面向中国客户的站点 AI 翻译 全球站点
The AI Hiveby IT Labs India
产品

The AI Hive 如何工作

一个受治理的共享记忆,供您的员工及其 AI 助手读取和写入,每一步都会根据谁在请求、在哪里工作以及被授予了什么权限进行检查。

1. 范围(Scopes)

七个范围,从全公司到私人

通俗地说:全公司,然后是项目,然后是客户,然后是团队,然后是个人,然后是个人的 AI 应用,最后是私人笔记。每一条知识只存放在一个位置,并由所有适用的人继承。

图 1 · 记忆范围和流向七个范围
Memory scopes and the three flows Memory is organised from organisation at the top, through project and client, team, user and agent. Personal memory sits apart and is private. Up arrows show promotion with approval, down arrows show inheritance, and the across arrow shows team knowledge being synthesised into project knowledge. Organisationcompany-wideProjecta programmeClientkept apartTeama group of peopleUserone personAgentthat person's AI appPersonalprivate notes, never promotedDOWN · INHERITUP · PROMOTEACROSSdefault offgovernance approvesproject lead approves3 people confirm Memory scopes and the three flows Memory is organised from organisation at the top, through project and client, team, user and agent. Personal memory sits apart and is private. Up arrows show promotion with approval, down arrows show inheritance, and the across arrow shows team knowledge being synthesised into project knowledge. Organisationcompany-wideProjecta programmeClientkept apartTeama group of peopleUserone personAgentthat person's AI appPersonalprivate notes, never promotedDOWN · INHERITUP · PROMOTEUP · governance approvesACROSSdefault offUP · project leadapprovesUP · 3 people confirm
记忆范围与三种流转:经批准向上、通过继承向下、经批准的综合横向流转。
范围这里有什么
组织全公司范围的知识:标准、政策、共享定义。
项目某个工作项目的知识,由参与该项目的所有人共享。
客户关于一个客户的一切,与其他所有客户隔离开来。
团队一群人共同学到并确认的内容。
用户个人自己的工作记忆。
智能体那位员工的 AI 应用的记忆,代表其工作。
个人私人笔记。仅本人可见,绝不会被提升。

现已可用 全部七个范围均已构建,并在我们自己的模拟环境中验证过。

2. 流程

三条受治理的流程在不同范围之间传递知识

向上:提升

经验证的见解只有经批准才会上移。

  • 当至少 3 人独立确认后进入团队级
  • 经项目负责人批准进入项目级
  • 经治理流程进入组织级

个人记忆是私密的,绝不会被提升。

向下:继承

当更高层级的范围发生变化时,每位受影响的人的 AI 都会收到更新。新成员立即继承背景信息,而不必花数周时间接受培训。

跨范围:综合

团队知识可以合并为项目知识。该功能默认关闭,需由管理员开启,并且始终需要审批。

3. 一个实际案例

从私人笔记到组织标准

一个示例场景,展示一条有用的见解如何在各个范围之间流转。其中的名称和细节均为虚构,仅用于说明。

  1. 一则个人说明

    一位工程师记下一条私人笔记:某家供应商的图纸在发布前需要进行特定的版本检查。这条笔记存放在她的“个人”范围内,其他人看不到,也绝不会自行被晋升。

  2. 她的 AI 应用提出建议

    会话结束时,她的 AI 应用会起草一份会话摘要。她阅读并确认之后,才会保存任何内容。这条有用的见解会作为给团队的拟议事实提出,而不是直接写入。

  3. 团队级:获得 3 次确认后

    当同事遇到同样的问题并各自独立确认时,该事实会达到至少 3 人的阈值。随后它成为团队记忆的一部分,团队中每位成员的 AI 都会继承它。

  4. 项目:由负责人批准

    项目负责人审核该事实并批准其用于整个项目。现在参与该项目的所有人都会从它开始,包括下个月加入的人。

  5. 组织:由治理决定

    如果这一见解值得成为公司标准,治理方会在组织层级批准它。每个早期版本都会保留,审计日志会记录谁在何时批准了什么。

4. 上下文包

为每个 AI 任务提供良好的起点

当 AI 应用开始一项任务时,它向 The AI Hive 请求一个上下文包:为它在组织中的位置继承的记忆,即它的组织、项目、客户、团队和用户。该包附带清单,因此应用以后可以低成本地刷新,而不必重新获取全部内容。

结果是减少重复交代背景。返回的记忆被标注为数据,看起来像指令的文本会被标出,因此检索到的记忆不会被误当成命令。

现已可用

每个结果背后

  • 混合搜索:语义加关键词,每条结果都会按权限重新检查
  • 客户隔离墙:只有在处理该客户或与其关联的项目时,才会返回该客户的记忆
  • 受众标签:internal(默认)、restricted(仅限指定人员,机器人不读取)和 external_ok(可安全用于对外回复,由负责人设置)
  • 版本管理:每次更新都会保留上一个版本
5. 空间

