NODES AI 2026 - How to Build the Next Wave of Graph-Based Agentic AI

This talk presents a practical approach to building graph-based agents using the Letria platform, a French company specializing in transforming unstructured data into actionable knowledge graphs. The speaker details the process of creating these agents to ensure compliance with the EU AI Act, which contains extensive regulations that can be challenging to navigate. He explains how to utilize OrientDB and ontologies to develop a knowledge graph that automatically assesses documents for compliance. The speaker emphasizes the importance of using a well-structured ontology as the foundation for accurate knowledge representation and discusses the automation of ontology generation from complex data sets. Additionally, he highlights the integration of natural language interfaces with the graph, enabling users to query and analyze compliance more intuitively.

[music] >> So, hello everyone. My name is Charles. I'm one of the co-founders of Letria. We are a French-based company located in Paris and today I'm going to give you a very practical talk on how to build a graph-based agents and how to of course power your agent with the best context leveraging graphs. Okay? So, I wanted to keep it super practical. We're going to go through

the entire steps from, you know, pure knowledge raw text to agentic reasoning. And I'm going to guide you through the entire step and of course at some point I will be leveraging part of our Letria infrastructure. So, just a bit of introduction about the company. So, we are a company specialized in helping organization turn all the unstructured data into actionable knowledge graphs and nurture these knowledge graphs

to agents. So, we all know that and I'm pretty sure that everyone in this virtual room is convinced that graph-based agents have a bright future and will definitely have better accuracy and better explainability than what we currently deploy at the vector-based agents. And so, right now I want to show you how we're going to build an agent to assess how your company, how your providers are compliant

to the EU AI Act. So, I don't know if you are all familiar with this regulation, but it's a very long regulation, hundred thousands of pages are containing a lot of different, you know, obligations, a lot of different compliance rules. And so, this is a nightmare for legal people. This is a nightmare for any CEO, any founder to, you know, make sure that each time you you

buy from a provider, you have a provider that it's abiding by the rules of the EU AI Act. So, there's a way. Of course, the traditional way would be >> [clears throat] >> chunk this entire AI Act, to vectorize it and to try and play with this vector space, but definitely what I want to show you is a very practical way to leverage OrientDB, to leverage ontologies,

to leverage knowledge graph and to use these very exciting pieces of technology to power your graph. Okay? So, I'm going to guide you going from this raw data to active agents in production that is of course giving a lot of value. So, right now I want to talk about the data. We all know that raw data are a mess, super complicated. We all know that we need

to, you know, context windows are not wide enough to integrate those hundred thousands of pages. So, this is where it starts and this is where we're going to kick off. So, at Letria, we will provide you with the ability to pass this data, very complex PDFs and to create the first layer, which is the ontology. Those ontologies are super important. Basically, if you go into our console,

you have a free tier and I will definitely give you a way to access this console and and use it freely. You have the ability to upload your files. So, this is the entire EU AI Act regulation that has been chunked into several you know, pieces. And I'm going to build an ontology in one click. Usually, ontologies would take you months to build, especially for a very

complex legal use case. Here, I am able to, you know, go and get either pieces or the entire EU AI Act regulation. Then I will of course power the use case that will just guide the entire ontology generation. Ontology will be the backbone of the knowledge graph. Those ontologies are becoming super super prominent because we've understood that we need that kind of schema to guide the graph

population, to make sure that we have the right classes, the right properties that have been vetted by the humans, by the expert, the legal expert here. And that we will have a graph that is consistent. And if I am to infer tens of thousands of pages and run it through this knowledge graph, then I am sure that the ontology will act as a bouncer to make sure

that everything that is entering the the club, the knowledge graph club, is fully compliant to what we already validated. Okay? So, here the use case would be basically I want to build an agent that can analyze technical documentation and provide an assessment on EU AI compliance. I can then play with, you know, the languages of my ontologies and then I can launch the ontology generation. So, this

will be a process that can take from a few minutes to a few hours depending on the size of your ontology and to shortcut it, I have already pre-built this ontology and you can visualize this ontology. So, this would be a total file that we output and of course I can show you a very visual way to assess the ontology. So, I'm using WebProtege, which is the

hosted version of Protege, which is a very famous and open source ontology visualization tool. And here you can see that all the >> that have been emulated are really precise. And those classes are typically specifically dedicated to the use case, which is EU AI assessment, making sure through an agent that all my providers are fully compliant. Okay? So, here you can see that I have a a

lot of different classes. Oh, sorry, I need to log again. I have a lot of different classes and if I click onto the class, I will have both the definition with also the parent categories, with also all the potential relationships. Okay? So, these are for the classes. These are for the properties. If I click onto a property, then I have a data property here, which is has

deadline and it was linking to a class, which is regulatory instrument and ranging through date. So, I'm expecting a very normalized way, normalized data point, which is a date. I will expect a duration. Here I might expect some tokens also. I don't know, maybe integers or percentage. It really depends on on this, but this will give the entire structure of the graph. Okay? So, once the ontology

has been validated, then of course I want to infer the data. So, I can take part of EU AI Act regulation or anything related to the EU AI Act regulation and I can build a graph. So, I'm going to select a file or multiple files. I'm going to select my generated ontologies and then I'm going to launch the build of my ontology. Of my knowledge graph, sorry.

