11-19 更新 实时
首页/ 技术栈/ 正文

OpenAI 与 Databricks 共推企业 AI

技术栈 来源:Databricks Blog 2026-11-13 23:30 0 阅读 机器翻译
OpenAI 与 Databricks 共推企业 AI

Databricks 联合创始人兼 CEO Ali Ghodsi 与 OpenAI 联合创始人兼总裁 Greg Brockman 讨论了企业 AI 落地的最佳实践。

OpenAI 联合创始人兼总裁 Greg Brockman 用一个简单的说法来描述没有公司数据的 AI:“这就像雇了一个拥有 20 个博士学位的人,却从不邀请他参加会议。”

企业希望在自己已经在 Databricks 中治理好的业务数据上使用 OpenAI 的前沿模型,并且希望把 AI 推广到多得多的员工。OpenAI 与 Databricks 建立合作伙伴关系,正是为了让这一切成为可能。OpenAI 构建人们每天都在使用的前沿模型、Codex 和 ChatGPT。Databricks 则提供这些工具在公司内部运作所需的上下文:经过治理的数据、业务定义和访问规则。Genie Ontology 是 Databricks 中的上下文层,承载这些定义并呈现公司数据之间如何相互关联,因此使用 Genie 的 ChatGPT 在开始工作之前,就知道“营收”或“活跃客户”在你公司里意味着什么。

Databricks 联合创始人兼 CEO Ali Ghodsi 与 Greg Brockman 在本周早些时候发布的一场对谈中,讨论了从聊天机器人向自主智能体的转变,以及企业需要什么才能从中获得真正的价值。你可以在下方观看。

三种关键的企业采用模式

大多数公司通过以下三种方式之一开始使用 OpenAI 和 Databricks:

为业务用户的工作提速

具备企业级治理的编码

能够变革业务流程的定制智能体。

下面将逐一介绍。

1. 借助 ChatGPT 和 Genie 插件加速业务用户的工作

过去,需要数据来做决策的员工不会写 SQL,因此他们要么拉上分析师,要么把数据粘贴到不了解公司定义的 AI 工具里。等待分析师拖慢了决策速度,而基于错误定义得出的答案又导致了错误的决策。

现在,他们可以使用 Databricks Genie 插件,直接在 ChatGPT 中自己提出这些问题。这让产品经理可以问出诸如“按套餐层级划分,周活跃用户采用率相对于目标的进展如何?”这样的问题。ChatGPT 会借助 Genie 理解公司的指标定义,找到相关数据,在 Databricks 中运行查询,并以可视化图形和图表的形式返回结果。该插件运行在 Genie MCP 上。曾经决策缓慢,如今员工可以快速行动,并且对底层数据更有信心。

Genie 使用 Genie Ontology,这是 Databricks 中的上下文层,用于存储公司已批准的指标、业务定义以及其数据如何相互关联。Genie 会应用公司自己的定义,因此 ChatGPT 中的收入与财务报告中的收入保持一致。用户只能看到他们已有权限查看的数据,而且每个答案都会链接到其在 Databricks 中的来源。

由于 ChatGPT 拥有来自 Genie Ontology 的业务上下文,它能做的不仅仅是回答问题。用户可以向 ChatGPT Work 询问推荐的后续步骤以及应涉及哪些人,通过 Slack 或电子邮件分享发现,并通过已连接的工具执行他们批准的操作。

要开始使用,请在 ChatGPT 中安装 Databricks Genie 插件。

2. 使用 OpenAI Codex 和 Databricks Unity Gateway 进行具备企业级治理的编码

AI 编码智能体通常一次在一个团队里扩散,而领导者会逐渐失去对谁在使用它们、成本是多少,以及代码和数据最终落到何处的掌握。Databricks Unity Gateway 让这一推广过程易于管理,使公司可以把 OpenAI Codex 提供给每一位工程师,同时保持可见性与控制力。开发者像如今一样使用 Codex。Codex 使用 Databricks 凭据登录,OpenAI 密钥保留在 Gateway 中,每一次请求都会被记录。工程负责人可以设定预算、按团队审查支出,并选择由哪个 GPT 模型处理哪项任务。