无论人在哪里工作,都会加载正确的记忆

人们在多个 AI 应用中,于代码仓库、云盘、聊天项目和本地文件夹里工作。The AI Hive 把这些空间逐一映射到公司的各个位置,让 AI 应用知道自己是在为哪个项目、团队或客户工作。

图 2 · 公司位置与个人空间示例
Company places and a person's spaces, linked in a space graph On the left, the organisation's places (org, project, team) and the company's declared spaces: a repository, a Jira project, a shared drive and a team chat channel. On the right, the spaces reported by one person's own tools. A solid line shows that her clone is the same space as the company repository; dashed lines link her own repository, drive folder and chat project to the Shipping project; her local scratch folder has no line and stays private. Organisation (places)Org: Example CoProject: ShippingTeam: PlatformCompany spacesRepository shipping-apiJira project SHIPShared drive "Shipping docs"Team chat #shippingDeclared by the company andlinked to places by mapping rulesPerson: AliceA coding assistant on two PCs,a chat app on her phone anda review bot she sponsorsHer spacesClone of shipping-apiHer repo ship-toolsDrive folder "ship notes"Chat project "Shipping"Local folder ~/scratchReported by her own tools,never guessedSpace graph• one address per space, many aliases• owner: company, person or client• every link has evidence behind itsame space (her clone = company repo)linked to a placeno line: stays private Company places and a person's spaces, linked in a space graph On the left, the organisation's places (org, project, team) and the company's declared spaces: a repository, a Jira project, a shared drive and a team chat channel. On the right, the spaces reported by one person's own tools. A solid line shows that her clone is the same space as the company repository; dashed lines link her own repository, drive folder and chat project to the Shipping project; her local scratch folder has no line and stays private. Organisation (places)Org: Example CoProject: ShippingTeam: PlatformCompany spacesRepository shipping-apiJira project SHIPShared drive "Shipping docs"Team chat #shippingDeclared by the company and linked toplaces by mapping rulesSpace graph• one address per space, many aliases• owner: company, person or client• every link has evidence behind itsame space (her clone = company repo)linked to a placeno line: stays privatePerson: AliceA coding assistant on two PCs,a chat app on her phone anda review bot she sponsorsHer spacesClone of shipping-api= same spaceHer repo ship-tools→ ShippingDrive folder "ship notes"→ ShippingChat project "Shipping"→ ShippingLocal folder ~/scratchno line · privateReported by her own tools, never guessed
示例。左边是公司的位置和空间,右边是一个人的空间。The AI Hive 知道她的克隆是公司仓库,而她自己的仓库、云盘文件夹和聊天项目属于 Shipping 项目。她的草稿文件夹保持私有。

明确声明,而非猜测

公司声明其空间和映射规则。个人自己的工具会报告其所在空间;The AI Hive 绝不猜测某人在哪里工作。

每个关联都有证据

空间与实际位置之间的关联要么明确声明,要么连同依据一并建议,建议须经确认后才生效。

私密内容始终私密

没有关联的空间(例如临时文件夹)保留在个人自己的记忆中,绝不会被共享。

测试中空间登记表和空间映射已构建并在测试中。

6. 索引

一个问题如何找到答案

The AI Hive 为它所知的每一类事物各保存一个可搜索的索引:谁和在哪里,以及是什么。每个索引都按关键词和语义搜索,每个条目都带有其受众,因此结果只包含提问者可以看到的内容。

