遍缺失，这正是「算得出 DAU、却推不动任何一次跟进」的根源。要让 Agent 真动手，数和动作都得说得清依据。"}]},{"type":"paragraph","attrs":{"id":"0924100b-d749-43ea-9ae1-24a832825e69","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"价值不在于「AI 给了你一个数」，而在于它同时告诉你：这个数怎么来的、用的是谁拍板的口径、执行了什么 SQL。"}]},{"type":"heading","attrs":{"id":"8c15ec19-dce5-4156-b8a2-24a3e8403c18","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"rgb(0, 128, 255)","background":""}},{"type":"bold"}],"text":"第三层：准确率 95.9%，而这不是靠更强的模型换来的"}]},{"type":"paragraph","attrs":{"id":"b56fc7a0-ed7c-48b5-a5f9-c7b5c34a13c6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"前面两层讲的是机制，这一层给结果。"}]},{"type":"paragraph","attrs":{"id":"7f39ebd7-c22e-4801-a2ba-ee48161b0b91","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"我们做过一组内部评测：同一批业务问题、同一套底层数据，分两条路径跑。"}]},{"type":"image","attrs":{"id":"a82a0a87-3042-449a-85f8-ed53e724c779","src":"https://developer.qcloudimg.com/http-save/audit-1920198/6cc5de9d5dc546500d844ef72068a061.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"766c3ad8-8968-44f2-b716-382818617952","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"DataBuddy 内部评测——走语义层准确率 95.9%，裸 NL2SQL 仅 83.5%。"}]},{"type":"paragraph","attrs":{"id":"ffc98fba-c0b7-4b0e-aa57-bc1ab8cf028b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"走语义层的那条，准确率 95.9%。"}]},{"type":"paragraph","attrs":{"id":"6cdb12a5-df58-4c2c-b24a-126bbcae7760","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这个数字的意义在于，它跨过了「业务愿不愿意用」的门槛。到了 95% 以上，业务方才敢把 AI 给的数直接拿去做判断；不到这个线，每个答案都得人工复核一遍，等于没省事。"}]},{"type":"paragraph","attrs":{"id":"e0e5f74e-f757-4e12-9a0e-6602a855443f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"另一条路径是不接语义层的裸 NL2SQL。这不是一个随便搭的对照组：我们在它身上把该做的算法优化都做了，schema 召回、few-shot 示例、SQL 自校验和重试全上。它最终跑到 83.5%，这个成绩其实已经高于业界公开的 NL2SQL 平均水平。"}]},{"type":"paragraph","attrs":{"id":"895744d2-0c5d-4bdc-8746-650eed5605df","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"但 83.5% 换成业务语言就是：每 6 个问题里有 1 个是错的，而且你不知道是哪一个。"}]},{"type":"paragraph","attrs":{"id":"2c050b27-3fe6-42e5-8895-9fc86cca1040","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"两条路径差 12.4 个百分点，恰好横跨了「能演示」和「能上生产」之间那道线。"}]},{"type":"paragraph","attrs":{"id":"7a0c1f32-5b14-4145-bb27-ff4a2207e23a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"真正值得琢磨的是这 12.4 个百分点从哪来。不是换了更大的模型，也不是调了更好的提示词，两条路径用的是同一个模型。差别只在于其中一条路径上，表怎么连、维度有哪些、指标怎么算，都已经写下来了。"}]},{"type":"paragraph","attrs":{"id":"5ef9329c-41c6-48f4-a28e-e3452908a64b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"准确率的天花板不在模型那一侧，而在你有没有把业务定义交给它。"}]},{"type":"heading","attrs":{"id":"259d9f97-005a-42ef-9d49-f5607011929d","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"人机协同：这些模型不用人肉写"}]},{"type":"paragraph","attrs":{"id":"fc15b3a5-d45d-4f44-a622-aef63c57cdab","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"讲到这里，一个很现实的顾虑会浮上来：听起来都对，但这不又是个「建三年都建不完」的工程吗？"}]},{"type":"paragraph","attrs":{"id":"54525127-2725-41cf-8ad7-bd117ad791c0","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这个顾虑没错。语义层和本体在业内被讨论了很多年，落不下去的主要原因从来不是价值不清楚，而是建设成本太高——得把散落在几百张表、十几个部门、若干位老员工脑子里的业务共识，一条一条录进系统。指望人从零手写，这事永远排不到优先级。"}]},{"type":"paragraph","attrs":{"id":"417a8b50-ea4f-41ad-a7be-dde2ec178a8e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"所以我们从一开始就把它设计成人机协同：在 DataBuddy 里，语义模型和本体模型的构建，都可以用自然语言交给 Agent 起草。"}]},{"type":"image","attrs":{"id":"a48e6cb3-85b5-4c85-96c2-3b059222d99a","src":"https://developer.qcloudimg.com/http-save/audit-1920198/8168f9652e4455cbbe6b05642857a40b.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"7c39c5ad-ac72-4389-84a2-097f4882cb20","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"人机协同建模四步闭环——人提需求、Agent 起草、人审改拍板、自动落地。"}]},{"type":"paragraph","attrs":{"id":"e5eef8a4-3f05-450e-a462-08ffcd01b24b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"落到产品里，这条闭环有两个入口。一个是 AI 智能建模，用一句业务描述，让 Agent 直接起草本体和语义模型；另一个是导入历史资产，把存量 SQL、老文件传上去，让 Agent 自动解析成结构化草案。两者指向同一件事：冷启动不再是从零手写，而是从一份草稿改起。"