要开始使用,请安装 Unity Gateway CLI 并运行 ug codex。具体步骤见 Connect Codex to Unity Gateway。

3. 借助 OpenAI GPT 模型、Model Serving 和 Databricks Apps 变革业务流程

那些可能改变核心业务流程的智能体,往往在进入生产环境之前就停滞不前,因为它们所需的数据处于严格的安全与治理规则之后。在 Databricks 上,OpenAI 前沿模型就在这些数据旁边运行,并受 Unity Catalog 治理。当由 Databricks 提供模型服务时,无需管理外部 API 密钥。团队可以在 Agent Bricks 中构建和评估智能体,也可以用 OpenAI Agents SDK 编写,并将其部署为 Databricks Apps。

要开始使用,请打开你的 Databricks 工作区,进入 + New > App,并选择 Agent - OpenAI Agents SDK 模板。智能体快速入门会逐步讲解。

OpenAI 与 Databricks 的合作关系如何衔接

这三种模式的背后是同一项合作。以下是关键事实、平台如何衔接,以及目前可用的其他集成。

关于 OpenAI 与 Databricks 合作的关键事实

OpenAI 与 Databricks 建立合作伙伴关系,帮助企业在 Databricks 中基于受治理的业务数据与上下文使用 OpenAI 的模型和工具。

2025 年 9 月,作为一项多年期协议。

涉及 OpenAI 的产品

GPT 系列模型、Codex、ChatGPT 以及 OpenAI Agents SDK。

涉及 Databricks 的产品

Lakehouse、Unity Catalog、Unity Gateway、Genie、Genie Ontology、Agent Bricks、Databricks Apps 和 Lakebase。

为业务用户提供答案与操作、在企业级治理下进行编码,以及可重塑业务流程的自定义智能体。

两家公司互为客户。OpenAI 是 Databricks 最大的客户之一,而每位 Databricks 工程师都将获得 Codex 的使用权限。

在 ChatGPT 中安装 Genie 插件,通过 Unity Gateway 路由 Codex,或在 Databricks Apps 中部署 OpenAI Agents SDK 模板。

OpenAI 模型、Codex 和 ChatGPT 直接接入 Databricks,同时有多款 Databricks 产品运行在 OpenAI 模型之上。Databricks 可提前获得新版 OpenAI 模型的访问权限,而 OpenAI 模型在 OfficeQA 上处于领先——这是 Databricks 面向真实文档工作的基准测试。

ChatGPT 是员工提问并完成工作的地方,Codex 是工程师编写和审查代码的地方,Databricks Apps 则托管各团队构建的定制智能体和应用。这三者中的推理都由 OpenAI 前沿模型完成。Databricks Model Serving 让 GPT 模型与企业数据就近运行,团队可以从 SQL、notebook 和智能体中直接调用,而无需管理单独的 API key。Unity Gateway 路由每一个模型请求,按团队设定预算,并保留审计日志。Genie 插件在 ChatGPT 和 Codex 中回答业务问题,并为它们提供来自 Genie Ontology 的业务上下文以便采取行动:公司的定义、其数据,以及每个人的访问权限。Lakehouse 在 Unity Catalog 治理下存储企业数据,而托管 Postgres 数据库 Lakebase 则为智能体保存记忆和状态。

其他 OpenAI 与 Databricks 集成

两家公司还构建了以下集成:

在 Databricks 上使用 OpenAI GPT 模型

最新的 OpenAI 模型,包括 GPT-6 系列和开放权重的 gpt-oss 模型,你可以从 SQL、notebook 和智能体中调用。查看支持的模型。

让开发者可以创建 Neon 无服务器 Postgres 项目、运行 SQL,并从 Codex 管理迁移。

面向 Codex 的 Databricks Agent Skills

官方说明文档,指导 Codex 如何使用 Databricks CLI 和 SDK。

Codex Data Science 插件

在 Databricks Lakehouse 上运行 Codex 数据科学工作流。

ChatGPT Data Analyst 插件

让用户可在 ChatGPT 中基于通过 Genie 连接的数据创建和编辑仪表板。

一键连接 ChatGPT

无需管理员配置即可将 ChatGPT 连接到 Databricks,并沿用每位用户现有的权限。

未来几周,我们将针对每种模式分别发布文章,介绍架构细节、设置步骤和客户经验。已经在 ChatGPT 和 Codex 中开展工作的团队,可以从 Genie 插件和 Unity Gateway 入手。在 Databricks 中构建的团队,可以从那里的 OpenAI GPT 模型入手。无论哪种情况,都适用相同的 Unity Catalog 权限和 Unity Gateway 成本控制。

OpenAI 与 Databricks 的合作是什么?

