How Genie Ontology Enhances AI Cowork | Databricks Genie One Demo

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: How Genie Ontology Enhances AI Cowork | Databricks Genie One Demo · Channel: Databricks · 13:39 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Task and workflow automation, Data, databases and embeddings, Privacy and data protection

Summary

Elise demonstrates how Databricks Genie uses an ontology to prepare an adoption review by combining live usage data, governed definitions, curated pages, and context from tools such as Google Drive, Jira, and GitHub. She explains how citations, permissions, and automatically learned ontology snippets help make the results traceable and trustworthy. The demo then shows Genie investigating growth, forecasting, recommending actions, creating a reusable skill, scheduling work, and building a shareable expert agent. The same context, ontology, and permissions are also available in Slack, enabling Genie to work where teammates already collaborate.

Key points

The ontology helps Genie interpret product and metric definitions, find relevant data, and incorporate context from third-party sources.
Genie combines manually curated context with information learned from dashboards, notebooks, and saved SQL queries.
Citations show which tables, queries, tickets, documents, and ontology elements informed an answer.
Metric views provide governed KPI definitions, while ontology snippets capture automatically learned business facts.
Genie uses Unity Catalog permissions to restrict access to data and ontology content.
Live operational data lets Genie investigate current trends, perform root-cause analysis, and create forecasts.
Skills can standardize custom workflows and be shared as Unity Catalog objects.
Users can schedule proactive analyses and create shareable agents that retain relevant context and guidance.
Genie can answer questions in Slack using the same ontology and permissions as in the app.

Chapters

Start	Chapter	Summary
00:00	Using Genie to prepare an adoption review	Elise introduces the demo and explains why answering adoption questions requires product definitions, reliable data, and context from multiple systems.
02:05	Ontology-powered research and trustworthy citations	Genie combines curated context, search, live queries, and personal memory, while citations reveal its sources and governed definitions.
05:40	Ontology snippets and investigating growth	Elise examines learned definitions and their authority, then asks Genie to explain a recent increase in new users.
08:17	Forecasts, workflows, and governance	Genie identifies possible drivers of growth, forecasts outcomes, supports reusable skills, and takes governed actions through connected tools.
10:22	Scheduling and creating a shareable expert agent	Elise schedules recurring analysis and turns conversation context into an agent that teammates can use and manage.
12:36	Bringing Genie into Slack	The demo concludes with Genie answering in Slack and summarizes how ontology, live results, and governance enable the AI coworker.

Quotes

Genie just went straight to the ground truth. — 03:38

Genie enforced our permissions through Unity catalog. — 04:08

Genie is not only doing search and it’s not only thinking about numbers, but it’s really combining the power of both. — 07:44

It’s the same brain. It’s the same ontology, the same permission structure, all in context. — 13:08

Concepts and entities

Genie, Databricks, Genie ontology, Unity Catalog, Unity gateway, MCP, ontology snippets, Metric Views, Google Drive, Jira, Slack, skills, Agent

Full transcript (original language)

00:00 [music] Hi everyone, my name is Elise and I’m a product manager for Genie. And a big part of my job is keeping a pulse on the product’s adoption. You know, what’s working, what’s not, and I actually use Genie to do a lot of that work powered by the ontology behind the scenes. So, I’m going to walk you through it. Of course, every number that you’ll see is completely made up for this demo, but this is really how we do this kind of work here at Data Bricks. So, I’m starting in the Genie1 web app, though I could just as easily be in the desktop

00:32 or or mobile app. And as you can see, it’s a simple interface. There’s a chat box here in the middle. I’ve got some suggested actions down below. On the left side, I can look at some of my prior chats as well as all the agents and the assets that are available to me. But more on that in just a second. I’m going to start by asking about adoption. And that sounds like a simple question, but it’s actually not. users move between web and desktop and mobile, Slack, Teams, and Genie needs to know that. We also have a bunch of outof-date

01:05 dashboards that are just flowing around and I don’t want Genie to pick those up. And of course, I don’t just care about Genie adoption. I also care about individual feature adoption, MCPS, skills, and Genie needs to know what those features are. And it’s even more complicated than that because it’s not just about knowing where all that data is in the lakehouse. G also needs to know that our plans live in Google Docs and our bugs live in Jira and our field team shares feedback on Slack. So to really answer this, the agent needs to

01:35 understand the right definitions of our products and our metrics. It needs to query the right data and it needs to pull the right context from our third party sources. So with that, let’s get Genie started. You’ll notice that in spite of all that complexity, I gave it a really simple prompt. I just asked it to create the review with a standardized template that every team uses at datab bricks. That’s because Genie already has a good sense for how I like to write these kinds of docs because I’ve already done it in the past. And it’s also going

