Podcast

Data Vault bez předsudků #2

speciál  |  vyšlo  |  délka  |  poslechů  |   |  mp3

Naším prvním hostem se stal Lukáš Mazánek, CDO Raiffeisenbank. Je velkým fandou Data Vault, datových produktů a především AI. A přesně o těchto tématech si s ním povídáme. Tak přejeme příjemný poslech.


Strojový přepis

English translation of the Czech recording, edited for readability. Speakers: Martin Šafránek and Richard Sklenařík (hosts), Lukáš Mazánek (CDO, Raiffeisenbank Czech Republic).

Intro and Lukáš’s path to Data Vault

Martin Šafránek (00:00:30): Welcome, friends, to the second episode of the Data Vault bez předsudků (Data Vault Without Prejudice) podcast. I'm Martin.

Richard Sklenařík (00:00:47): And I'm Richard.

Martin Šafránek (00:00:48): Richard, what did you learn over the summer on your project with Scalefree?

Richard Sklenařík (00:00:53): Well, I'll start with the classic: I found out how much I don't know and how much I can't do. That's the downside. The upside is that I learned yet another business – how a reinsurance company works. Fascinating stuff. In the Czech Republic there's maybe one, maybe two reinsurers – it's quite a unique business. So yes, great. And what about you, Martin? What did you do over the summer?

Martin Šafránek (00:01:23): Something similar. Over the summer I started building a RAG knowledge base for the Czech public education sector. It will soon go into a public repository, and every school will be able to use it for free under the Apache License.

Just over the summer, next to my regular job, I put some 250–300 hours into it. It's fully configurable and customizable, it supports lots of languages, lots of document storage methods, different search strategies with an advanced statistical evaluation layer on top of it – and it's all tailored to schools.

Take one organization that runs both a primary and a secondary school, with a single document holding the school rules for both schools. You have to get a precise answer from exactly the right part of that document.

Very interesting, very exhausting – and the biggest miracle: even though I built it, it actually works.

Richard Sklenařík (00:02:15): Well, that's a bit of a parallel. Just like Data Vault, right?

Martin Šafránek (00:02:19): Of course. But today it's not about me or you, Richard. We have a very special guest – someone with a long history with Data Vault at a financial institution, who for the last two years, as Chief Data Officer (CDO), has been looking at data products from a completely different angle than the two of us. So, Lukáš Mazánek, CDO of Raiffeisenbank Czech Republic – welcome.

Lukáš Mazánek (00:02:45): Thank you for the invitation, gentlemen. I'm really looking forward to this discussion.

Richard Sklenařík (00:02:48): I'd like to start by explaining why Lukáš Mazánek is here, why he's our guest – and there's one simple reason. When I was working on a project for Raiffeisenbank Czech Republic, Lukáš was, in my view, the one who decided on Data Vault and backed it quite strongly.

I've always wanted to know: what led you to it?

What convinced you to take that path?

Lukáš Mazánek (00:03:26): At that time I was Head of Data and Information Systems, which included the data warehouse.

For a long time we had been trying to solve one problem: how our data warehouse could keep adapting to a constantly changing business environment, and what that meant for the bank's finances. Our data warehouse was, of course, a significant item in the budget, so the strictest criteria applied to it.

I started looking for a way to change our approach to warehousing, so that we could respond to a fast-changing world. And Data Vault was a fairly simple answer to all our questions – at least from my point of view.

Data Vault has several unique properties that fit exactly into the puzzle. So that's it.

Richard Sklenařík (00:04:44): Nice. I remember our lunches together, when you'd always throw exactly these questions at me: "And can that Data Vault of yours do this?" Did you have any other sources of information about Data Vault – apart from those lunches with me?

Lukáš Mazánek (00:05:03): I'm naturally curious, so I always try to study a topic from books, professional articles and discussions with experts. I put the picture together over quite a long time.

One of Data Vault's strengths is that its methodology is very well described – and it's broad, I'd say almost complete. It starts right at the top, with the business, and takes you through the whole process down to the technical solution and the consumption of the data.

So I studied the methodology, and I also read the book "The Elephant in the Fridge" – if I remember correctly, it was you who gave it to me, Richard.

Richard Sklenařík (00:06:07): Definitely.

Lukáš Mazánek (00:06:08): Then I certainly read some articles by Patrick Cuba. I have to say he's a very readable author who can explain things really well. After all, his book The Data Vault Guru is, in my opinion, brilliant – a must-read for data people.

So I was equipped with the academic knowledge, let's say. What I always miss, though, is the real-world view. That's why I kept trying to pull this information out of you, and that's why we eventually started by defining a proof of concept. Because that's the best way to start.

Richard Sklenařík (00:06:54): That's true. That was the so-called RTDS – the real-time data store.

Lukáš Mazánek (00:07:01): Yes, exactly.

Richard Sklenařík (00:07:04): Right – and the most interesting part was that Data Vault showed it was ready for challenges like processing real-time data and daily batches in parallel, without any special tricks.

That was the reason we went for it, because the existing data warehouse couldn't support real time at all.

Lukáš Mazánek (00:07:34): That's right, and at the same time it was our biggest worry: whether we could process real-time data in a way that would still answer warehousing questions. That's why we started with RTDS and chose exactly that use case – we knew it was a knockout criterion. If it didn't perform, it would be of no use to us.

Richard Sklenařík (00:07:58): And back then you also managed to train some people at the bank, is that right?