}]},{"type":"image","attrs":{"id":"37a82c79-c811-4bb9-9e04-4ed1c8e7bcc6","src":"https://developer.qcloudimg.com/http-save/audit-1920198/7ab4e674d980b29695ec9b1c0d3d1edb.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"bd5f9a0e-f212-4826-82f9-567a73b6abcb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"语义工作台——管理各个业务域的实体、指标与模型，可直接被 Agent 消费。"}]},{"type":"paragraph","attrs":{"id":"841a40cf-e888-4434-b914-87d1e194e182","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"但 Agent 起草的方案不会直接生效。它把读到的元数据、历史 SQL，连同人给的业务需求，整理成一份完整的语义建模和本体建模草案：事实表和维度表怎么 Join、建哪些指标、建哪些实体、实体有哪些属性、实体之间是什么关系。然后它会停下来，等人审。这一步是我们刻意不做成「全自动」的地方——能从数据里推断的交给 Agent，只有业务才能拍板的留给人。"}]},{"type":"image","attrs":{"id":"5d7621c8-6f78-4511-8026-3be558df360b","src":"https://developer.qcloudimg.com/http-save/audit-1920198/95cd36309e25d16847dfb127d9a9bc23.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"22899632-dc3a-46da-a234-2fde8dbd3e31","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Agent 起草的建模方案逐项可改，确认后再一键落地，拍板权始终在业务方手里。"}]},{"type":"heading","attrs":{"id":"cc6a0a92-00f8-47f4-a8a4-9c60b87e4f53","textAlign":"inherit","indent":0,"level":1,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"最后：那句话，现在能被答对了"}]},{"type":"paragraph","attrs":{"id":"a4e2669a-21d1-4612-a02f-f242cdb8c4df","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"回到开篇那个问题——「上个月华东区客户的 DAU 是多少？」"}]},{"type":"paragraph","attrs":{"id":"0bf91265-5486-4408-bba2-294bc33f6e85","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"现在再问一次，链路是这样的："}]},{"type":"paragraph","attrs":{"id":"3eea58f8-0438-4f93-9f4e-6d1054a16b64","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"本体先回答「算的是谁」：「客户」是一个已声明的对象，判据是有成交转化，跟「商机」严格区分，「华东区」按客户签约主体所在地归属。语义模型接着回答「怎么算」：DAU 按自然人去重、有效操作即活跃、自然月口径，一段唯一逻辑，可预览 SQL，可追到底表字段。"}]},{"type":"paragraph","attrs":{"id":"a5f41c36-ab84-4b3e-9c7f-bb2771226d62","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"于是 AI 给出的不再是一个孤立的 12.7 万，而是客户 DAU 8.3 万（留存该看这个数）和商机 DAU 4.4 万（转化潜力该看这个数），还顺手告诉你上月有多少商机转成了客户。"}]},{"type":"paragraph","attrs":{"id":"8902d319-ec12-4063-8dae-c7782ba6456e","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"再往下走一步。因为这两个对象各自挂着规则和动作，Agent 可以对客户侧活跃度下滑的年框大客户，按规则同步客户成功团队；对商机侧高意向、却超 7 天没跟进的，直接派发跟进任务写回 CRM，全程受权限约束、留下可审计记录。从「答一个数」到「据此办一件事」，正是从取数工具走到决策中枢的那一步。"}]},{"type":"image","attrs":{"id":"8a144911-baa0-456f-b745-f0bacbe0e604","src":"https://developer.qcloudimg.com/http-save/audit-1920198/f3a195666dc35c82f6725b443b20d751.webp","extension":"","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":"","aspectRatio":0,"status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"e5b8b1a5-d7b2-4eab-9389-419ad59b194b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"本体先答「算的是谁」，语义再答「怎么算」，两层合力把一句话落成一次受治理的业务动作。"}]},{"type":"paragraph","attrs":{"id":"8687217e-2727-4671-818d-acc19822db80","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"一句话之所以能被答对、还能被接着执行，是因为问题被拆成了两层，而这两层都被分别补上了：语义模型管住口径，本体模型定住对象。开篇那句话反过来说同样成立：口径对了、对象也对了，你才第一次拿到一个既准确、又问对了人的答案。"}]},{"type":"paragraph","attrs":{"id":"d01ea773-ab1c-487e-ae93-f44620ff3253","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这件事的价值，其实比「让 AI 答得更准」要大得多。"}]},{"type":"paragraph","attrs":{"id":"66cbfef0-b788-45cb-ad4e-08fd5e93dd60","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"当企业把对象、关系、规则、动作显式地写下来，它做的是一件过去从没做过的事：把那副只存在于资深员工经验、制度文件、部门口头默契里的业务骨架，第一次变成机器可读的资产。过去这副骨架每次人员流动都会损耗一点，每换一套系统就得重新口口相传一遍。"}]},{"type":"paragraph","attrs":{"id":"3bd7751b-7042-40d6-9f08-d44172da6220","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"而一旦写下来，模型可以换、Agent 可以换、前端可以换，这副骨架都还在。"}]},{"type":"paragraph","attrs":{"id":"d5eed956-16c2-4885-9784-9de92402a364","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"企业 AI 的竞争，最终拼的不是谁的模型更强。护城河不是买来的工具，是谁先把自己的业务骨架写了下来。"}]},{"type":"paragraph","attrs":{"id":"a68b5805-2e63-413d-9ecc-fe250484e49b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"从统一 Catalog 到语义模型再到本体模型，让数据可理解、可推理、可治理，DataBuddy用「语义 × 治理」为企业数据筑起 AI 护城河。"}]}]}