Vibe Your Data Model, Supercharge Your Genie

Summary

Vibe Data Modeling uses AI agents to generate complete, validated business data models from a plain English description of an organization, replacing weeks of manual meetings with an iterative loop that takes minutes.
A real customer engagement produced a 679-table enterprise model spanning 21 domains in days, demonstrating that agentic data modeling can operate at genuine enterprise scale.
Connecting a well-structured data model to Genie Metric Views eliminates the token overhead of join definitions and ambiguous metrics, making Genie faster, cheaper, and dramatically more accurate.

Vibe Your Data Model, Supercharge Your Genie

Nobody loves data modeling. It takes weeks, requires dozens of meetings, and by the time you ship a schema the business has already changed. But skip the model and your Genie Spaces pay the price: wasted tokens on join definitions, ambiguous metrics, and answers you can't trust. In this video we show you how to break out of that trap. We introduce Vibe Data Modeling, where AI agents generate complete, validated business data models from a plain English description of your organization and iterate on them in minutes, not months. Then we connect the dots to Genie: a solid data model paired with Metric Views eliminates the token tax, letting Genie focus on understanding your questions instead of navigating your schema. The result is a Genie that is cheaper, faster, and dramatically more accurate. We will walk through a real customer engagement that produced a 679-table enterprise model across 21 domains in days and show how it feeds directly into Genie Spaces that actually deliver.
Talk By: Cary Moore, Sr. Specialist Solutions Architect, Databricks ; Roberto Bruno Martins, Senior Solutions Architect, Databricks ;

Chapters

FAQs

What is Vibe Data Modeling?

Vibe Data Modeling is an agentic workflow where AI agents generate complete, validated business data models from a plain English description of an organization. The agents propose, review, and refine the model in an iterative loop, compressing what traditionally took weeks of meetings into a process that runs in minutes.

Why does a poor data model hurt Genie Space accuracy?

Without a clear data model, Genie must spend tokens resolving join definitions, reconciling ambiguous metrics, and navigating schema relationships instead of focusing on answering business questions. As the speakers explain in this video, a good data model is worth 10,000 tokens — when the work is done up front, Genie does not have to rediscover the structure on every query.

How large a data model can the Vibe Data Modeling approach handle?

A real customer engagement described in this video produced a 679-table enterprise model spanning 21 business domains in a matter of days. The model was built agentically and fed directly into Genie Spaces, demonstrating that the approach can operate at genuine enterprise scale.

What are Genie Metric Views and how do they improve performance?

Genie Metric Views replace manual instruction rules with structured semantic definitions of key business metrics, giving Genie a clear, governed vocabulary to work from. According to this video, pairing a solid data model with Metric Views reduces Genie iteration cycles for real customers and produces answers that are cheaper, faster, and more trustworthy.