Lukáš Mazánek (00:08:04): That's right. From the very beginning we set a condition: every member of the team working on Data Vault had to go through official training. I think Scalefree provided it for us directly.

It was my decision, I pushed for it – I thought it was important that we build the elephant in the fridge, not a giraffe. Because, of course, incomplete knowledge of the methodology means you get sort-of Data Vault – which means no Data Vault at all.

That's why we wanted everyone to go through the training. For us it was a paradigm shift, a fundamental one, and we felt it was very important that people came to the work properly prepared.

I think it paid off brilliantly, because the proof of concept turned out very well, and to this day RTDS serves the real-time needs of our system – for marketing, I believe?

Richard Sklenařík (00:09:24): Yes, that's right – real-time marketing is implemented there, or rather the data foundation for real-time marketing.

Still, that's where it stopped. In the end, Data Vault never made it into the bank's data warehouse itself. Once the proof of concept was over – and RTDS is now only in maintenance mode, not being developed further – the people unfortunately couldn't use the knowledge they had gained any more.

Lukáš Mazánek (00:10:01): I wouldn't be that strict. The experience those team members took away is visible in various places across the DWH.

The fact that we didn't continue developing Data Vault – or rather rewriting the whole data warehouse in Data Vault – is to a large extent my failure, let's say.

With a change this big, it's extremely important to get buy-in from everyone on the team, and the data warehouse team at Raiffeisenbank Czech Republic has around 100 people on average. It fluctuates – we know how to shrink the warehouse team quite a bit, but also how to grow it when there are acquisitions.

Either way, the team as a whole needs to believe in the idea to be able to make big changes. I didn't fully manage that. We had many excellent experts on the team who were very experienced in traditional Bill Inmon warehouses and traditional warehousing.

Lukáš Mazánek (00:11:35): Although they accepted most of my arguments, where we parted ways was the technical view of how Data Vault stores and processes data. As a manager – and as a data-positive person who cares about what data means – that didn't matter so much to me. But for most of the technically minded people and architects, it was a crucial point.

I heard one of what I'd call the most notorious objections: that it has too many tables. And plenty of other technical objections – I don't want to say those technical objections were irrelevant. I just mean that, from my perspective as a manager, they weren't that essential.

I look at it a bit like a black box: if there are Data Vault implementations in organizations much bigger than ours, and those instances meet the requirements and are among the best in the world, then honestly, how the data is technically stored didn't matter that much to me.

But as I said, if you're making a change this big, at least the core team has to be convinced as a whole – and I simply didn't manage that.

Richard Sklenařík (00:13:32): I might add the consultant's perspective – the analyst's, the developer's and so on – because I went through this myself. I have to say the Data Vault approach is so different from the classic Bill Inmon approach, or rather from how the Inmon approach was implemented at Raiffeisenbank, that you can't do both in parallel. You'll do either one or the other.

When I talked to people and collected feedback, the reaction was basically simple: "It would be great if we could start building part of the warehouse in Data Vault now." But as you say, they were experts, experienced people – and people like that usually have an irreplaceable role in the existing team. They couldn't just disappear from it overnight, because they were needed there. I think that also contributed to it not working out.

But let me ask you something slightly different. You already mentioned that one of the books you read was The Elephant in the Fridge. I don't know if our listeners know it – if not, I warmly recommend it.

Richard Sklenařík (00:15:05): The thing is, that book isn't primarily about Data Vault. It's primarily about metadata – about semantics, taxonomy, ontology and so on. And you might ask: fine, but how does that relate to Data Vault?

Very simply – because that's the foundation, the core, the place where you have to start if you really want to use the Data Vault methodology, and use it correctly. That's the starting point where you have to be able to get hold of your data, find the business objects, find the business keys and so on.

Lukáš Mazánek (00:15:46): Richard, you're absolutely right, but I'd still put it differently. Taxonomy, ontology – those are words that scare most listeners straight away, and lots of managers feel it's something intangible, some academic approach we don't need here.

I'd rather start by talking about how Data Vault is connected to the business, and the ontologies, taxonomies and the meaning of data will follow from that naturally. It's simply like this: at the beginning there is always some business architecture. That business architecture is usually made up of capabilities, those capabilities are part of value streams, the value streams have stages, and of course there are information concepts in all of it.

So if I understand the business well – and that should always be my aim in any organization – and I have the business well described architecturally, because if I don't, I can't make changes, I can't manage the change – then I can start deriving, for example, the information concepts from that business architecture. And, quite surprisingly, those information concepts start to show up, to a large extent, in the hubs.

Martin Šafránek (00:17:41): Lukáš, let me stop you there for a moment. Some of our listeners don't have such a deep data background, and you've used several terms that I think deserve an explanation. So could you tell us: what is a business capability, what is a value stream, and what is a business or information concept?

Lukáš Mazánek (00:18:00): So I made exactly the mistake I criticized Richard for. All right, back to the start. So that we're not talking at the level of "it would be nice to make more money", we need to say: this company creates value in a certain way.

And there are business architecture tools we can use for that. We use value streams, for example – which basically describe the path from the first step, through the individual stages, to the point where I've delivered some value. That value can be, for example, a signed contract.

So at the start of such a value stream there's client acquisition, then gathering the client's information, then drafting the contract, and finally signing it. So I have a value stream that delivers a specific value. And in that value stream I've already named several concepts.

By a concept we mean a general mental image of something – in this case, a document or a signed document. Those are concepts.

Lukáš Mazánek (00:19:34): And they can be expressed by a set of data – then they're information concepts. So if we're talking about a specific document, at the end of the day that document is a pile of text that captures its content – and that is the content of the information concept, which in this case is the document.