And then of course, I will have a very interesting But what's super interesting is that 100% of the nodes and edges that have been populated are fully compliant by the the ontology. And just so you know, we have published a leaderboard on all the models and how they are reliable on the text-to-graph task, which is not an easy task. So, we have benchmarked all the models, all

the commercial frontier models, but we have also fine-tuned our very own smaller models. And so, if you go to our website, then I'll I'll send you some um URL to access it. Then you will have a live leaderboard on how those models are behaving in terms of text-to-graphs. And you can see that you have very, very astonishing results. For example, Claude 4.5 Sonnet is really good in

out at output reliability, so compliance to the ontology. At the same time, if you check 4.6, which is the newest model, you can see that it's really, really bad if you add some specific thinking like Claude Opus 4.6, which is one of the best model in terms of reasoning, is really bad at text-to-graph. So, again, please ask us for advice when you launch this text-to-graph because the

model that you use behind will have a major impact on the quality of your graph. But here I have a graph that has been fully automatically enriched, that is fully compliant to my ontology and that it can then of course upload onto and plug onto my OrientDB. So, here I am on the Neo4j OrientDB platform, the cloud platform, and I have imported my entire graph. You can

see this is the same exact graph. But the interesting part, of course, is that I can leverage all the cool features of the OrientDB agentic platform. Here I can create an agent in one click or in one prompt. I can build an agent saying, "Hey, I want to perform EU AI Act assessment making sure that the documents from my providers are fully compliant." You can just type

everything you want and it will automatically create an agent. Here, the agents will will be fully configured with a prompt instruction that has been automatically generated with all the agent tools that have been automatically of course set up. So, this is a great, great way you know, create a natural language interface on top of your knowledge graph. So, we all know that it's pretty complicated to apply

those Cypher queries. Not only Not everyone is able to generate very powerful Cypher queries. So, here you have a way to very easily query the graph that we built in one click with the Letria platform. So, I'm going to ask a question which is list all the fines when you don't comply with the EU AI regulation. So, I want to, you know, make sure that I'm fully

aware of the fines, of the you know, penalties when you have some kind of infringement. And here you will broken down all the tooling and all the steps to query the knowledge graph. So, it took a bit more than 10 seconds. I can check, of course, very easily what we applied, the type of agent tool that we called They called the natural language to Cypher tool which

is there. And then they also give me the query, the specific query. So, it went from one natural language query to one Cypher query and then with the reasoning it applied this and I have an answer in natural language that can that is very well-structured and can help me chat with my knowledge graph very easily. So, just created a graph rag agent in a matter of a

click. This graph rag agent is ontology-based. The knowledge graph behind is automatically enriched from text and it's fully compatible with the ORADB platform. Okay? But I want to go further. I want to go further. Here it's a conversational interface. This is great, but I'm pretty sure that we all want to way more complex interaction with the knowledge graph. Talking to a graph is one interesting way to

interact, but there are many other UX that we can implement. And here I have enabled a very, very cool feature of the ORADB platform. I have made this agent externally accessible. Which means that I can query it from third parties. And also, what I did is that I enabled MCP server. So, basically I created an MCP server very easily using the ORADB Okay? So, imagine I can

call this agentic graph rag from other tools. Okay? So, I can apply both natural language interface, but I can also apply other very interesting stuff. So, here what I decided to do is to use the cloud interface So, with a free plan. So, again, everything is free here. You don't have any, you know, vendor lock-in. You can recreate this on your own for absolutely zero dollar. Here

what I did is that I added a setting which is a connector. And the connector is the Neo4j MCP. Okay? So, I just copied the MCP URL to make sure that I'm fully, related. And here what I can do is basically I can chat with this graph. I can leverage this graph, but in a very interesting and very intuitive way. Let me show you an example. For

example, I can drag and drop any of my technical documentation, so my company's documentation or my provider documentation. And in a few seconds I can make sure that I will have a gap analysis and make sure if I'm fully compliant myself, my company, or if my providers are fully compliant. So, I'm going to make a bid on this and I I want to take the the opportunity

to run this analysis on the Neo4j cloud agreement, guys. So, of no worries. I've already checked and and Neo4j legal team has done a a great job. You guys are are fully compliant, but for the sake of this demo I just wanted to show you how interesting those motion can be like I just can save this agreement in HTML. So, I really, really raw data. Like I