02:05 to be really good at navigating all those nuances that I just laid out because we have the ontology enabled. And I can actually see that thinking as Genie works. I can click in here. Looks like it started with some potentially relevant data assets and then it went into the ontology. It’s looking at some of our pages as well as some of our ontology snippets. Again, it’s combining the context that we curated by hand with whatever it’s learned from our dashboards and our notebooks and our

02:36 saved SQL queries. And I think I saw a snippet from GitHub. Yes, that’s actually Genie reaching into the search index so that the context graph can span our first party and third-party data after it went through the ontology. Looks like it also did query some of that data live. So, it found some documents in Google Drive, pulled some tickets in Jira, and because that’s using MCP behind the scenes, all that data is going to be super fresh. And then finally, it ran some SQL against what looks like our usage tables. And

03:07 with that, I already had a dock before I could even finish talking about the thinking trace. And as you can see, it’s it’s super robust. There’s some nice visualizations. It knows to click in on the things that I’m going to ask about, some highlights, some low lightss, tickets, recommendations. I notice that it’s also writing really concisely and that’s probably because I’ve already saved some personal memory asking Genie to write in a particular style when I create docs. And I can

03:38 actually see that if I click into the citations, it’s part of my personal memory. And notice that Genie did all of that fairly quickly. You know, rather than grinding through an agentic loop of trial and error and eating up a lot of tokens along the way, Genie just went straight to the ground truth. That is really the genie ontology at work. Now, one thing I do want to call out is that everything here is data that I already have access to. There’s no separate genie specific access model for every

04:08 source that you see here. Genie enforced our permissions through Unity catalog. So, I only see, you know, the tables and files that I have access to or the emails in my inbox. And the same goes for the ontology snippets that are built on top of our data. So, that’s awesome. But how do I know that I can actually trust it? For that, I’m going to dig further into these citations. This is everything that Genie actually used to handle my task again, we see tables, queries, tickets, and docs. We can also see that it ended up using several parts

04:39 of the ontology. So, for example, I think I saw in the thinking trace there were some metric views. And as as Chad mentioned, metric views are certified KPIs. They allow you to create one governed definition for something like, you know, weekly active users. So the whole company counts it in the same way. The pages that I see here are also the they’re kind of the qualitative equivalent essentially a set of governed wikis in Unity catalog. And it looks like Genie did use some of those pages that I saw earlier in the thinking trace. That’s not surprising since Genie

05:10 often favors the assets that a human has certified before it explores. And if I click into one of these, I can see that, you know, my teammate Rea seems to have added a bunch of details about the various Genie surfaces. So, not surprising that Genie ended up finding some value in this. And you can see that this is a lot of rich data. So, it’s hard, if not impossible, to try to curate every concept for the whole company by hand. And that’s why we also have ontology snippets. These are facts

05:40 about our business that Genie has learned and evaluated automatically. And in this case, it looks like we have a snippet about excluding internal workspaces. Uh there’s a snippet about how we differentiate a new user from a returning one. And there’s a snippet for how we define user activation. Now, our data science team has been doing a bunch of work on growth lately, so I bet that this snippet comes from that work. But let’s click on it to find out more. Here I can see the full definition of the snippet. We’re looking at the share of

06:12 newly provisioned users who are active for two or more days in their first week. Makes sense. And yes, it does look like Genie pulled this snippet from a dashboard that was authored by our colleague Jeffrey. And he is indeed the data scientist who’s working on Genie growth. And that’s actually super special. Imagine all the different teams at data bicks that are probably building dashboards related to the usage of their individual products. Genie found the right one for our team. And and with this really high authority score, it’s clear that onto rank even seems to know

06:42 that Jeffrey is one of our more trusted users, probably based on, you know, his track record as a data scientist, as well as how fast this dashboard is climbing in popularity. So when Jeffrey does great work, the whole team benefits even before they’ve codified these concepts as Unity catalog semantics. Of course, the data team can also formalize some of these snippets if they’d like to. so they can easily promote them to UC Semantics with Genie Code. Now, if I want to triple check authority on this topic, I can always click into his

07:14 profile. And here I’m seeing that he’s engaged with several of the domains that our team cares about. A domain is just a tag that groups assets logically. And it marrows Genie search space, which helps to optimize latency and cost. I can also click into the ontology snippets, and I see that Genie’s actually pulled a ton of ontology snippets from Jeffrey’s work. I can click in, I can learn more. But in general, I know I can trust this first pass. But I do have one follow-up question. You know, why did new users

07:44 spike so much in the past few months? Let’s ask Genie. Now, this kind of work is really only possible because Genie is computing the results based on our live operational data. In contrast, a lot of generic agents are are just reciting information that already lives in a document, but there’s no document with this week’s numbers. That’s the whole reason I’m asking Genie. So, Genie is not only doing search and it’s not only thinking about numbers, but it’s really combining the power of both.