Now, if I have the architecture of the bank – the business architecture – mapped like this, I can start mapping it to the IT architecture or to data solutions. In this case we're talking about Data Vault, which, as listeners surely know, stands on three basic pillars: hubs, links and satellites. The hubs are the places of truth, let's say, about the information concepts.

Richard will certainly add the technical view afterwards – I'm looking at it from the business side. From my perspective, a hub is the place where I've decided that there is a business-relevant information entity which I need to identify uniquely by a business key.

Data products, semantics and AI

Lukáš Mazánek (00:21:08): And around that information entity, the individual pieces of supplementary data start to orbit like satellites.

And here we get to one of Data Vault's nice properties: I can attach any number of satellites to such an entity through the business keys, without fundamentally changing the architecture of the data warehouse.

Then there are the links, which let us connect individual hubs, or information concepts, together. Here we see the reference back to the business architecture or the data architecture, because at the end of the day – and now we're getting to ontologies – I need to describe the business or the data in a way that says specific concepts have some relationship to each other.

For example, a customer has a product. That's a standard ontology triplet, which can be captured in RDF format – subject, predicate, object – meaning that something has a relationship with something else. So the customer has, or signs – there's always a verb in the predicate – a contract, a product, anything like that.

Lukáš Mazánek (00:22:46): And that connection, that relationship, can be captured as links in Data Vault. In my view, this is one of the key differentiators between Data Vault and, let's say, standard relational databases – technically, or rather in terms of the meaning of data.

Because in standard relational databases, paradoxical as it sounds, relationships are often implicit. They usually aren't explicitly modeled or captured. Of course, in the relational world I can always build table A, a junction table and table B, capture the relationship in the junction table and describe it somehow – but that's not usually how it's done.

Usually there's just a foreign key that captures the relationship, and its meaning is usually inferred from the content of the connected entities.

For me that's a big problem. You have the same thing in natural language: imagine what a sentence without verbs would look like. You simply can't convey any information that way.

Lukáš Mazánek (00:24:18): And that's the change that a data architecture based on ontologies, or knowledge graphs, brings me – or Data Vault, which stores data explicitly in this way. I can come and look at which relationships, which verbs, I have available for these specific concepts.

And now you start to see how the picture comes together from the top down and back again: things are connected to each other and work as one whole. In my view, that's probably the biggest difference between Data Vault and traditional relational databases. And I've already forgotten what the question was.

Martin Šafránek (00:25:07): That's completely fine, and we can show it with a nice example. A bank has a client – someone bound by a contract with the bank. And then there's a prospect – someone who might become a client, who wrote an email or was approached somehow. So now we have two groups of parties.

The bank has a different relationship with each of them; each may be recorded in a different system; we have a different set of information about each. For the prospect we may only know a phone number or an email address, which isn't ideal – it's a temporary identifier. And the bank may have to look at each of them differently from the GDPR and legitimate-interest perspective.

That's exactly the moment when the business architecture, the value stream, flows down into the IT architecture and then all the way to the individual hubs – where I'll have a client hub and maybe a prospect hub, with different business keys.

Lukáš Mazánek (00:26:10): This is a lovely discussion, so if I may, I'll pick up on this example. Of course, there are many ways to model this situation. I'd almost rather think of the prospect or the customer as a role of a natural person – or a legal entity – which generally means some party, some counterparty. I'd even want to capture third parties and, let's say, any other counterparties of that organization.

And that's another nice thing that ontology, or semantic modeling – and therefore Data Vault – lets me do: I can say that something is in a certain role, and model those relationships. So we come full circle back to how important it is, first of all, to understand the business. It's really crucial to understand who a prospect is, when they become a client, under what conditions, how they change, how the individual concepts differ – but at the same time how they're similar.

So I'm already building a mental image, an ontology of it. Here I might even get by with a taxonomy, where I'd simply say there are some categories of natural persons. But roles would probably be more suitable – so I'd really use an ontology.

Martin Šafránek (00:28:02): Great. And a step away from ontology and taxonomy, we're close to the topic of data products.

I had the opportunity to be part of a number of discussions where you, Michal Čech and others at Raiffeisenbank created a very detailed methodology for implementing data products: what a data product is, not only from the technical point of view but also from the business point of view, how a data product carries semantics, and how it fits into the overall map of concepts in the company.

Lukáš, my question – because I know data products interest you a lot, and data products solve many problems and also bring some new ones: what's your view on data products being the next stage in the evolution of traditional data warehouses?

Traditional data warehouses suffer from a number of problems that can't be solved: ownership, making the business accountable for data, and the fact that the data warehouse has thousands and thousands of tables, and the contract is stored in thirty of them, each time with a slightly different semantic definition.

What's your view?

Lukáš Mazánek (00:29:20): I think it's extremely important to start by answering the question: why have data products at all? For me, the answer is: because of AI. It's hugely simplifying, but if I go right down to the core, for me it's AI.

It's exactly the same reason to have Data Vault – again, the strongest reason is AI.

Because the world has changed, and changed quite quickly, and today everyone is flirting with generative AI and agentic systems in some way. Machine learning has been with us for quite a long time, but GenAI is a fairly new phenomenon. AI is absolutely, vitally dependent on high-quality, semantically meaningful data, and that data has to be AI-ready.

That's a requirement on data management, on data governance, which enforces further properties of the data so that it can be AI-ready. So why data products?