I just in one click extracted the information from the website. Okay? And I can just drag and drop this cloud agreement. And then, of course, I can just say So, use So, I can be, you know, in a bit a bit I can directly tell it to use the the the tool. Or I can be super allusive. Just, you >> if this provider is compliant Again, the

EU AI Act is a series of rules that are all interconnected. And of course, if you think that you can chunk these thousands of pages, put it in the vector space and just drag and drop this HTML file and get reliable, explainable results, of course, we all know that it's asking for very complex reasoning, that it's asking for dependencies through third parties, through amendments. And here the

graph is playing an amazing job of making sure that they broke down all of the rules, all of the articles, stored the metadata of the AI Act and then can give you a fully activable brain on the EU AI Act. So, here I'm going to launch this analysis just to show you guys how it's going to perform this, but of course, it can take up to a

few minutes sometimes. So, I've already preloaded this, but what's interesting is that first it will do the heavy lifting. Like, okay, it's HTML. I want to, of course, make sure that I can, you know, pass and and extract the text. Okay? So, now that I've extracted the text, I'm going to query the Neo4j EU AI Act regulatory assistant. So, it went and asked a query to the

agent. The very same agent that is here, right? So, it's talking directly to the agent. But it's not going to ask one question. It's going to ask multiple questions. So, again, I'm just to driving the request and then it will automatically chat with the ORADB agent. It will automatically ask some question it's considering that the context is not good enough, it will go back and forth. So,

sometimes I have up to 10 discussions between those agents up to the result which is a very, you know, qualitative answer. So, I'm not going to go through those entire steps, but imagine here it's asking for a question. So, here they are replying that you need additional tool, you need additional context to query this. That's okay. It won't stop with a no. It will go and ask

for another query, another way to access the the knowledge graph, which is amazing. So, there are no broken queries here. You will always have those kind of motion. It will never end up. Okay? This is why it can take up to a few minutes. Let me get back to a recent one. A recent one basically was the same, the same motion that I showed you I want

to check on ORADB EU regulation. And here it passed the HTML and went to also get some additional context online. So, this is something that you can, you know, activate or not or not. You can say stick to the knowledge graph. I only want the single source of truth or truth, which is the knowledge graph. Or you can allow any other tool to go and fetch on

the web, right? So, here they they got some additional information from the AWS ORADB marketplace listing, which is good because there are a lot of additional information such as the DPA file, which is super important. And here it looked as the EU AI Act. Okay? And then it's giving me a very precise very cool way to assess it. So, yes, it is compliance. Yes, you need to

double-check because it's an infrastructure and it falls under AI system builder deployer. So, it's a very precise section of the regulation. So, what is amazing is that in matter of, you know, 20 minutes of course, I I shortcutted the processing time. In a matter of 20 minutes we managed together to go from this very chaotic and very complex legal to an ontology that has been one-click deployed

and and created. The ontology was created through the Perseus website. Sorry, I need to get back to my Perseus. It's here. Let me show you again. This is the generated ontologies. You can download the file. You can see the evaluation. We are We are giving you an evaluation as for the quality of the ontology. So, everything is embedded within the platform and automated. Then we went through

the Ontology-based graph generation. Then we plugged the ORADB instance activated the ORADB agentic tools and then linked through the MCP servers to cloud. So, we can very easily and intuitively through your WhatsApp, by voice, talk to a knowledge graph. So, respectful of time, we still have 10 minutes to go for the Q&A. Um do not hesitate to ask anything on this. Very happy to uh show you

the engine behind and and give you my overall feedback on the on the entire process. Zack, up to you. Zack or maybe Martin, could you maybe the questions? Oh, okay. Sorry. This is text questions, right? So, how to make ontologies for different subject matters? Is there any framework? So, is your question about having multiple ontologies at the same time? So, the ontology in our case is use

case So, it's not a domain based ontology. Doesn't mean that it's not reusable between your different agents, but of course we want to make it them these ontologies super granular, super easy to deploy and super easy to share with subject domain experts. So, again, for the subject domain expert, advise and recommend is first sharing the evaluation. So, the LLM as a judge that I displayed at the

end is a very cool way to embark your domain expert in in the ontology journey, but also you can share protege visualization tool with them. You just upload the total file. You can ask them for comments. You can ask them for additional classes, renaming, merging. This is a way to interact. So, you can see it as a semi-automated motion. We do the heavy lifting, 90% of the