OpenAI 与 Databricks 携手合作,使企业能够在 Databricks 中受治理的业务数据和上下文上使用 OpenAI 模型和工具。OpenAI 提供 GPT 模型、Codex 和 ChatGPT,Databricks 提供数据平台和治理能力。

如何将 ChatGPT 连接到我的 Databricks 数据?

使用 Databricks Genie 插件,它运行在 Genie MCP 上。Databricks 管理员启用 Genie,用户随后将其连接到 ChatGPT。每个人只能看到自己有权查看的数据。

Databricks Unity Gateway 如何管控 OpenAI Codex?

Codex 通过 Unity Gateway 发送请求,而不是直接调用 OpenAI。开发者使用 Databricks 凭证登录,而 OpenAI 密钥保留在 Gateway 中。负责人可以设置预算并查看各团队的支出,且每个请求都会被记录。

我可以在 Databricks 上使用 OpenAI 模型构建 AI 智能体吗?

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(OpenAI and Databricks: Enterprise AI on your governed data)

Ali Ghodsi, co-founder and CEO of Databricks, and Greg Brockman, co-founder and president of OpenAI, discussed best practices for rolling out enterprise AI

Greg Brockman, co-founder and President of OpenAI, has a simple way to describe AI without company data: “it's like hiring someone with 20 PhDs and never inviting them to a meeting.”

Enterprises want to use frontier OpenAI models on the business data they already govern in Databricks, and they want to roll AI out to far more of their employees. OpenAI and Databricks formed their partnership to make that possible. OpenAI builds the frontier models, Codex, and ChatGPT that people use every day. Databricks provides the context those tools need to work inside a company: governed data, business definitions, and access rules. Genie Ontology is the context layer in Databricks that holds those definitions and shows how the company's data fits together, so ChatGPT using Genie knows what “revenue” or “active customer” means at your company before it starts working.

Ali Ghodsi, co-founder and CEO of Databricks, and Greg Brockman, discussed the shift from chatbots to autonomous agents, and what enterprises need to get real value from them, in a conversation released earlier this week. You can watch it below.

Three key enterprise adoption patterns

Most companies start with OpenAI and Databricks in one of three ways:

Accelerate work for business users

Coding with enterprise-level governance,

Custom agents that transform business processes.

Each one is covered below.

image3.png

1. Accelerate work for business users with ChatGPT and the Genie plugin

Historically employees who needed numbers to make decisions couldn’t write SQL, so they looped in analysts or pasted data into AI tools that didn't know the company's definitions. Waiting on analysts slowed decisions down, and answers built on the wrong definitions led to the wrong decisions.

Now, they can ask these questions themselves in ChatGPT using the Databricks Genie plugin. This lets a product manager ask questions like, “How is weekly active user adoption tracking against target by plan tier?” ChatGPT will use Genie to understand the company’s metric definitions, find the relevant data, run the query in Databricks and return results back in visual graphs and charts. The plugin runs on the Genie MCP. Where once decision-making was slow, now employees can act quickly and with greater confidence in the underlying data.

Genie uses Genie Ontology, the context layer in Databricks that stores a company's approved metrics, business definitions, and how its data fits together. Genie applies the company's own definitions, so revenue in ChatGPT matches revenue in the finance reports. Users only see data they already have permission to see, and each answer links to its sources in Databricks.

Because ChatGPT has the business context from Genie Ontology, it can do more than answer. Users can ask ChatGPT Work for recommended next steps and who should be involved, share findings through Slack or email, and carry out the actions they approve through connected tools.

To get started, install the Databricks Genie plugin in ChatGPT.

2. Coding with enterprise level governance using OpenAI Codex and Databricks Unity Gateway

AI coding agents usually spread one team at a time, and leaders lose track of who uses them, what they cost, and where code and data end up. Databricks Unity Gateway makes it easy to manage this rollout allowing a company to give OpenAI Codex to every engineer while keeping visibility and control. Developers use Codex as they do today. Codex signs in with a Databricks credential, the OpenAI key stays in the Gateway, and every request is logged. Engineering leaders set budgets, review spend by team, and pick which GPT model handles which task.

To get started, install the Unity Gateway CLI and run ug codex. The steps are in Connect Codex to Unity Gateway.

3. Transform business processes with OpenAI GPT models, Model Serving and Databricks Apps

Agents that could change core business processes often stall before production, because the data they need sits behind strict security and governance rules. On Databricks, OpenAI frontier models run next to that data under Unity Catalog governance. When Databricks serves the model, there is no external API key to manage. Teams can build and evaluate agents in Agent Bricks, or write them with the OpenAI Agents SDK and deploy them as Databricks Apps.