Lukáš Mazánek (00:30:50): Because of AI. In my view, it's a fairly suitable way to decentralize the data architecture in an organization. And here again you have to ask yourself: do I need to decentralize the organization?

Because if I don't – if the organization is small, or not at a level of maturity where it can afford decentralization – then maybe a centralized data warehouse without data products is the best solution. Or a centralized data warehouse with data products, and so on. So you need to ask the right questions, and they will lead you to the answers.

In our case, I believe Raiffeisenbank Czech Republic is mature enough and ready to be an agile, decentralized organization, at least in the area of data. So we decided to use data products as a way to shorten the path of data from the source systems to the consumer, and to give data a meaning that both a machine and a human can consume in essentially the same way.

Lukáš Mazánek (00:32:22): And this is the approach for moving the organization into the desired state. So a data product is one of the tools a data architect has available. Can it be done differently?

Yes, it can – but it brings some specific advantages and some specific disadvantages. You hinted at some of them, Martin, and there are many more. I'd also warn that the data world has its fashion trends too. Right now data products are, as they say, all the rage, so they're used for everything and everything is claimed to be a data product.

So again, I'd warn: you don't necessarily need data products. If you're not ready for them, or you don't need them, there's no shame in not having them.

On the other hand, I think sooner or later AI will force every organization to make its data AI-ready. How to reach that goal is, of course, up to the organization – it should be the most efficient path. I still believe that both Data Vault and data products are suitable tools for organizations like ours to achieve that AI-readiness.

Richard Sklenařík (00:33:53): And when it comes to data products and Data Vault, do you see Data Vault directly supporting them?

Lukáš Mazánek (00:33:58): I actually see a huge connection there. As Martin said, we designed a fairly detailed methodology for data products – how to introduce them in our organization, again from the business all the way down to the specific implementation.

And within that methodology we found that one of the key questions is: how big should a data product be? What should be in it? Because, in theory, a data warehouse can be a data product. We have it as a data product – yes, it's universal, everyone uses it, it's very popular. But we all somehow feel that's not the way to go. By the way, since we're on the subject: we iterated our way to the conclusion that our guide is the "aggregate", as defined in domain-driven design. So our data products revolve around an "aggregate".

Lukáš Mazánek (00:35:29): The aggregate has a root, and around it data satellites orbit, so to speak. And it's all connected by relationships. And here we're slowly starting to feel that Data Vault is a very good foundation for building specific data products around the individual hubs. Because when I look at the hub map of a particular organization's Data Vault – if I wanted to go about it in a completely technical or mechanical way – I can make a map of the hubs, look at how they're connected to each other, and pick the ones with the most connections, because those are probably the most important.

And then I can think: OK, here I have some big business value, something really important is happening here. So, OK, it's the client, and it's there for some purpose. What purpose? Fine – so I carve out this part of the client together with something else, and together that gives me a data product.

The Data Vault methodology, the way it stores data, guides you very naturally to what the data product is and where its boundaries are – so that it's not too big, not too small, and its structure doesn't change too often.

Lukáš Mazánek (00:37:00): So that it's stable in some way, because a data product has to have a data contract, and what you don't want is a data contract that changes every week. So, in my view, the nature of Data Vault maps very easily onto data products, and I don't think even the transformation of the data will be that complicated. Building data products on top of, let's say, a relational database that already has some history – and Raiffeisenbank has a history of several successful acquisitions – isn't a simple transformation.

By the way, it usually turns out that the business meaning, the business use of the data, doesn't follow from the data itself.

So when we create data products, we start with something we call storytelling. It's a real methodology that lets you describe what you want to achieve with the data product in fairly simple statements, very similar to the subject–predicate–object principle. So I build these sentences out of concepts and relationships.

Lukáš Mazánek (00:38:32): Then those concepts go into a business glossary – they become known to the business, and the business starts to understand them. Around that, I can build a business knowledge blueprint. The business knowledge blueprint already contains concepts and relationships. And with a bit of luck, those concepts and relationships would show up in Data Vault as hubs and links. So I'm a big fan of both Data Vault and data products precisely because they beautifully connect the business, the meaning of data, technology and, let's say, the way data is communicated.

Richard Sklenařík (00:39:14): And you've set me up nicely for a little commercial break – sorry, Martin. You asked me what I took away from the summer I spent on the Scalefree project.

Among other things, we use dbt on that project, and we use exactly the features related to data products. That's what you mentioned, Lukáš: in dbt you can define on your models that a contract exists, whether the contract is enforced, you version it, and dbt makes sure you don't break the contract if you happen to want to change the structure, and so on.

And if you want to try it out in combination with Data Vault, there are libraries that implement the Data Vault methodology for dbt and put it into practice very easily. There are several creators – the Scalefree company I mentioned and, to be fair, the British company Business Thinking, which came up with its own library; they're very similar. In fact, we're planning another guest who will tell us a lot about this library, I think – but that will be more of a technical session.

Richard Sklenařík (00:40:44): And if I'm not mistaken, a week or two ago Patrick Cuba announced, I think, his own version of dbt libraries that support Data Vault. So that was just an almost-commercial aside.

And one more thing caught my attention in what you said, Lukáš. I have to say it's been haunting me ever since – that's why I handed over to Martin, because I wasn't able to react. I liked how you started comparing it to language syntax. Taxonomy, ontology – that's basically what it's about.

So let me just add: yes, links are like the verbs in our speech. And if you're looking for adverbs, those are the satellites on links. Just a small note. Yes, every link has its satellite, and if it doesn't, you actually know nothing about that link.