Full transcript

[00:08] I guess we're live here. Yeah. So, you can hear me. Thanks, Allison, very much. Uh, I'm Carrie Moore, like you said. This is Roberto Martins. >> Hi there. >> And, uh, we'll give you this here for about 30 seconds. Please read. Uh, we're not making any promises. No future
[00:20] forward-looking statements. I don't think it really applies to us anyway, but All right. So, this is going to be our agenda today. Um both Roberto and I have a lot of experience of developing these kinds
[00:33] of things for customers. Um but it's still hard, right? You still have to have a lot of conversations. You have to understand the business, do the vibe data modeling. We'll show you how it works through. We go customer story. Uh
[00:45] Roberto's going to give you an app and then uh I'm going to touch and end on a stronger better genie. There you go. Oh, fantastic. Um data modeling is hard
[00:57] work. Yeah. So if you've enjoy data modeling that means you like to talk to thousands of people deal with lots of opinions and um you know it's still hard right because you still have to talk to the business you have to convince them
[01:10] that this is the only way I still have customers who have different entities that they call the same thing across different parts of the organization um the hard part isn't actually building the schema or the um primary keys
[01:24] foreign keys all the relationships those are you know there's still work, but they're not hard work. Um, but also if you don't do that work, your reports aren't going to work well. And
[01:36] even worse now, your AI isn't going to work well if you don't give it context. If you make it too ambiguous, if you put all the things in there, it's just going to be terrible. So really, and I talked to many customers, they've never really done the
[01:49] full job of getting themselves a solid their normal form in their silver layer. I also like to say that like a good data model is worth 10,000 tokens. If you do that work up front, your AI is not going
[02:03] to have to do it. Your genie code's not going to have to do it. Who's built a report on bad data? What do you get? A bad report.
[02:16] All right. So, this is where they break, right? You got manual approaches. You have to talk to different people. Um, you have to code it out. it's all sort of abstract and done in the background. You really don't have a way to present
[02:28] it to people. Um, people coming from migrating from data warehouses, they may, if you're lucky, have a good silver model, a good thermomoral form, uh, source of truth and gold layers and star
[02:40] schemas, but most likely they're either dealing with acquisitions, they're bringing different data in. Um, so it's not that clean. And usually they think, "Oh, I'm going to move to the cloud and now I'm going to start from scratch and
[02:52] I'm going to rebuild everything. it will be all perfect. So, we're going to help you with that vibe data modeling. So, what does that mean? Um, basically, uh, one of our
[03:04] colleagues here who's not here is Amar Ali and this really is his brainchild. But what it is is it takes, um, information about your organization. and you tell it I am this kind of company
[03:17] and we'll actually go and build out all those entities, build out all those attributes, build out all those uh uh relationships on the fly in the background using these rules. >> Can I take it from you?
[03:30] >> Sure. Okay. So, what is in this thing? It's a whole lot of effort, trial, research that translated into hundreds of rules and a very structured workflow
[03:44] following decades of experience of good data modeling work. And that is all captured in this agent agentic workflow that uh
[03:56] goes through a very the detailed instruction of your business or it can even start from a very simple one and you iterate on top of it and it looks at
[04:08] it understands the industry understands the breakdown understands what structure it should create does an initial design reviews connects
[04:20] complements reviews again and this translates into hundreds of rules that eventually give you a full version of the model which is we call the ECM the
[04:34] enterprise corporate model or the M and the MVM. So you get both a large comprehensive version and a quick version that uh gives you the minimum
[04:46] viable model and then you can find your space in there. And just as an example, these are a few rules about specifically the topic of tagging and a quick peak at that is not
[04:59] showing sorry of the kinds of groups of rules that we have built into this model. All of this was with a lot of research and trial and
[05:13] uh it turned out to be exactly uh this uh uh a gentic workflow that you
[05:27] instruct the model what to do. It comes back to you with a proposal which you review, pick where you need to improve, where you need to evolve, what what you didn't add to it before, and iterate
[05:40] again. Each of these iterations takes what used to be weeks discussing different variations and turns them into hours. And that's pretty much the basis of what
[05:54] we want to do here and support the modelers to add the intelligence to review and quality assure the whole idea. So thanks. Yeah. No, it's like I said, it's
[06:08] a lot of work that went into this. Um, but this is a real customer story. Um, so uh the our friend Almer Oh Jesus, I'm sorry. our friend Armor uh was talking
[06:20] with this customer. It's a mining customer and surprisingly I asked the question, so why did they have to do this? You know, they're a mining customer. They've been in business for years. Why do they need to do this now? And the whole answer was because they
