Live from NODES AI 2026: From Data to Knowledge to Action - The Graph Intelligence Platform

This talk explores the evolution of software engineering in relation to graph technology and the impact of artificial intelligence on application development. The speaker discusses how the human brain can be likened to graphs, and asserts that everything from infrastructure to individual interactions can be represented as interconnected data points. They outline five significant shifts in application design, including transitioning from static logic to dynamic reasoning, moving from simple request-response models to goal-oriented approaches, and the shift from monolithic systems to composable architectures. The importance of a knowledge layer for enterprise systems is emphasized, which includes metadata, domain data, and decision traces that inform agentic systems. The speaker emphasizes that as applications become more autonomous, understanding and leveraging graph structures will be crucial for improving decision-making processes.

[music] >> Hello everyone. I'm super excited to be at the first nodes AI conference and it's been an amazing day. Such great content. I think just kicking it off with the panel discussion was great. I do want to start by thanking all our speakers who had such amazing content to share with all of us. Of course, I had my own list of sessions that I loved. I

love customer stories. I love to hear how everybody is implementing graphs in real production workloads. So, the stories from Booking, Adobe, and then various other companies Cisco, The New York Times, as well as Bayer, Banque de France, and Daimler Truck were some of my favorites. Of course, I think everybody should share what their favorites were. And we are going to have all of these recordings available on

YouTube, so other others can listen to it on demand over a period of time. With that, you know, everything in the world is actually a graph. And then once you spend enough time in graph space, you realize that starting from human brain and how it is structured, how neurons interact with each other, how they store information, that's the reason we call our memories and you know, in

agentic systems that talk about memory and how memory should be depicted. Human brain is best example of how you store memory and how interactions happens and it is a graph. Take a step back. Look at our infrastructure. Look at the bridges and the streets and everything. It all itself is a graph. You know, like the way we depict it and all. The best way to represent your

city's infrastructure is actually in a graph form and you can do the shortest path and longest path and all of those things. And then if you take a step farther back and if you look at Earth from outside, especially with the mission to moon right now going on, this is so real. If you really look at Earth, you will see lights across the world, the way the

infrastructure works. It's clusters of information together. It also is a graph. It's all interconnected communities that exist to across the globe. And then take it from outside Earth to like individual people using their phones and the shopping journeys they have. What is the journey, whether it's a patient journey or a or a shopper's journey to buy things, it's nothing but a graph. It's it's like a tree

structure that allows you to go ahead and understand what people are doing, what they like, and all of that stuff. And the challenge historically has been all the stuff, all the information in the world looks more like a graph, all the organizations try to depict it in rows and columns in form of Excel sheets. I know Excel and and spreadsheets are easy for humans to understand sometimes,

so they try to put like this complex interconnected information in the world in these rows and columns. Actually, that's not the best representation. We all, the community of graph enthusiasts, realize that. We all know that most information in the world is so interconnected. The best way to represent it is in the graph format. And you know, additionally, along with all the information the way it is, the

world is also changing. And if you look at AI is like literally changing how we build applications, how the applications are going to be designed, and all. And and I think uh the most important thing in this is historically, the way we build applications was they were predefined workflows. Like I always say like we basically took human process, converted into visual interfaces, and then that was the

application that we built. But in the new world, that's not how applications are going to be built and run, especially with the power of large language models and agentic systems. Your applications are going to become more dynamic. They're going to be context aware. And they're going to be more autonomous. It's like going from cars that were all manual driven to automatic cars to now self-driving cars. And

the applications will also be coming become more self-driving. They will be autonomous. They will run on their themselves. And this big shift that is coming with these autonomous applications is is there are five key characteristics of this. Number one is you're going from static logic to dynamic reasoning. What does that mean? We had hardcoded logic in our applications that drove all the business decisions. Now we're changing

that. We are actually moving into more dynamic reasoning where large language models with the reasoning capability can take inputs, take what task they want to complete and use that particular action that needs to happen and use the information it has, the tools it has, and then make that decision. Same way, if you look at the the applications are moving from this request response to goal-oriented design. What

that means is historically we have been like put some data in, you get some information back. As a human, it is great, but in the self-driving or autonomous systems, what you want is no, I need to go from point A to point B. Now figure out what the paths look like. Same way, I want to detect fraud. Here are the characteristics of fraud that we as as

an organization care about. Now go figure out how to identify fraud. So, that's what is the key thing is it goes from request response to goal-oriented. The next big thing that will happen is you go from monolithic to composable systems. You basically, instead of building one tightly coupled system, you basically build tools, capabilities, skills that a reasoning model can take and then automatically leverage that to go

ahead and build the the application logic dynamically. And then finally, the other thing which is key is you go from code-centric, which is everything being in code, to prompt-driven. What I mean by prompt-driven is you will basically be able to give some general prompt or parameters that you're interested in and let the system figure out how to go ahead and make all of the decisions happen with

all the information it has accessible to it. And then finally, it all is going to move from CRUD, which is create, update, delete kind of a like you know, application, into context to context-driven applications. Basically, the logic is set by the reasoning goals and rich connected context that you have. So, if you have an agent you're building, you give it all the context it needs. If you

are building an agent for like you know, your web experience, you want to give it all the information about the users that are coming in, what they have done, how they have done it, anything interactions that have happened with that particular user in the recent past, you have to provide all of that context. And then that context becomes the input for the system to be driven. So,