So that's quite an important thing for me. And speaking of language syntax – AI language models have already come up. A big topic – or is it big? Hard to say.

Lukáš Mazánek (00:42:00): It's definitely big. Fundamental, in my view. The big turning point came at the end of last year, when the first Claude Code models were released. From that moment on, it's been fairly easy to vibe-code basically anything, which will clearly affect practically every area of business and every organization. And at that point, of course, the question of what it will cost starts to come up.

Because there are organizations that, in their excitement about AI arriving, have already let their employees go, while others are sensibly waiting to see what it brings. Right now I feel a slight cooling of that enthusiasm. It seems to me that, especially on the upper floors, there's a bit of a question mark: yes, it's interesting, yes, you can easily create software or solutions, but what will it actually bring us as a company?

How will it pay back?

Lukáš Mazánek (00:43:30): So the question has two sides again: benefits and costs. As for language models, their use in our bank is growing like a rocket. That's good – it's thanks to the people and teams working on adoption.

But of course, price tags are starting to appear, and at that point questions about local models come onto the agenda – large or small language models. So we're working on that too: the idea of having our own language model and powering our solutions with it, whether for reasoning or for coding.

With small language models, you simply train the model to make decisions on a specific business topic. That's an easy entry point – the small language model – and at the same time it connects straight away with everything we've talked about here.

Lukáš Mazánek (00:45:01): Because fine, you have a language model, and that language model is capable of some reasoning. But to give good answers, it needs good inputs. That means the data has to be structured, described, fresh, it has to have governance, it has to be continuously looked after.

What's becoming clear now – and I think this will be the main differentiator between the winners and the rest of the market and the industry – is that whoever has the best-prepared data will be able to make the most of the leverage of AI and language models.

Because today the situation is that, in my opinion, you can buy a model so good that you don't actually need it. You can get by with a not-so-good model.

Lukáš Mazánek (00:46:20): I've been playing around…

Richard Sklenařík (00:46:22): (cutting in) That's something like an overqualified employee.

Lukáš Mazánek (00:46:24): Pretty much. I've just installed Qwen 3.6 on-prem. It's good. I'm not saying it's Claude Code – you can't do absolutely everything with it. But let's be honest about what things look like in organizations today.

What gets automated is a situation where someone copies something from one column to another, so that a robot picks it up somewhere and it appears in some other Excel sheet, where someone else does something with it again. Processes like these need to be automated, we need to get rid of them, because those people can do more interesting work – more rewarding for them and for the organization. But to do all that, I need to understand the business, and I need to have described data that is AI-ready.

And once I have that, it can be automated on a relatively simple model. I really don't need a top-tier model – something much, much simpler is enough. A language model, I mean.

Richard Sklenařík (00:47:31): But I wonder if it's partly a problem of trust. When you have a relatively simple process and hand it to AI, you're able to validate the AI's output yourself, because you understand it – you can validate it, you can trust the output. If you really wanted to push those big, complex language models to the limit, the things would be so complex that you couldn't validate them yourself, and you'd have to trust that it's fine, that it's correct – in an organization like a bank. Maybe that's the snag. Is that why you can't use the big language model?

Lukáš Mazánek (00:48:18): I think that really is the problem – or rather, it's a perceived problem. But in my view – I don't know how to put it – we're not there yet. You see, we're solving trivial problems, for which trivial but well-described data and a really small language model are enough.

And I can build trust in it on a set of, let's say, golden test scenarios. I can run repeated regression tests, I can measure it – which, by the way, is another well-known, well-worn truth: what I don't measure, I don't manage. Every agentic system has a self-improvement loop, and I can't build a self-improvement loop if I don't measure, because then it's hard to improve.

So it is a problem of trust, but in the wrong place. In my view, you need to start, and start small – you need to learn to work with the data properly. I can have the data prepared according to the Data Vault methodology, I can have a simple, well-described data product, and I can start with just that one data product, for a specific business problem, with a small language model.

Trust, context, ontology and domain-driven design

Lukáš Mazánek (00:49:48): Get people on board, prove that it works correctly – and there are methodologies and methods for ensuring that correctness, because hallucinations are, of course, a real problem that exists and has to be dealt with. But on problems this small, I don't run into that distrust.

I agree: yes, a large language model has lots of problems – but that's very far away. I think before that come local large language models, and only after them the top-tier language models somewhere in the cloud. But we're not there.

Richard Sklenařík (00:50:26): I think there would also be a security problem there.

Lukáš Mazánek (00:50:31): Of course – security. In general, security now seems to me one of the most crucial areas, because as large language models become more and more capable, security becomes more and more important.

And again I'd stress: start small, think big – on specific data, with a small language model, on-prem, locally. That way I cut off a lot of problems that I'd have to solve with a large language model and that I don't need to solve. Because with a large language model you need to audit how the model is put together, since it will certainly have some bias.

A hundred percent – because even for people it isn't easy to say where the truth lies. So the model is roughly as biased as whoever trains it and the data they give it, which means it needs audits. It's known that some language models prefer, for example, Western companies over Eastern ones. Now, if I use such a model to simplify work in procurement, I'm already influenced by that bias.

Lukáš Mazánek (00:52:03): But as I say, these are problems we don't need to deal with at all. Let's use nice data with a small language model, or a local large language model, for our purposes, and let's show people that it works. Let's leave that for the next five-year plan.

Martin Šafránek (00:52:21): I'll certainly pass that on. We're talking about three components an organization needs to use AI effectively today.