job, and then you can apply the 10 remaining percent on making sure that your domain expert are fully vetting the ontology. Yeah, Conrad. Uh you insert something interesting. Jesus indeed applied an interesting talk on ontology building and also leveraging use case based ontologies and an iterative process. This is very interesting to understand and assess the quality of the Where's the leaderboard? Okay. So, the leaderboards, we we

will deploy our really really soon, but right now you have all the data available through our text to graph benchmark. So, I'm going to copy paste this link. This is here. Uh you can access the leaderboard. You can also download our scientific paper. So, we um uh documented and published a paper that was accepted last year at ISWC. And this paper is about how we created an

evaluation data set. So, we manually uh annotated the text to KG bench. Uh and the text to KG bench is a public uh text to triple data set that was that needed some improvements to put it that way. So, we spent a year with a computational linguist internally recreating, reformatting the ontologies, and re-annotating the data set to make sure that we have a very robust text to

triple evaluation. And so, we run it through any of the new models that are live and in GA. There's a question from Varun. How is the graph schema being created? Are you using a LLM or rule based algorithm? So, we do leverage LLMs. Uh we leverage those LLMs and there's an agentic framework again uh that we have fully documented. I'm going to copy paste I'm going to

copy paste the article on this, Varun. It shows you all the steps. So, building a semantic network YAML file, then creating the definition, mapping the hierarchy, mapping the relations. Everything is you know, broken down. So, we post-processing everything, making sure that we are fully compliant. Those LLMs are incredible in a way that they've learned on very famous ontologies and very you know, high level ontologies. So, we

do ask them to start you know, top categories, top classes by reusing very common classes. This is a way to make sure that we you know, follow some some rules of course. And then once you go into the hierarchy, once you go granular, then it really applies the very specific classes to your use case the very specific mapping to your use case and definitely the relations are

totally abiding by your events, your attributes. There's a question about text to Cypher. How to improve text to Cypher natural language queries to Cypher without using Pad Neo4j embedding provider? So, here again, I don't know. I don't know much about this to be honest. Probably you can leverage schemas. We realized that if you add the schema into the of course the the the prompt, then it will

provide you with better results. You can pretty sure leverage a lot of different models, not only paying models. You can open source model such as someone advising. So, there's a question from Priya saying, how does the graph traversal work during inference? Is it writing Cypher queries under the hood or is it using some sort of similarity search? Also, do you have a benchmark on latency token consumption?

Okay. again, this one is not there's no similarity search on these specific examples because we only use Okay. If we were to you know, vectorize those triples, which is something that in our one of our product, our Grapher product, we would have similarity search involved. But here, there's no vector support. Only a graph query. So, these are Cypher queries under the hood. Okay. And this is leveraging

uh what is embedded within the the OrientDB platform. So, this is leveraging the natural language to Cypher tool that is embarked and integrated within the OrientDB tool. There are plenty of other ways to do natural language to Cypher. You can implement yours. There are plenty of papers on this. So, you can you can check online. You're not vendor locked in to OrientDB. You can use other tools.

So, can you share details on the underlying engine for ontology creation? So, yes, this is the article that I shared. Uh you can go and get it. There's also a paper that is linked to this. We published a paper a year ago on ontology building. Uh we benchmarked a lot of the methods. So, very happy to discuss with you guys about this. I'm going to write down

my email address. You can ping me anytime. [email protected]. You can definitely ping me if you have any question about how you want to how we built those ontologies automatically. Also, I'm in useful work. Can you share the material KG creation script for the EU AI Act? So, here I use the Perseus platform. So, again, I'm going to write down. I have a lot to share, guys. So,

here is the link sign up freely on Perseus. On Perseus, you would be able to upload your existing total files if you have ontologies in house. And then of course, you will be able to launch the generation. So, that would launch the KG creation script. Okay. So, this is a of course very easy to to use and you're not forced to use the console. If you want

to work through our API SDK, then there's a technical documentation that is available on our website. Uh how is the ontology being evaluated for correctness? So, yeah. We have a detectors. We are running first a pitfall detector. So, here I'm linking the repo for pitfall detector. We open source this detector. So, here we're going to go through all the you know, major issues. So, could be any

type of issues. Could be semantic issues, logical issues, structural issues, redundancy issues, naming issues. You name it. So, we have around 20 issues that we are analyzing. So, this is part of the evaluation process first. And and and once it's done, then we apply LLM as a judge for more overall assessment on structural quality, semantic quality. a bit more qualitative. So, you have two evaluation processes. Pitfall

detector that I just linked and then LLM as a judge. Thanks. Incredible work. Thanks a lot. It's 5:00 p.m. my time. Uh we reached the 30 minutes. So, I'm going to have to leave you guys. But we can uh follow up uh this conversation, you can directly uh reach out by email. >> [music]