[06:33] had never done it. They had basically through acquisitions, through different entities across the organization, they didn't have a single silver model. So that was the impetus of why they wanted to do this in the first place.
[06:46] Um, so basically I've got a mining industry. I want a data model. What does that mean? Um, you basically have to go out and figure out what all the entities that a mining industry has. What are the
[07:00] operations? What are the different components and all of that basically comes into a final model. I mean this is where all of those attributes are established. What is a customer? What is a material? What
[07:13] are the types of materials? What are the activities as far as mining gold or mining uh carbon? All of those different things are basic activities that have to be collected in the model. That would be a lot of work for any one of us to have
[07:25] to pull together. We may have gone out and gotten some other external models and tried to correlate those, but they don't ever match your source systems. They don't ever exactly match your business. So having this with something
[07:37] you can actually interact with was very helpful for them. And so this is the uh a vessel table. This is like basically the outcome of what you get in the end. Um it's really cool to see some of the stuff about like
[07:49] the ontology, the genie ontology and all of those things that are sort of all coming together. Uh but this is what you would an example of what you would get from that model. As you can see, it has all of the relationships, all the
[08:01] entities and the attributes uh there. Um you'll see at the end we have all of these published about 40 industry models. So you can go and look at them for yourselves. But basically they were really happy. They didn't like, you
[08:13] know, it wasn't a bad thing. It was actually kind of fun. They were able to tell the the thing about its industry and it was able to go out there and build it.
[08:27] >> And that is where we were around
[08:41] talking about iterations. We're talking about feedback. We're talking about putting the human in the loop and getting a review and creating a workflow around that. What we actually had was uh
[08:53] an agentic script that was several thousand lines long and uh needed some technical knowledge of of data bricks just to run it. People could use it definitely. Did it accelerate?
[09:05] Definitely, but why stop there? Let's think about how we actually do this. We when we do data modeling, we usually draft something, discuss it,
[09:20] review it, change it, and then go over that again until either everybody is happy or we manage to tell the the people who are not to shut up.
[09:33] I'm pretty sure and some of you have been through that or at least thought about it. Anyway, uh what if we can actually very quickly iterate over these
[09:45] different ideas, show where they go, show why they are they they are actually a great thing or a dead end, why they fit or don't fit the business. Well, for that we need something better than a
[09:59] huge job technical. We need something that can actually create, inspect, compare, and refine a model directly at the speed that humans work. And uh that means giving them a better
[10:15] visualization than that prototype we just showed you, allowing them to navigate and annotate what they see, controlling how that is progressing towards a better model. And this is what we're capturing with the app that I'm
[10:28] just about to show you right now. And let me do the out tab. There it is. So bear with me. This is not even an alpha yet. Uh I hope this
[10:41] will go fine, but it's a live demo. So you know how that goes. So the app starts pretty simple. You just start with a list of business. I created two here as an example. Uh I'm going to go
[10:53] into them later, but there are some things I want to show you. We do we did talk about industry models and uh we actually have already loaded some industries in here and uh it's already
[11:08] connected to the repository. We're going to give you the QR code to the repo later where you can download any of these. I have already downloaded the mining one and if you pick any other
[11:22] let's say banking for instance when you click on it you can get some initial preview of what you're going to get and you can download it and then you get
[11:34] something like this. Uh let me go into the mining that's already here. And you can see that it downloaded the minimum viable and the ECM. You can navigate and explore the industry model
[11:48] directly from here where you can see all the domains in there. You can go into one of them. Look at the tables, understand the primary keys, the foreign keys. If you want to go even deeper. Oh,
[12:01] where this is going? This is going from the m the hollow cycle to the bench and you can navigate it there. Actually this so you can read all the
[12:13] detail. By the way, all the content here was generated by the agent and uh you can even see and the whole diagram of the model in here.
[12:28] You can click on any entity. I strongly recommend that when you're navigating such a bu a big model, you go to a specific domain and uh there are some interesting tricks
[12:40] you can do even like to to be able to better see things. So you can click on an entity. It's better when you do it bigger and you can see what it relates to. You can click on a relationship.
[12:52] You can make annotations anywhere and they are contextual. I'll show later why that matters. So you I'm here just playing with the mining industry model. But what you can do as well is