We're talking about models – ideally local, which removes a whole pile of issues. Look at Anthropic's videos and official documents about how long your queries, your prompts, are retained, because Anthropic has to analyze misuse of the more capable models. Very interesting – I had a very long conversation about it on another project at a large financial institution.

So, ideally, a smaller language model.

The second thing we need is data.

Third, what connects them is a description of how the data is stored, what it means semantically and, above all, how the organization works with that data. And that's exactly the point where the language model simply doesn't have the information.

It's trained in general, it knows in general how a bank works, it can infer how the data is probably used. But without machine-readable information about data products with a very strong semantic layer – the information Lukáš is talking about here – or a Data Vault model, again mapped to business processes, AI in that organization simply won't deliver as many business benefits as it could.

What really enables AI to deliver those business benefits is metadata. Not only describing the data, but also explaining how the company works. Do you agree?

Lukáš Mazánek (00:54:00): Yes, it would be hard to disagree. Rather than metadata, I'd talk about context. I think what's changing a lot right now – and maybe it's not so visible – is the need for high-quality context, because we'd like to approach machines and people in the same way.

We talk about agents, and they should be interchangeable, but of course there's one big difference.

People have something called implicit context. As we sit here, we simply know where we are and why we're here. So when I use a word, all of us here weigh it in the context of the fact that we're recording a podcast right now.

But an AI agent doesn't have that by default. It has no implicit context, which means I need to supply that context. I build the context from data and metadata, of course – but it changes the approach to data management.

Lukáš Mazánek (00:55:32): In the past it was very much about ownership of specific entities. I think in the future it will be more about context: how I build it, who looks after it, who decides what goes into it. And how well the context is built largely determines how successful the agent is.

So I need to change data management so that the metadata and data are in a state where I can build context that lets the agent – whether a human or a machine – make the right decisions for a specific business use case. In my view, that's a new discipline of data management.

And again it brings us back to how I store the data so that it's AI-ready, and how I'll move it around. So: Data Vault, data products. It keeps guiding us there, which is why I say that, from my perspective, the strongest driver of these changes is AI. I remember the days when you basically couldn't say the word ontology out loud. Today a proper strategy has to contain the word ontology at least once. And that's a good thing.

Lukáš Mazánek (00:57:02): Yes, it may be a bit of a marketing term, but it shows that the whole organization, even at the very top, understands that data has to have meaning – without meaning it has no value – and that the organization realizes data is, to a large extent, the most important thing it has.

Human resources belong there too, of course, I agree. But imagine our bank lost all its data overnight. That's simply irreplaceable. So the company is changing, and we have to adapt to that change – and that change is meaningful data, business-meaningful data.

Martin Šafránek (00:57:51): And since we're on changes and challenges, I can't help but recall a discussion I had with Martin Hajný from your bank – just for listeners, Martin Hajný is one of the AI leaders at Raiffeisenbank Czech Republic.

Some time ago I had a debate with him about the bank's internal methodologies and manuals that keep the bank running, and how readable they are for people and for machines. For twenty long years we created documentation readable by people. And people have, as you said, Lukáš, implicit context. Those methodologies are used for regulatory reporting to the Czech National Bank, for a risk model, for approving loan applications – and that's tens and hundreds more pages of methodology.

And the reality is that even when we have context, even when we have the data and everything else, we still need to give AI agents a machine-readable methodology as well – those thousands of pages provide better context, describe the origin of the data and affect its business meaning.

And here we can run into a number of problems – I started running into them myself when I was building the knowledge base for the Czech public education sector – because AI may not clearly understand even such basic terms as "client". Without the right context, the AI agent doesn't know which client we're talking about.

Martin Šafránek (00:59:25): Are we talking about a retail client? A retail client doesn't have to be an investment banking client, and so on.

And again, in the knowledge base for schools, there's the topic of school supplies. For example: what supplies should I bring to school? Without context – are we talking about supplies for regular lessons, for primary school, for which year? Maybe they're supplies for entrance exams.

These simple words just don't work without context, without the methodological background.

Lukáš Mazánek (00:59:58): That's true. Domain-driven design helps me enormously here. I think it should be required reading for all architects. The meaning of the data is, to a large extent, hidden in the domains, because the data then sits in a specific context – the context of that domain.

And by the way, this is one of the big challenges a data architect faces. Because from management's point of view, it's best to have everything exactly once. First, it's cheapest when you only have it once, and second, you obviously want to have it right – that's why you have exactly one.

The problem, of course, is that data changes its meaning depending on the context. I think I had an example in some talk: when you say the word Jaguar, it can mean the big cat, it can mean an operating system, it can mean lots of different things. And the context helps understanding.

Lukáš Mazánek (01:01:31): So if I say Jaguar here, it'll be a bit puzzling, because the context of this podcast has nothing to do with cars or zoology. But if I talked about a jaguar at the zoo, people probably wouldn't picture a car – or at a racetrack, the animal – though maybe someone would.

So context is more and more important, and among other things it comes from the specific domain I'm working in, because the broadest definition of a domain is that a domain is formed by the boundary of a language. Basically, the people in a domain speak a language they understand – just notice it.

When people in retail or corporate banking talk, they have specific names for specific things that they understand.

One word is a complete description of the whole situation for them, while we have no idea how anyone can start a fairly complex process based on a single word. So the language, the meaning, depends on context and is enclosed within the domain. So if I go back to the question, I actually try to look at everything from the domain-driven design perspective, and design the solutions that way.