To get started, open your Databricks workspace, go to + New > App, and choose the Agent - OpenAI Agents SDK template. The agent quickstart walks through it.

How the OpenAI and Databricks partnership fits together

Behind the three patterns is one partnership. Here are the key facts, how the platform fits together, and the other integrations available today.

Key facts about the OpenAI and Databricks partnership

A partnership between OpenAI and Databricks that helps enterprises use OpenAI models and tools on governed business data and context in Databricks.

September 2025, as a multi-year agreement.

OpenAI products involved

GPT models, Codex, ChatGPT, and the OpenAI Agents SDK.

Databricks products involved

The Lakehouse, Unity Catalog, Unity Gateway, Genie, Genie Ontology, Agent Bricks, Databricks Apps, and Lakebase.

Answers and actions for business users, coding with enterprise-level governance, and custom agents that transform business processes.

Each company is a customer of the other. OpenAI is one of Databricks' largest customers, and every Databricks engineer is getting access to Codex.

Install the Genie plugin in ChatGPT, route Codex through Unity Gateway, or deploy the OpenAI Agents SDK template in Databricks Apps.

OpenAI models, Codex, and ChatGPT connect directly to Databricks, and several Databricks products run on OpenAI models. Databricks gets early access to new OpenAI models, and OpenAI models lead on OfficeQA, a Databricks benchmark for real document work.

ChatGPT is where employees ask questions and get work done, Codex is where engineers write and review code, and Databricks Apps hosts the custom agents and apps that teams build. OpenAI frontier models do the reasoning in all three. Databricks Model Serving runs GPT models next to company data, so teams can call them from SQL, notebooks, and agents without managing a separate API key. Unity Gateway routes every model request, sets budgets by team, and keeps an audit log. The Genie plugin answers business questions in ChatGPT and Codex and gives them the business context from Genie Ontology to act on: the company's definitions, its data, and each person's access rights. The Lakehouse stores company data under Unity Catalog governance, and Lakebase, a managed Postgres database, keeps memory and state for agents.

Other OpenAI and Databricks integrations

The two companies have also built these integrations:

OpenAI GPT models on Databricks

Recent OpenAI models, including the GPT-6 family and the open-weight gpt-oss models, that you can call from SQL, notebooks, and agents. See the supported models.

Lets developers create Neon serverless Postgres projects, run SQL, and manage migrations from Codex.

Databricks Agent Skills for Codex

Official instructions that teach Codex to work with the Databricks CLI and SDKs.

Codex Data Science plugin

Runs Codex data science workflows on the Databricks Lakehouse.

ChatGPT Data Analyst plugin

Lets users create and edit dashboards in ChatGPT over data connected through Genie.

One-click ChatGPT connection

Connects ChatGPT to Databricks without admin setup and applies each user's existing permissions.

We'll publish a separate post on each pattern over the next few weeks, with architecture details, setup steps, and customer lessons. Teams that already work in ChatGPT and Codex can start with the Genie plugin and Unity Gateway. Teams that build in Databricks can start with OpenAI GPT models there. The same Unity Catalog permissions and Unity Gateway cost controls apply in every case.

What is the OpenAI and Databricks partnership?

OpenAI and Databricks work together so enterprises can use OpenAI models and tools on governed business data & context in Databricks. OpenAI provides GPT models, Codex, and ChatGPT, and Databricks provides the data platform and governance.

How do I connect ChatGPT to my Databricks data?

Use the Databricks Genie plugin, which runs on the Genie MCP. A Databricks admin turns on Genie, and users then connect it to ChatGPT. Each person sees only the data they're allowed to see.

How does Databricks Unity Gateway govern OpenAI Codex?

Codex sends its requests through Unity Gateway instead of calling OpenAI directly. Developers sign in with a Databricks credential while the OpenAI key stays in the Gateway. Leaders can set budgets and see spend by team, and every request is logged.

Can I build AI agents with OpenAI models on Databricks?

Yes. Agent Bricks lets you build and evaluate agents, or you can also write agents with the OpenAI Agents SDK and deploy them as Databricks Apps. No external API key is needed when Databricks serves the model.

Which OpenAI models are available on Databricks?

Databricks serves recent OpenAI GPT models, including the GPT-6 family, and the open-weight gpt-oss models. The supported models list shows what's currently available.

See the Databricks Genie plugin at work inside ChatGPT.

Get the latest posts in your inbox

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 Databricks Blog(OpenAI and Databricks: Enterprise AI on your governed data),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