I think that's how you should think about the key application changes that are coming. So, in short, software engineering is going to completely change and it goes from in five ways, right? One is going from static logic to The second is going from request response to goal-oriented applications. You go from monolithic, like you know, big application that you're building to composable system or composable capabilities that reasoning

models can bring everything together. You go from code-centric to prompt-driven. And finally, you go from CRUD applications to like context where context drives And so, if you think about all of these key constraints and if you want to go from this predefined workflow and simple data models to this dynamic context-aware applications, what do you need? What you need is knowledge. You know, basically, as a context for

the agents, what you want to provide is knowledge that your enterprise has. If you are building like you know, very generic system that can be built on top of like all the knowledge on the web, it's very easy. You just go and ask your like you Claude or or Gemini any questions, it will take all the information on the web and give you the answers and those

systems are going to be pretty simple. But in an enterprise environment where your IP is your information and knowledge within the organization, you want to represent all of that information in a way that agents can actually benefit from or your intelligent applications can benefit from or your tools can leverage. So, if that is the case, the knowledge layer, which basically brings all your information, which is mostly

connected data, to those agents is super critical. And there are three broad areas or components of this knowledge layer the way we think about. Number one is ontologies. It's mostly metadata, metadata that gets interconnected information about your domain data, where it is located, the full semantic map of your whole organization. Which means if your data is in Databricks or Snowflake or any other system, where is the

data, how it's interconnected, what tables do what, what decisions are made by which system, so you can basically make any decision based on this definition that you have within your environment. The second is access to your domain data or data, which means it could be like just the reference data, it could be the full domain data. We have organizations I will give you examples of that they

are putting all of their data into graph and then making it accessible. In other cases, you may have just information that may be in other systems and you can basically understand where it is and access that. And the third area that is really important in agentic systems as enterprise knowledge is all the interactions and engagement that's already happened with the agentic systems, which includes your long-term and

short-term memory, which includes your decision traces. So, we like there's a lot of talk we already had today about context graphs. So, context graphs are nothing but decision traces. How are decisions made? What were the policies that were applied for those decisions? What were the overrides that happened? So, that's the other information in the knowledge layer you need. And then finally, what are the business rules you

want to apply for any kind of an action you want to take. Final goal of the the agentic system is to make decisions, take action, and that action requires some boundaries to go ahead and do that. And that is where what you provide as a as a full context in form of a knowledge layer for your And then at Neo4j, I think our main goal is how

do we make it super easy for you to build this knowledge layer to power your agentic systems. And we have been investing heavily into different tiers. So, if you think about our whole platform, our goal is to go ahead and give you a tiered architecture, which gives you flexibility about components you want to compose together to build your agentic platform. So, at the lowest level we have

is our core database and analytics capability, which is the core database you all have loved. 15 16 years of innovation, everything going from like you know, replicated graph for high availability. You can take the database, cluster it into three machines, and you can get high availability to like fabric where you can have different graphs, and you can query them as if it was a single graph, or

going into infinigraph, which is like like unlimited scale for your graphs like you know, from tens of terabytes of size. And then like you know, things that we are working on for virtual graphs and all that. The second tier is AI-powered tool. Just the database is not enough, right? The next thing you want is you want to be able to go ahead and use different tools to

get your data in. So, whether you want to query the data, you want to visualize and analyze the data, you want dashboards. All of these are now AI-powered. So, if you use this in Aura now, you can just say, "Hey, give me a dashboard that is focused on this particular information like patient journey." It will create all different like you know, visualizations for you out of the

box. Similarly, data import is an important thing. You want to be able to get data in. Document intelligence is something we are working on, which will give you ability to take unstructured data, create a data model, and ingest it. And then finally, ontology management or like you know, models model management and all of that is basically something that that we are investing heavily in right now. So,

that's the AI-powered tooling layer. And then at the top of it is graph AI, which is with Aura agents. I'm super excited about Aura agents. You all may have looked at Aura agents session that Ed did. If you haven't, please go ahead and listen to it in on-demand version of it. We are super excited about it. We are also going to announce like you know, something after