Data Vault methodology, data products and vibe coding

Martin Šafránek (01:03:05): Thank you for adding that and explaining it. That talk was from Data Dreamland 2024, if I remember correctly – we prepared for it together. I later took the liberty of using that lovely example at an AI conference.

There's one more topic hanging in the air: the Data Vault methodology and the methodology for data products.

My view is that the Data Vault methodology – even if I look at the thousand slides you need for the Data Vault 2.1 certification – gives data products very little space. Maybe a handful of slides, and all the methodology says is: here are the layers of the data warehouse; put data products either into the Business Vault or into an Information Mart. That's all.

Then I had the chance to work with you and Michal Čech on implementing data products at Raiffeisenbank, and I suddenly realized that the data product methodology not only can be much more detailed – it has to be.

So the question is: isn't that a bit of a pity? And a second question: what's your feeling about people in the Czech Data Vault community – aren't we sliding too much towards seeing Data Vault as a technical, architectural system, rather than as a means of communicating with the business?

Lukáš Mazánek (01:04:34): I certainly feel the problem I described at the start: the focus is too technical – and in my view that applies to Data Vault as well. However much Data Vault is, to me, the ideal of completeness and of tying data to the business, when I talked to experts about data products, the answers were still largely technical.

I badly missed the link I outlined here: between how nicely the business knowledge is connected and stored in Data Vault, and how easily I can build a good, working data product that delivers business value.

In my view, that's still a debt that Data Vault has too. I think it would be worth emphasizing in the methodology, precisely because it's the enabler for AI. What I'm saying is: we understand the business, we can store data so that the business finds its way around it, we can move that data, and we have data products for it that are tied to Data Vault and to the data stored in business terms.

Lukáš Mazánek (01:06:07): And at the same time, both the way the data is moved and the way it's stored are semantically meaningful. That means it's readable for a machine as well as for a human. To me that's an extremely sexy message – and I still don't hear it. I don't hear it loudly enough.

So if I may make a wish, I'd ask Dan Linstedt to add it.

Martin Šafránek (01:06:32): We'll certainly send him the transcript. I think, yes, this is one of the areas where the Data Vault methodology should keep evolving. The methodology hasn't stood still: an update – or rather an extension – is being prepared, since the Data Vault 2.1 standard was published two years ago.

And it will all revolve around one theme: the more AI systems we have around the data ecosystem, or the ecosystem in general, the more complex the questions of trust, credibility, controllability and security become. So Data Vault will move in that direction.

But I do feel that a detailed data product methodology is missing from Data Vault and should be added.

What do you think, Richard – am I being too bold?

Richard Sklenařík (01:07:25): I don't think you're being too bold. It's true that this is something I've missed in the Data Vault methodology from the very beginning, already in version 2.0 – something I call implementing regulations.

It's also what we answer today when people ask what the difference is between the 2.0 and 2.1 methodology: should I learn 2.0 or 2.1? Basically, it's still the same methodology. Version 2.1 only developed things that weren't missing from 2.0 – they just weren't described, weren't highlighted, weren't covered in that much detail, or weren't mentioned at all.

But that doesn't mean the methodology described as version 2.0 couldn't handle them. It could handle them back then. And it's the same with data products: yes, you can implement data products with the Data Vault methodology today – no matter whether you take the older 2.0 or the newer 2.1 – but you have to figure it out yourself.

That's the unpleasant part, and that's where implementations and the consulting firms that use Data Vault start to diverge, because suddenly it's down to folk creativity.

Richard Sklenařík (01:08:55): So yes, it would be good if this appeared. But as I say, it's perhaps at the level of those implementing regulations. Just like Patrick Cuba's book, The Data Vault Guru, where he made available templates for loading the Raw Data Vault, procedures for building the Business Vault and so on – things that weren't described in that much detail in the core 2.0 methodology.

Lukáš Mazánek (01:09:30): I can only agree, and I hope it makes it into the new methodology. Those implementing regulations – or, let's say, a more detailed methodology – are in my view a key prerequisite for wider adoption.

I was surprised myself at how far it is from reading Zhamak Dehghani's Data Mesh to building your first data product. There's a long list of questions along the way that the literature leaves unanswered. Along the way there really are lots of questions like: fine, we'll use knowledge graphs – but which ones? Will it be a business knowledge blueprint, or some other approach? Will we store the data in some particular kind of store, or something else?

There should be more of those specific questions about how to solve exactly this, and more best practices saying that in this case I should go this way. And something else struck me – again, I'll come back to the business side of data products.

Building the data products wasn't hard. What was hard was making sure they had business value.

And to give them business value, it turned out that three quarters of the time had to go into business questions and questions about the meaning of the data. Then it got programmed – sure, but that's the very end. That it has the right number of layers and matches the model – even the one Zhamak Dehghani proposed – that it has its planes: OK, fair enough. But for it to have value, I need to do something at the start.

Richard Sklenařík (01:11:14): We've been talking about AI and about Data Vault. One thing that hasn't been said, and that I'd like us to mention: Lukáš, you talked about AI and Data Vault in the sense that Data Vault fits perfectly into the concept as a data source for AI.

But on the other hand, we first have to build that data source – and that's something AI is, of course, used for today as well. I think that should be said here too. If you decide to use artificial intelligence to create the data model and implement the data platform, you have to decide which methodology the AI will build with. The easy answer is: any – you have AI, it can handle everything. Yes, fine, the huge public models can do it.

