GraphRAG: Driving Agentic AI Value for the Enterprise with Nikolaos Vasiloglou

This is Textron tv. Hey everyone. Welcome back here to Textron tv.

Yeah. I got a first time guest here for you. His name is, uh, Nicolo or Nick VA Glue.

I hope we got that right. Nick is the VP of research for ML at Relational ai. And we're gonna talk a little bit about, of course, ai.

'cause we don't spend enough time talking about AI as it is. M welcome to techron tv. It's a pleasure to have you on here, my friend.

How are you? I'm fine. Thank you, Allah, for hosting me today and happy new year.

Happy New Year to you. Um, Nick, before we jump into things like graph Rag and and relational ai, let's hear a little bit about Nicholas's story. Um, if you wouldn't mind share with our audience, kind of, you know, I said you're the VP of research ml, but tell us your kind of path to here.

Yeah. So I came to the US uh, uh, probably 20 years ago. More than 20 years ago.

I did my PhD at Georgia Tech. And when I was, uh, graduating back in 2009, uh, that term, big data and machine machine learning was becoming very hot. That was exactly what I did in my PhD.

So, um, I started working as a machine learning software engineer, started a small company and I worked with a bigger startup called Logic Lock those days. Um, and we did, uh, very scalable library in CPL plus plus about a machine learning. We were calling that, or statistical learning those days, which we did apply a lot to, uh, retail, you know, building demand forecasting systems.

Uh, and then I moved on to security where we applied graph techniques, uh, very successfully on, uh, botnet detection. We have a trilogy of papers, kind of like, uh, part of the rigs, uh, where we, um, we kind of saw how you can use machine learning. We discovered some boats that they were, uh, unknown before.

Uh, really cool stuff. Um, and then, uh, um, my startup got a acquired by, uh, logic Lock, and then Logic got acquired by Infor. And then after that acquisition, we started the company with, uh, uh, a relational ai.

So it's been a journey of about 15, 16 years now with, uh, the people from, uh, uh, what was then called Lolo in our relational ai. Of course, many have joined, um, working on something, uh, that, uh, started as machine learning and, uh, now we know it as, uh, uh, ai. Um, also wanted to highlight that during my journey, I spent a lot of time, uh, publishing at conferences, but also organizing.

I started a company, uh, conference called ML Conf, which was kind of a big hit for, uh, uh, industry. And, uh, I just came back from Europes, which is the, uh, biggest, uh, conference of ai. This is everything you hear about AI in all the cheering awards and Nobel Awards.

They all came from that community, which I had. Uh, really, I had the pleasure to, to be part, you know, I wasn't the smartest, obviously. There, there are other people smarter than me.

You Never wanna be the smartest. There's the lesson I learned in business. Never be the smartest person there.

'cause you'll never learn anything that way. Right. Always better to surround yourself with smarter people, and that's how you learn.

But, uh, interesting. Excellent. What was the name of that community, by the way again?

It's called, uh, it's called Res. It's started, it's about 35 years old. They started as nips, but then they remained into Europes.

It's Neuro Information Processing Systems. It's a conference that, uh, competes with Taylor Swift because they sell out, competes who is gonna sell out first, you know, when they announce the conference, it was good. It can sell out in minutes.

So Burning Man, really, Taylor Swift and Europes sell out very quickly. I hope that tickets don't cost as much as the Taylor Swift tickets, but where, where was the, the most recent conference? Where was it held?

It was in Vancouver. It was in Vancouver. Ah, That's beautiful there.

Good. So you heard it here first. Thanks for that tip Nip.

I, Nick, I'm gonna keep my eyes open on that one. Um, so you, you know, tied into your journey, of course, was this journey that led to relational ai, but for people out here now, okay, they, they heard what you said, but maybe they're still not sure exactly what it is relational AI does and yes. Problems it solves.

What, how, what would you tell 'em? Uh, you know, it's kind of interesting. We, we started with a temporary name as relational ai.

We spent, you know, we're doing AI in the relational world, but then we liked it, people understood it, and we kept it. So just very quickly, um, I like, uh, uh, there, there's several definitions of what we're doing. I think the, the thing that people would understand the most is the digital twin.