08:17 And it looks like we have a result. The spike is pretty broad. Looks like there’s dozens of accounts and they’re adding new users. A few months ago, we opened the Genie apps for both Teams and Slack and that seems like it might be correlated. They’re now our fastest growing surfaces because it tends to lead to a lot of groups of users adopting. So growth is happening more quickly than before. It makes sense. So if it’s real, the next thing I want to know is where it’s going. So I’m going to ask Genie to do a quick forecast and add any recommendations.

08:48 That forecast would normally take me a while to model by hand, but forecasting and and root cause analysis are built right into Genie, so I can just ask. But what about an analysis that’s totally custom to our team? For example, maybe we build a a standardized launch scorecard. You know, every time we ship a feature, it grades the first few weeks against an adoption curve of our past launches. Well, that’s what skills are for. It runs the same workflow consistently. And once I’ve created a skill, I can also share it with my

09:18 teammates so that they can take advantage, too. And of course, a shared skill is simply a Unity catalog object. are always going back to that nice centralized governance. Nice. Now I have my forecast. With that, I think I’m ready to send this out to our exec team and answer any questions that they might have. I’m also going to open a ticket to graduate the Teams and Slack apps to G since that beta is is

09:50 clearly working. So, Genie is writing to Gmail and Jira as me with my credentials. So, behind the scenes, this is all governed via Unity gateway. Our admins can decide which users and groups get access to, you know, which MCP tools, and they can even set super granular policies, for example, blocking rights to particular Slack channels. Now, I don’t want to have to ask for these follow-up analyses every week. So, I’m going to ask Genie to always investigate anomalies proactively when I’m doing this kind of reveal.

10:22 I also don’t want to have to remember to run this every week. And for example, I might want this to happen when I’m out of office. So, I’m going to put this on a schedule. Of course, this adoption review isn’t the only time people ask me about Genie adoption. So, I’m also going to turn this into an agent that can answer questions about it using the context from not only this conversation, but also the other conversations that I’ve recently had about Jeanie’s growth.

11:00 I’m going to name my agent Great. Now I have an agent. This is an expert co-worker that other people can use to ask questions and take action related to Genie’s performance. And unlike an OKR doc that’s going to go stale pretty quickly, this agent remains up toate and can compute answers live.

11:32 And to that end, I’m actually going to share this with the other PMS. And now I’m going to test it. I’m going to click into the agent. And as you can see, I’m going to be able to chat directly with it. This agent has inherited the context and the guidance that I gave Genie in the last conversation. So when I ask one of these sample questions, I’m not going to have to specify what product we’re talking about because the agent is already customized to this topic. And more broadly, you can probably imagine how

12:03 other teams might use something like this. Think about anyone who fields a ton of questions. You know, IT, HR, finance. Now, everyone in the organization can create and collaborate on these expert co-workers in one single place. From here, I can I can further manage the agent. Maybe I want to tighten the scope. I want to sharpen the instructions. I can also certify it. So, when I go to the agents list page, I’ll see it marked as certified and the Geneontology will prioritize it. Now, one last thing. Everything that I just did started in the Genie app, but a lot

12:36 of my teammates live in Slack. So, I’m going to hop over here into my demo channel and ask Genie one last question. Genie is now answering right inside this Slack thread using the same ontology and the same permissions as if I were in the UI. It’s probably also going to have to do a web search uh to get those industry benchmarks that I asked for. And this is honestly one of my favorite features. Most of my teammates aren’t going to leave Slack to open up another app. They’re in here all day, every day. So instead of making them come to Genie, we

13:08 bring Genie to them. It’s the same brain. It’s the same ontology, the same permission structure, all in context. There we have it. So, in just a few minutes, I prepped for an exec review. I forecasted where our product’s going. I kicked off some new work. And I taught Genie how to do all of this even more effectively for the future without me

13:39 always having to be in the loop. That is a highquality AI coworker. And it’s only possible because Genie understands our business through the ontology. It computes results instead of just reciting docs. And it enforces our organization’s governance along the way. We’d encourage you to try this on your own data and share your feedback as we continue to shape this road map together.