[13:07] if you go to the businesses, you see that I created a fictitious mining company here uh and I actually kickstarted it from the mining model. So
[13:20] these models here are exactly the same I was just navigating but I added a lot of specific context which you can get from your internal documents
[13:34] your process descriptions. You just copy paste them in here and that gives you the additional sauce to make that generic industry model yours.
[13:46] And uh then you run the vibe through that. I deliberately made this just the clean uh copy and I created a copy of it where
[13:58] I already built a run applying all this detailed description and generating a V2. So when you look at the run,
[14:10] you can actually see here it came from the V1 corporate model, generated the V2, it went through all these steps that the agent does and you can even navigate in
[14:23] detail at what it did at every point. And you can see everything that was thrown in there, all the instructions. I just applied the business description. By the way, there is no mining company
[14:36] called Teranova that I know of. >> If there is, I'm sorry guys, but uh and with that then you get your V2 model. And what you get when you do
[14:50] that? First thing you can see at the bar here at the top, you can see that the agent did a substantial change from the original industry model. It's now
[15:02] becoming your model. You can see that the agent thinks there's still some amount some amount of work to be done in terms of uh organizing things, getting deeper into some things. I'm going to
[15:15] talk later about this part. You can see how much effort it put into this. And then for each domain, you get a clear indication of where there were changes, where you had new domains being added.
[15:29] And uh in this case there are none here but you can see everything the model the model is the agent is suggesting you do. And if you go inside any of these you
[15:42] can also see which tables have changed. In this particular case I picked one that didn't change that didn't have anything new. So exploration got some new uh tables in
[15:55] here. And you can even go down, of course, this one everything is new, but you can go down to the field level. And what you do from here, you go back, you you start reviewing what you got out of
[16:09] this. And uh you can check all the feedback that the AI gave you. So here are all its suggestions. And you can even add feedback yourself
[16:22] as you go through it. I added a couple. For instance, the logistics uh domain was not welldeveloped here. So I'm putting a big note here and probably would add even more. And the way you do
[16:34] that I think it's pretty obvious that logistics is not great, right?
[16:48] domains are great. And then I can just for instance say
[17:09] and that captured the feedback and I keep doing that and I can control the progress of the review and I can mark which tables in each domain have been reviewed and the feedback will tell me
[17:22] by whom. So I I also added we also added the the control of the flow in here. And then when you get this review progress here to a level that you're happy with,
[17:34] what you do then is okay, it's time to create a new version. And then you prepare a new vibe. You pick which feedbacks you want to
[17:46] incorporate as instructions. If you need to, you can always add more as well, but you build the instructions directly from the feedback people gave into reviewing
[17:59] that model. And then you start a new run. I'm not going to do that now. The last run took five hours. I don't think we have that long. But uh
[18:11] uh how many of you can say you managed to do uh a completely new version of a model that has
[18:24] entities in five hours? >> No longer. >> Yeah, I never made it either either. That would usually take me weeks. So, you're going to put
[18:38] a couple of days of effort afterwards certainly, but it's going to be reviewing and discussing and refining, not just trying to track what relationships I forgot, what impact that other area had in this one that I that I
[18:51] didn't catch and things like that. It's going to be really refining the model. And that's I think it's the most important message I wanted to to give. And this is what we want this this uh
[19:04] workflow based on the app to do. And uh if that works, you get something like this. Uh a self-optimizing cycle, a virtual cycle, let's call it a flywheel.
[19:18] If you if you if you if you like uh the Amazon style, I did work there for a couple of years where you iterate and you deliver orders of magnitude faster
[19:31] than anybody is used to. And as you saw, the bulk of the hard work is done by the agent. So you get a model a model that has all the primary keys, all the foreign keys, all the the the
[19:43] descriptions and relationships at all levels and is get is checked thoroughly to be consistent and cover the business definition that you gave it. And uh we're going to be
[19:57] publishing the app version pretty soon. The industry models and the agent are already published. We're going to give you the QR code later. And uh with that about the benefits please Gary.
[20:09] >> Yeah. Yeah. Thank you. Thank you. Thank you. So the uh the one point I do also want to add is that one of the iterations basically suggested to us that we hadn't been tagging all the PII properly. So you could just go in there and just say an established new rule.
[20:21] You could do it through a new vibe and it would go through say the healthcare model and start to attach these tags to um all of the uh attributes and make sure that they're they're properly tagged and that will inform your
