Daniel Davis explains why most AI proof-of-concepts failed in 2025: they lacked proper contextual grounding. His company TrustGraph builds context graphs that transform unstructured data into embeddings and graph structures, then retrieve relevant sub-graphs for AI systems. He describes it as a Supabase for AI context.
Chapters
What is context graph actually? And who's faking it?
Oh, almost all of them. Gartner just released an article about the topic and everybody in that list is faking it. I know, 'cause they've actually, some of them have come to us asking us for help. The term graph itself, there's a lot of confusion around it. I did a talk a couple weeks ago to a data engineering group here in San Francisco. And I actually asked, you know, how many people here, you know, know graphs, know graph systems, work with them? And I saw like one sheepish hand in the corner. My estimate is that about 10 of the audience had worked with graph systems, which really surprised me. Graphs are just kind of a mathematical field where you are connecting points of information to other points of information.
And—
It's really fundamentally not any more complicated than that.
Is context graph the same concept as a knowledge graph then?
A context graph is simply a graph structure that is optimized for usage with AI technologies. We're referencing papers that were written in the '60s, '70s, '80s, and in the '90s. There were a— there was a tremendous amount of research.
Why do you think it matters more now?
I think it's because bluntly, 2025 was the year that AI POCs failed. I even had somebody reach out from a mining company in Brazil and they said they had done a bunch of AI POCs and they all didn't really get the results that they wanted. And it's funny 'cause if you look at the timelines and it seems the answer has converged onto Context is all you need.
Let's say if I were to build a context graph from scratch, right, based on my RAG knowledge and my AI knowledge, so let's say I'm a company and I'm gonna take my company's documents and data, right? And I will, you know, chunk them and run them through an embedding model, like, you know, and then store them in, store the vectors in like PyCon or Quadrant and pass that context when an agent needs it, right? Do like similarity search and stuff, fetch the top results into the prompt and maybe add a knowledge graph on top with it using Neo4j.
Everybody's doing it a little bit differently. This is where also we gotta talk about again, like EF versus property graphs, because you mentioned Neo4j. They are a labeled property graph. Property graphs are when you have nodes and edges, and then you add lots of attributes to them. For instance, if I use my example of Fred has 4 legs, so if I wanted that to be a relationship, it would be Fred has legs 4.
Yeah.
I could add properties to that, like Mark told me, Mark told me on Friday. RDF graphs focus more on much more complex structures. The dominant use case for these systems was fraud detection and KYC. Anytime there's a transaction, you're querying a graph system to try to determine, is this a fraudulent transaction or not? There's a transformation process. So if you have unstructured data, You have to transform that into graph structures. So the way that we did that initially in TrustGraph is a very, very flat structure. During the ingest process and the transformation process, we would create vector embeddings for what we were actually storing in the graph and mapping those to the graph so that when somebody made a request, that actually gets pulled through a semantic similarity search. Which finds these embeddings, and that's what we use for the basis of subgraph retrieval, which is essentially you're now retrieving DAGs is what you're doing from the graph system, and that gets repackaged into the context, and that's GraphRAG, HybridRAG. What happens when the data changes or doesn't change? I'm not a big believer in fact or fiction, black and white. I think the closest that we can get to saying something is factual is if through continued observations across a large span of time, that it's always observed the same way. You have that initial context. Fred has 4 legs. Well, does Fred have 5 legs now? Now you start going—
Right.
Here was the initial observation. Here's the second observation. Has it changed? Every time you do this, you start cataloging that when was it observed, were there's a change, were there not a change. If there is a change, you start creating that branch. Well, this is how it changed. Then you start making calculations. How did it change? What if all of a sudden Mark told me Fred had 5 legs and he was just making a joke? He lied to me. We as humans can violate our world models. Like my, my shirt is purple. I could tell you it's green and I could tell you very confidently, no, I'm wearing a green shirt. Mm-hmm. I'm violating my world model 'cause my eyes are telling me it, it's, it's a purple shirt. So measuring these things over time, how they change, that can start to get us into why it changed.
I— TrustGraph is, is taking a graph video of your existing data and monitors it over time.
Just because something is old does not mean it's still not the best source. Of information.
Is it different from Neo4j, what they're building?
TrustGraph 1.8 supported Neo4j. TrustGraph 2.0 does not.
Why?
We're more like Supabase for AI context.
Oh, I meant—
And that, so, 'cause Supabase basically abstracts all the storage that you need for creating a web app. So TrustGraph, it abstracts all the context services, storage and retrieval that you need for building highly reliable and accurate AI agents. So we take unstructured data, we structure that as a graph, and then if you have tabular data, you know, lots of rows and columns and numbers, and that is stored as well because I got a secret for you. Don't, don't, don't tell anybody. Graph databases are just rows and columns too. I think the people I know at Neo4j are probably gonna unfollow me on LinkedIn after this.
Do you have a doomsday prediction for AI?
Venn diagram of the movies Idiocracy, The Matrix, and Terminator. In the middle is like, You are here. And I sometimes feel like if you, I think the movie Idiocracy is the one that I look at and go, oh no. Oh no. Is this where we're going? Because, should I say this? If you just look at our current government administration and President Camacho's administration from Idiocracy, are they really that different? It goes from being a comedy to a dark comedy to being a documentary. For me, that's what scares me is that I see two things happening and they're kind of like converging. I see people blindly accepting the information they're told based on who shouts the loudest. And at the same time, I see government power being consolidated in the hands of the few, being able to make decisions just unilaterally with no oversight, no checks and balances. When those things start—