First of all, we are, uh, a native app on, uh, on Snowflake, which means that if you have your data on Snowflake, you just turn on, you know, you just go to the marketplace, find relational ai, we're highlighted, and you turn that app on and you can immediately start using the app. It's, it's a very quick and efficient way. Uh, sometimes we call ourselves the knowledge graph co-processor of a snowflake as just to explain to people just, you know, so if you don't know what a knowledge graph is, if you don't know, um, you know, even what AI is, um, I wanna say, like, I give you a simple example.

I think it's better to understand, for people to understand, you know, if I was asking any, any of our friends here today to look at their, uh, enterprise database and explain their business, it will be extremely hard the way that you store the data in a, in a, in your enterprise database. It serves multiple purposes. But one of the purposes that it doesn't serve is, you know, having an overview, an explanation of your business.

You know, it doesn't have concepts, it doesn't have, you know, it doesn't show. People are understand diagrams. They understand concepts.

You know, if I have a supply, if I'm a supply chain, a CPG company, like Broker Gamble, somewhere like that, I have, you know, vendors, I have factories, I have, uh, transportation, uh, I have warehouses. And people try, you know, they, they like drawing arrows and explaining their businesses as a big graph. This is what you can do if you have relational ai.

And the nice thing is you tie that on your enterprise database. So you don't, you don't move to a different system. You go to your database and you create this nice, let's call it relational modeling.

People call it knowledge graphs. There's different names for that where you describe your business. Now, once you have it like that, you can do magic because that diagram and that kind of type of modeling can help you run optimization problems like, you know, optimize your supply chain.

'cause you just set it constraints that this can only go here, it cannot go there, and I can only see 20 packages from here to there in a week. All the, all this expression of constraints. You can do that naturally in our system.

And then if you want to do ai, you wanna start asking questions and have, you know, a language model generating your business analytics and all that stuff. We do know these days that the best way to do that is by having an knowledge graph. So all the effort you do to model your business as an knowledge graph then, you know, automatically translates to an, uh, uh, uh, a planner.

It can translate to a language model that takes that and generates ways for you and a lot of other stuff. Like you can, you can, because we're gonna talk about that later. You can build agents, you can, you can do a lot of interesting things.

So, uh, you know, if I had to summarize, you know, native on Snowflake, you don't have to move your data. You don't have to worry about security parameters. Easy way to model your business, build a digital twin, and then build intelligence applications with the model that you got.

I think that's the most important aspect. Absolutely. Look, you know, we, we covered graph, graph databases, knowledge graphs and stuff, a, a fair amount here on Textron through, whether it's on the videos or in articles or, or what have you.

Another thing we've been covering, because it's become a very hot term, is the term rag, right? And now, you know, part of, uh, relational AI is, is combining graph rack. Graph rack and how that helps address AI limitations.

Make us smarter a little. Nick, I I said before, right? I don't like to be the smartest person.

You're smarter than me on this. Go ahead. Talk to me.

What do we mean by graph rack? Yeah, so, uh, you know, let's start with the problem. I mean, the, the, the beauty about Gen AI was that we can, people want to ask questions and get answers, okay?

And the question could be as simple as what's the capital of Greece? Or, uh, prove, how do I prove the firm theory one is like mm-hmm. Uh, one, uh, one word.

The other is 200 pages of, of complicated math. Okay? So this is the problem we're talking to.

I want, I have a question and I need an answer. Okay? So, uh, the, when you have a single point question, which is, you know, when was Apple founded?

Who was Alexander the Great Sister? Something like that. It's, it, it boils down to a retrieval.

And we don't, this is called rug retrieval or, uh, augmented generation, which means that you take your documents, you assume that your knowledge is in documents. That's the other thing. Like where's my knowledge?

Is my knowledge in documents? Or is it in a database? Is it on images?

Where is it? Okay? When your knowledge is not in the database, you use something called you know, rug, which means that I take my documents, I use a term called vectorization, which means that I collapse them to a vector, you know, a set of numbers.

And, uh, and then I retrieve the most relevant stuff. I put them on a language model and it gives me an answer. It's plenty of articles.

It's extremely simple. You can code it in three lines. You know, it's, it's very easy.

Even if you're not technical, you can do that. Okay? Now, there are some other questions that I call them multipoint.

Okay? So, which is, has Bill Gates ever worked at the same company with Steve Jobs? You know, uh, it could also be, let me give you a more complicated one.