this session that Steve is going to talk about around Aura agents. Super excited about that. But Aura agents allows you to build agents in minutes once your data is in the We also are integrating our capabilities into different agentic ecosystems, whether it's AWS, GCP, whether it is Azure, or whether it is LangChain and the various different open source capabilities and and communities. We integrate ourselves into all

of these ecosystems. So, if you choose to use a different framework agentic platform to build your agentic systems, you will have Neo4j next to you to like you know, build these systems seamlessly in general. And then finally, we are doing a lot of innovation with memory and context graphs. If you haven't seen our package that we have launched in in Neo4j Labs to build your own memory,

there's great demo that Will has built on context graph. Take take a look at it, and we are trying to make it super easy for you to build all of these agentic systems based on Neo4j. So, that's how we think about our investments. Along with that, there are five key benefits that our customers are seeing by using Neo4j and graphs in the agentic AI systems or agentic

AI platforms. Number graph rank by far is better than just traditional rank. If you are doing I think Intuit speaker said it best when I was at the Seattle Summit. His thing was vectors give you similar content, similarity searches, versus graphs give you related content. And so, when you can combine your related and similar information, that's when you get the most accuracy. And that's why graph rank

has been performing outperforming traditional rank by almost 3x in many of the studies that we have seen. The second thing graphs are great at and providing a lot of value is the semantic map. So, you can use Neo4j as a map of where your information is, how it is interconnected, what decisions can be made by which which system. The third, memory is a graph problem. As I've

said before, the way human brain keeps like I went for a vacation. I went there with my family. We went and saw three shows. I know exactly the three shows that I have. Then we went to see Columbia University. This was New York. So, we went to Columbia University. If you think about what I remember from my vacation are these key entities and how they are related

and whom I went with. I don't remember a reel of 5 days. And that's how our brains work, and that's exactly how graphs work. Like what are the key entities and their relationships and what you need to remember, and how can you like you know, revisit them whenever you want to make decisions. So, I think memory is a graph problem. Then context graph, decision traces. This This

is all interconnected decisions that are happening in enterprise. So, that's also a graph thing. And then finally, you want to have multi-hop reasoning. Every complex problem or even simple problems requires multiple hops. Like a customer that has come and complained about something. They recently bought a product where they were shipped the product. The shipping has not arrived even after the deadline has passed. Like all these are

interconnected scenarios. And in multi-hop reasoning, you should be able to go ahead and reason through multiple hops, and graphs are best way for those kind of traversal queries. And finally, if you think about your enterprise architecture, which all of us are developers, we are building applications. We are in our world, but you have to fit into the whole enterprise where it is. And the knowledge layer powered

by our Neo4j's graph intelligence platform is the backbone for the agentic intelligent applications or agent agents or tools that you're building for these agentic And then it will interact with your orchestration layer on the left, and you can use any of the standard orchestration platforms. And your data may be coming from various different platforms, whether it's databases, whether it is lake houses, whether it is coming from

applications and SaaS products and all. So, I think bringing it all together through this graph layer actually is going to be a game-changer, because the agents require all of that information interconnected, easily to access. And then finally, what we are seeing is this three very common architectural patterns. The first one is the semantic map or semantic layer pattern I call it, wherein you're using the knowledge graph

for understanding where what data is. In this case, EA, they have Snowflake, they have basically put all their key tables, where they are, how they are interconnected with other tables, that information in the knowledge graph, but the data has not been moved. So, it's zero copy architecture, where you can use reasoning model to then go ahead and create the queries and query information The second is where

you have the semantic map along with some domain data. So, many of our customers, in Rocket Mortgage's case it is, a lot of interactions that have happened with the customer on on different chat experiences and SMS messages and all, combined with the Salesforce marketing cloud information about the customer. They bring all of it together, put it into Neo4j, but some of the information may be in like

you know, existing systems. Some of it is in graph, and they bring it all together to build the agentic experience on top of it. So, you have a combination of it's a hybrid approach where some data lives in the existing platforms, and some data actually lives in in the knowledge layer. And the final, the third is where it's like, "Hey, the organization believes are for that use

case you believe that that's the graphs are the right way of representing information, and you move whatever data is required for that agentic system into graph, and then graph becomes the primary store. You can call it context store, you can call it context graph, you can call it knowledge graph, whatever term you use, but you have all the information for your agent directly into it. So, that's

the patterns we are seeing. So, in short, I think AI is going to change how we build applications. It's actually going to give a huge amount of opportunity for us to leverage graphs as the data model for agents because it's all interconnected data that you have that is required for agents to make right decisions. And so, I think we have a huge opportunity as a community to

help every organization to be successful with their agentic deployments. I'm super excited about the whole day, and I want to pass it I first want to thank everybody for attending. I know we have thousands of people attending across the globe, and I'm super excited for that. And with that, I do want to pass it on to Steven Chin to share some of the key thoughts he has.

Thank you. >> [music]