[20:34] ontology. that will inform your Aback. Um, all very powerful stuff that you can do just by uh generating the vibe. Um, so but the reason I wanted to really
[20:46] be here was um there are some benefits given a BI background. Um, you spend a lot of time building that star schema, that snowflake schema, etc. just to make sure that the the data mo or the reports
[20:58] and the um and the uh analytics your BI was going to work well and efficiently on that model. So I sort of came across this working with a lot of different customers. They
[21:11] would try to put a lot of different tables into a genie space and then they would try to iterate iterate iterate on the instructions. Um, basically trying to tell Genie what they could have told
[21:23] if they had just modeled the data correctly. Do this, don't do that. Um, and so with that without that model, you have to describe the joins. Um, obviously you're going to do and put all the context in there. You're going to
[21:35] add all the dimensions descriptions and other things like that. Maybe you're going to actually go in and put in some uh you know example SQL about how to calculate a specific metric. Um you're going to try to give it information
[21:48] about levels and other things like that, right? So you could overwhelm yourself by trying to instruct Genie, but you could just tell it with a model.
[22:00] Um so what I have basically come to be is the genie semantics or the metric views is what we've been primarily used. It's hard to keep up with all the different name changes here being a database company. Sorry. >> Let's just keep it chinny
[22:13] metric fuse for now. It's still working. And by the way, I didn't mention it, but the agent already generates metric field definitions. >> But anyway, so given that BI background,
[22:26] the metric views are a fantastic addition because now I can pre-build all the relationships. I can put in all the joints. I can add different metrics and measures. I can tell the things about the levels. um I can tell it pretty much
[22:39] everything that I want to be able to report on and put it in a metric view. Of course, my background tables are going to be your normal starflake schema. Um you're going to have all of that proper relationship in the
[22:52] background, but what this does is it allows you to add that level of instruction so that you don't have to do that in your um your genie model.
[23:05] And so what it does um and I don't know you've probably seen um a lot of our uh Genie best practices out there in blogs. It gives you a lot of things about how to start with the right data, add instructions, do some
[23:17] feedback, um you know, put in some pre-SQL, put in some joins. So on the left hand side is a customer who actually did that. I have C customers here, Mo and Moit and Nan. um both real
[23:31] world customers of data bricks and they did on the left hand side they went through these five iterations and they improved from 38% to 46% to 85% and to 93%. Obviously that was several
[23:44] iterations. It's a lot of work. On the right hand side though they replaced all of that with just one metric view. So from 0 to 93% just by putting the metric view in place. And now you can add those
[23:57] additional instructions. Now you can add those are different like don't do this do this context that is then going to greatly improve but you're using that context window for that information not for something that you would easily
[24:10] describe in a metric view. >> Uh one important addendon here and shameless plug my customers Alando is presenting tomorrow exactly about this topic of using metric fuse and jinny and
[24:23] how it improved their performance. >> Great plug. Thank you. Because I could talk up here all you want. You're not going to believe me. You're going to believe my customers. So anyway, um so
[24:35] yeah, so the model makes it the answer trustworthy. So why do this? Why spend all these iterations with your business users? I create a a very solid third normal form silver view that you can
[24:48] then um you know build those uh curated gold tables on top of put all of that context and awareness in there. Well, the reason is because now it's going to
[25:00] become more and more important to be able to build valid genie and AIdriven agents. They're going to be able to use that information and you're not going to have to teach it from scratch. you're going to be able to give it that business context from the beginning to
[25:14] the end. We're actually going quite quickly here, so >> that's good. It gives them more time to ask questions and throw rocks at us. >> Um, so anyway, so uh we put this
[25:26] together. It's a public repo. It's on our industry data models. You can use that QR code. It'll take you to GitHub repo. And uh what that is is 40 industry models. both the MVM which is the
[25:39] minimal viable model what keep talking the minimal viable model and then the um what is it extended there's an ECM so basically it's the everything model and the minimum viable model and so uh
[25:53] Roberto is taking us to the um GitHub repo you can see we've got some example uh graphics um different entities different relationships it's an entire readme about what the model is and does
[26:06] and how to do it. Um, and so, you know, follow that link. Um, we also have, uh, and our shameless plug, a four blog series that is going to talk about where
[26:18] this came from, how to use it, why it's important. Um, and so that's also going to be slowly released um, on our blog. So, watch for that as well.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.