So I want to invest in the stock market. Give me a stock that if I invest, it's gonna help Rwanda. Okay?

How do I answer that? Now, what happens is that in, in the typical drug, the first one that I described before, it's more like you, you treat your information, your documents as, as a pile of, you know, you stuck them, you know, it's a pile of books. Like you, you, you set them, but you, you don't exploit any relations between them.

Okay? And I wanna take people back, uh, the, the older, uh, uh, people from our audience back in the nineties where the web came out and you wanted to do a search and you were using Alta Vista, okay? And, uh, people might not remember how difficult it was.

'cause Alta Vista and all these searcher senses, they were just doing a keyword search. They were trying to find WebP pages that they matched the, they weren't, you know, they weren't very successful until Google came. And what Google did said, well, you know what?

The web is not just documents, web pages. It is also, you know, they have links. They're linked together, okay?

The hyperlinks. And they took advantage of that. And voila, here we have Google, okay, so, and no one knows Al Vista no more.

No figure. I loved Alavita, but that's another story. Go ahead.

Yeah. Anyway, they were pioneers for their time. 'cause there were of course, other problems except for just the, the two about.

Sure. Now, so now graph of what it does, it says, well, you got all these documents, but before you ask a question, can I try and create links between them? Okay.

So you use the language model. Basically, I want describe the process. It's an it process to describe links between documents that they didn't exist before.

So if you find Bill Gates in this document and you find him in another document, you know that these two documents are connected. And if there's another document that talks about, you know, uh, you know, just to give you an answer, bill Gates has still jobs never worked at the same company, but there's other least, uh, where can happen, you know? So you start extracting edits and relations and, and basically you build something which is, it's not exactly a knowledge graph because it's not perfect.

So let's call it the graph of knowledge. It's a graph where different pieces of knowledge are. So what happens is that when you are asking the question, you know, you are able to retrieve pieces that they can be in completely different documents.

And let me give you the, let me answer the example of, uh, uh, of, uh, stock market in Rwanda. Okay? So if you take documents out of the web, you will find that, uh, an NASDAQ stock is Microsoft.

Okay? Now we know that, uh, uh, uh, bill Gates sits on the board, okay? And we know that, uh, bill Gates also has the Bill Gates Foundation and the Bill Gates Foundation invest in malaria.

And malaria is the prominent disease in, in, in Rwanda, okay, rda. So if, if, if you, if you wanted to, to solve that, to ask that answer, you would have to navigate over a graph and give that, okay? So basically, if the query has the term Nasdaq and Rwanda, and we know that there is a path that connects them, okay?

So now we can navigate that graph and answer that question. But of course, there's another case, which is what if my documents and my information is not into documents, but I have a relational database, you know, your typical enterprise database, that basically what graph, I guess it boils down, it boils down to, uh, you know, taking a natural language question and translating it to a formal query language, whether this is SQL or this is zq, well, or this is the language that we use or sparkle and it answers the question. So that's another way of, of, you know, of thinking about graph.

Uh, so I hope I didn't confuse the audience, so No, no, I, you know what, that was illuminating. Amazing. Thank you, Nick.

You know what I'm sitting here thinking, I don't think we told people the website. Is it relational ai? Relational ai, Relational ai, Tom?

Uh, I think they both work. I think both work. com.

Um, uh, yes, using the relation ai, yes, but you can, as I said, you can always find us through Snowflake as well. Absolutely. So, you know, we we're there, it's time of year.

You're doing your year in review, your year looking forward, you know, this type of thing. And certainly 2024, like 2023 for that matter, a uh, generative AI was, was all the rage, right? That was the big story.

A lot of people say the big story this coming year though will be agen ai, right? Because now we're moving in agentic AI and then also physical AI robots and, and all of these things. Um, how early are we?

Right? How I've been an entrepreneur many, many years. There's, I always wasn't always our media, I it startups.

And the lesson I learned is if you just have a nice to have product, you're not gonna be successful. You gotta have a must have, must have products make money. Is graph rag must have at this point, or will it be must have in the near future?

I think it's becoming, I think it's maturing. Uh, it has matured enough and we, we do have some, uh, uh, indications in the market that, uh, uh, people want it more and more. Um, but I wanna tie that to, to what you called about EnTec ai.