图 3 · 索引以及问题如何得到回答示例
Indexes, the link graph, and how a question is answered Catalog indexes for people, agents and bots, places, spaces and topics on the left; memory, knowledge and source indexes on the right; a link graph in the middle; short index cards loaded at session start; and the five steps of answering a question: question, who and where, filter by permissions, search, and an answer that links to its source. Index cards, loaded at session startA short card for the person and their current place, sent with the space manifest. Details are fetched only when needed.Catalog indexes: who and wherePeoplenames, e-mails, accounts, ask-me-aboutAgents · task bots · workerskind, master or sponsor, purpose, placesPlacesorg, projects, teams, clients, membersSpacesspace IDs, aliases, owner side, linksTopics and glossarysystems, customers, terms, codenamesContent indexes: whatMemory indexitems from people, agents, botsand workers, stamped withtheir memory addressKnowledge indexdecisions, actions, facts,procedures, digests, who to ask(from space content)Source indexcleaned mail, chat, documentand issue text, with access rules,for finding the originalLink graphperson member of team / projectagent / bot / worker works for personspace linked to placeitem created by actoritem created in spaceitem / record counts toward placerecord about topicrecord evidence from sourceEvery index: keyword + meaning search,every entry carries its audience,results filtered to what you may seeAnswering a question1 Question“What did we decideabout canary deploys?”from her coding app,in shipping-api2 Who and wherecatalog indexes:“canary” = topic“we” = Alice’s placesShipping = 3 spaces3 Filteronly places andspaces Alice maysee; client walls;groups of 3 or more4 Searchmemory + knowledgeindexes, keywordand meaning, insidethat scope5 Answerfollow links to thesource: “blue-green,5-min canary — 3 Oct,#shipping (Alice);7 people agree” Indexes, the link graph, and how a question is answered Catalog indexes for people, agents and bots, places, spaces and topics on the left; memory, knowledge and source indexes on the right; a link graph in the middle; short index cards loaded at session start; and the five steps of answering a question: question, who and where, filter by permissions, search, and an answer that links to its source. Index cards, loaded at session startA short card for the person and theircurrent place, sent with the space manifest.Details are fetched only when needed.Catalog indexes: who and wherePeoplenames, e-mails, accounts, ask-me-aboutAgents · task bots · workerskind, master or sponsor, purpose, placesPlacesorg, projects, teams, clients, membersSpacesspace IDs, aliases, owner side, linksTopics and glossarysystems, customers, terms, codenamesLink graphperson member of team / projectagent / bot / worker works for personspace linked to placeitem created by actoritem created in spaceitem / record counts toward placerecord about topicrecord evidence from sourceEvery index: keyword + meaning search,every entry carries its audience,results filtered to what you may seeContent indexes: whatMemory indexitems from people, agents, botsand workers, stamped withtheir memory addressKnowledge indexdecisions, actions, facts,procedures, digests, who to ask(from space content)Source indexcleaned mail, chat, documentand issue text, with access rules,for finding the originalAnswering a question1 Question“What did we decide about canary deploys?”from her coding app, in shipping-api2 Who and wherecatalog indexes: “canary” = topic,“we” = Alice’s places,Shipping = 3 spaces3 Filteronly places and spaces Alice may see;client walls; groups of 3 or more4 Searchmemory + knowledge indexes, keywordand meaning, inside that scope5 Answerfollow links to the source:“blue-green, 5-min canary — 3 Oct,#shipping (Alice); 7 people agree”
示例。目录索引(谁和在哪里)与内容索引(是什么)由关联图连在一起。问题被解析为位置和主题,过滤为此人有权看到的内容,经搜索后给出带有来源链接的答案。

测试中索引卡片以及带路由的采集功能已构建并在测试中。

已构建并在测试中。未通过测试的内容将被移除。

7. 连接器

设计上只读

The AI Hive 从您已在使用的系统中学习,但连接器只读取这些系统。它只会把数据发送到您设定的地方(webhook、邮件、推送)。

已构建并在测试中 测试中

  • Jira Cloud
  • PostgreSQL 视图
  • IMAP 邮箱:仅限选定的文件夹,并验证发件人域名
  • Git 仓库地址与项目的对应关系(不会读取代码本身)

连接器将拟议的事实信息送入审批队列。在有人批准之前,它们发现的任何内容都不会成为共享记忆。已在测试副本上测试,尚未在真实账户上测试。

已构建并在测试中。未通过测试的内容将被移除。

同样已构建并在测试中 测试中

  • Microsoft 365 邮件、Teams 和云盘
  • Google Drive
  • GitHub

已构建并在测试中。未通过测试的内容将被移除。

完整情况请参见集成页面。

8. 调用方权限

AI 应用的构建目标是权限不超过其背后的人

每个请求都带有谁在请求、哪个应用在调用,以及他们在哪里工作。The AI Hive 给 AI 应用的是该人自己的权限与限时授权中较小的一个。其构建目标是让智能体看到或更改的内容不超过其用户;已由我们自己的测试检查;未经独立审计。

机器人和数字员工拥有自己的身份和一位真人负责人。它们可以在授权范围内读取和提议,但其构建目标是无法审批;已由我们自己的测试检查;未经独立审计。如果某个应用行为异常,紧急停用开关会立即暂停该应用或机器人。

由此带来的控制措施

  • 受治理的写入:没有角色、限时授权或审批,任何内容都无法进入共享范围
  • 按层级(组织、项目、团队、客户、用户、机器人)设置功能开关;下层只能收紧限制
  • 会话摘要在保存之前须由本人或机器人的负责人确认
  • 对每次允许和拒绝的审计

了解 The AI Hive 如何在您的组织中发挥作用

我们正在为首个试点寻找少数设计合作伙伴。目前还没有外部团队在使用它。请告诉我们您的团队和他们使用的 AI 工具,我们将一起确定试点的范围。