But here's the trick to remember: the Data Vault methodology is, as you mentioned, complete – and it also works with a very limited variety of building blocks: hubs, links, satellites, plus some specializations of these object types, and so on.

Richard Sklenařík (01:12:45): The same goes for the approach – how I need to analyze the data so that I can break it down like this, store it, and so on. And in my opinion that's a wonderful thing, because you can take your local, closed, small language model and train it on the Data Vault methodology – because a small language model can handle that, can't it?

Lukáš Mazánek (01:13:17): That's a great question. Vibe coding is trendy. Of course I vibe-code from the moment I get up until the moment I fall asleep – in between I go to work and to lunch. So I certainly have some experience with it.

With vibe coding and large language models, the problem isn't that you don't have a methodology available.

The problem is that you have a huge number of methodologies available. If you ask a language model to design a solution – a data-oriented one, say – it will typically combine several methodologies. So you get a mix of different approaches, none of them pure. I don't want to say it's bad, but there are lots of traps, and most people fall into them.

I always try to pick one specific methodology, stick to it, and keep my solutions as pure as possible with respect to it. And this is where, in my view, Data Vault has a huge advantage: as I say, it's fairly complete from top to bottom, and it lets me force the large language model to follow the methodology. Only when it identifies a gap is there room to make decisions of its own.

Lukáš Mazánek (01:14:49): I think this approach has a huge impact on the quality of the vibe-coded solutions that get delivered.

And then there's one more area I'll touch on briefly: having a language model trained – or training one yourself – on a specific methodology. And I agree, I think that's a way for larger organizations to approach solving their problems. A language model trained on specific methodologies, specific security patterns, specific solutions in specific languages makes it possible to make the delivered software dramatically more efficient and of much higher quality. So – great question. I think this is the future.

Wrap-up, upcoming events and outro

Martin Šafránek (01:16:06): Lukáš, Richard and I thank you very much for coming to the studio. To finish, we'd like to ask you to sum up the hour we've just spent debating in two or three points: what matters most for a bank right now, in the age of AI, very capable models and huge amounts of data? Where does its competitive advantage lie?

Lukáš Mazánek (01:16:36): First, thank you for the invitation. I love talking about data with data-positive people.

Second, I'll sum the question up for myself: what are, for me, the most important features of Data Vault, and why do I recommend using it?

First, the completeness of the methodology.

Second, how closely the methodology is tied to the business.

And third, the semantic description of data.

These three properties that Data Vault offers make it possible to use the full leverage of artificial intelligence. I believe artificial intelligence will change the world further, and if we want to be part of that, we need to give data meaning.

Martin Šafránek (01:17:41): Lukáš, thank you once again. Now I'll ask Richard to give our listeners, in two or three sentences, an outline of the Data Vault events coming up in the next two months.

Richard Sklenařík (01:17:56): Probably the most important is an invitation to the jungle. Don't be alarmed – it's an invitation to the Prague University of Economics and Business (VŠE). And no, it's not an invitation to exams during exam period. This time it really is a data jungle: Data Festival at VŠE takes place on 3 November this year, and Martin and I are planning, as every year, to have a small stand there representing Datavault Builder, one of the vendors of automation tools for building a Data Vault.

Then, also in November but abroad, for anyone interested: Scalefree holds, as every year, a one-day conference in Hannover at the end of November. It's a great networking opportunity.

There will of course be interesting talks too, no doubt about that, but in my opinion the networking has much higher added value, because you can watch the talks later from the recording – but the people you won't meet from a recording.

Richard Sklenařík (01:19:27): And there's one more event we're missing, at VŠE again, this time partly – or mostly – internal.

It's planned for October: our colleagues from Scalefree will be here again as part of a study program. I have to admit I don't remember the exact names, but as part of the master's program there will be a day devoted to Data Vault, within the various methods for building data warehouses and so on.

And the colleagues from Scalefree will be here for it. The event is primarily for students of that program at VŠE, but I think there will also be a chance to meet people who are involved with Data Vault here in the Czech Republic – so it will be a small gathering of the Czech Data Vault community.

Martin Šafránek (01:20:45): Exactly. And the date – if I'm looking at the right e-mail and the right year – is 22 October. If you'd like to attend, please write to me or Richard on LinkedIn and we'll try to arrange your participation. Gentlemen, today was really good, really rich, really demanding, and thank you for a great chat.

Lukáš Mazánek (01:21:10): Thank you for the invitation.

Richard Sklenařík (01:21:11): Thanks a lot.

Martin Šafránek (01:21:12): Take care.

Lukáš Mazánek (01:21:13): Goodbye.

Data Talk (01:21:17): Thank you for listening all the way to the end, and thanks also to our long-standing partners, the members of the Data Talk Club: Sazka, TV Nova, Direct Technologies, Good Data, Mython, Colors of Data, Bistreet, FLO, Carl Data Company and Intex. Thank you very much for your support – and may data be with you.

Resources mentioned in the episode

  • The Elephant in the Fridge – John Giles
  • The Data Vault Guru – Patrick Cuba
  • Data Mesh – Zhamak Dehghani
  • Domain-driven design (Eric Evans)
  • Data Vault 2.0 / 2.1 methodology and certification – Dan Linstedt; training by Scalefree
  • dbt model contracts; Data Vault packages for dbt: datavault4dbt (Scalefree) and AutomateDV (Business Thinking)
  • Datavault Builder – Data Vault automation tool
  • Data Dreamland – Scalefree conference, Hannover

Odebírejte Data Talk

Apple Podcasts Spotify Deezer Overcast Podcast Index RSS Feed