Okay. If that's, that's fine. 'cause they, they're actually, yeah, no, go ahead.

Connected. They're actually connected. So as I said, I came back from URIs and 90% of the conference was Azen ai, even, uh, one of the key players from Open AI who got the Test of Time award, this is an award that you get, he got for a paper he published 10 years ago.

And it turns out, it turned out to be very influential, which is what OpenAI is doing. I Suki, uh, he talked about that the future, like the oh one model is, you know, the is is a way of doing genetic ai. And, and many, many other people talked about that.

Uh, I'm actually preparing, uh, a, a, a review of the conference, which I'm gonna, um, talk more about that. Um, so let's, can, can we define what an agent is? Okay.

Uh, if I, if I may go ahead. So The best way to understand that is if we follow the, the very famous paradigm from Daniel Kaman, the, uh, the noble, uh, Laureate, uh, uh, in economics who Coursely passed away, uh, this year, we, we do have a blog post, uh, explaining, uh, his theory through the, the lenses of, uh, uh, advent AI on our website. Um, so we have the system one, that's a system two thinking, you know, uh, if, if, if you know, a rabbit crosses the road as you, you are, uh, driving, you know, you immediately hit the brakes.

You don't think about it, oh, there's a rabbit. Uh, I might get, I might kill it. You know, it's something that happens instantly.

And then when I ask you to prove the, uh, Ethereum a pgo Ethereum or solve a task, you know, assemble a a, a furniture, this is the system too. Like, you have to understand where I'm gonna, like, you start planning, you start trial. You, you know, trying something, failing, trying again, like it's a search process until you get there.

Okay? This is the Slow, there's a famous book called by Daniel Kahneman, thinking Fast and Slow. Now in, in practice in Ingen ai, the language models, in the beginning, they were trained to answer a question immediately.

You ask something, it starts generating, and you expect this to be correct, or you expect to be as, as, uh, as complete as possible. So that's kind of like the system one thinking immediately, I give you a que, I ask a question, five plus five, you say 10, you don't analyze it, you know how to do it now. And as they kept, you know, increasing the size and the chaining data, um, it was becoming better and better, but then it saturated.

And then they realized that, well, what if I ask the language model several times and it gives me an answer, then I correct it. And I say, well, uh, I think you made a mistake there. Try again and try again.

And this is kind of like the people, you might, people might heard of that as chain of thoughts, which basically is, I'm trying to solve something, but in iterations and I'm searching different directions. So the language model using itself as a tool is, in my opinion, the definition of an agent. You know, uh, and it starts searching, like when it plays tests, uh, you know, uh, it's smart enough to think of some moves, but it tries, you know, several thousands of different moves.

Uh, and it might even play a small game between itself for some time until it gives you a massive, now the importance of that AAD is, is huge. I'm, I'm gonna give you a reference from a, a talk. Um, when a researcher was trying to solve the poker game, he saw his journey.

He says, I kept training with more and more data and I was getting to know my, um, better results. But I realized that if I was giving to my model 20 seconds to think the results were going through the roof. And he says, by using 20 seconds to think and to do this adjunct thinking, you are scaling a hundred thousand x compared to training my model with, you know, giving it more and more games, uh, past games to get trained.

So we realized that by giving more time for the model to search, we can actually, uh, solve much harder problems. Absolutely. And I, I, you know, so I use open ai, GPTI, and I've been using the 1 0 1 model, um, and you, and you're right, I see it's telling you what it's doing as a thing.

It's you, you actually do see that. And I see a big difference. Someone told me oh three is coming out now.

Um, I'm not, I haven't seen it yet, but, we'll, we'll see. Anyway, Nick, we're way over time, but I'd lo I find this fascinating and I could sit and listen all day. We're gonna invite you back for more.

You know, you're going to be, have to become one of our go-to smart people here around ai. And, uh, I want to thank you. I want to wish you tremendous success with relational ai.

Ai, it sounds like a fantastic concept and graph rags, and as this becomes bigger and bigger, thanks for joining us today. I appreciate it. Thanks for everything.

Thanks for, we'll see you soon. Nicholas Siglo, VP of research ML at Relational AI here on Text Drug tv. We're gonna take a break.

We'll be back in just a minute.