Everyone's Talking Context Graphs. Here's What One Really Is | Daniel Davis, TrustGraph

I asked all the major AI services a really simple question. Who are former NHL head coaches who are not currently coaching a team? Every single major AI service made mistakes. I asked it because a friend of mine and I were chatting. You know, what are the top 10 songs that they've played live over their career? They actually have a website that has all this data. Not a single one of them got it right. Perplexity gave me a top 10 list that wasn't even in numerical order. I think most people are using AI on questions they don't know the answer to. I mean, that's kind of the whole point. Legitimately scares me that I don't see people fact-checking the information that's being presented to them. You can hire somebody to get whatever you want in a Wikipedia article. Wikipedia does not fact-check anything. Their criteria for what goes into an article is who made the edit.

It's such a highly weighted source on AI search engines and even on Google, right? I was trying to do validation AI give me the correct information. Then I double-checked that data through Google. It turns out that data is wrong, and that data, that blog post, whoever wrote it probably used AI and generated wrong data. We are facing this, this compounding misinformation era. Welcome back to Two Sided AI. Today I have Daniel Davis, you know, the co-creator of TrustGraph. They're the open source context backend for AI agents that speak been trending on GitHub. Daniel, welcome.

Thank you. Great to be here.

This is like the hottest term in AI infra right now. And now every company seems to be a context graph company, right? What is context graph actually? And who's faking it?

Almost all of them. I can tell you for certainty, everybody that was in— so 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. So it's not a radical new concept. Well, it's okay. First of all, there's been some people that even claim some of this was plagiarized for them. I don't want to go into that fiasco, but this is not a new term.

Mm-hmm.

If you look back in academic literature, you can find the term context graph going back 20 years. Now, those references tend to be more just what we think became the term knowledge graph, which again is kind of not an old term either. That's kind of something that especially Neo4j really pushed that term really hard and kind of did and didn't catch on. That's kind of another topic, whether, you know, knowledge graph people really have a strong sense of what a knowledge graph is. And the first instance I can see somebody using the term knowledge— excuse me, context graph in a way that we think aligns with the usage that we believe is what a context graph is, was Vicki Frohan in 2019 when he was at Calibra. And I stumbled on that talk by accident., and, uh, he gave it in 2019. I forget what conference it was at. And funny enough, he only gave it one time. I know because I just recently did a podcast with Vicki where we talked about all of this stuff. Uh, I really highly recommend it. Vicki is really, uh, a truly an expert on AI, cognitive science, human psychology, and so many of the topics that are not being discussed in this field that are really paramount to pushing this technology forward in a way that makes sense to work with us and understand us and how we're gonna use it. That being said, 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, And that's kind of been my experience this year. I've given my "What is a Context Graph?" talk 3 times now in the Bay Area, all to different kind of, you know, data engineering groups. And my estimate is that about 10% of the audience had worked with graph systems, which really surprised me. I didn't expect it to be that low, except, except they have and they didn't realize it. Level set. Graphs are just kind of a mathematical field where you are connecting points of information to other points of information, kind of a, if you start really going up the ladder, set theory, et cetera. And it's really fundamentally not any more complicated than that. A knowledge graph is a very use case agnostic, simplified, well, I shouldn't say simplified. That's not the right word. Write the word generic, I guess, is a more, is a better word, maybe. I also feel, I think we're gonna get into syntax and diction a lot today, more than likely. I've given talks on that too. I am quite pedantic when it comes to words. A knowledge graph is very generic in that there's so many different ways you can structure it. It can be directional. It doesn't have to be directional. You have, you know, what's more of a semantic web, RDF style graph. You have property graphs. So many different ways you can structure a graph. And then you have the data engineering world that works a lot with DAGs, you know, directed acyclic graphs. It's still a graph. It's a graph. A DAG is just a, a sub— it's really a subgraph. It's nothing more than a subgraph where you have node A connects to node B and there's a path. Well, that's just a subgraph that you could extract from a knowledge graph. And that's actually what the entire field of subgraph retrieval is in knowledge graphs, where you build this huge graph structure and then, you know, you take a starting point and you start retrieving nodes and edges related to that based on various different parameters. So DAGs kind of are a more specialized version of just a knowledge graph. So they're all in the same family of really this mathematical structure. And there seems to be just so much confusion about that. And I don't really have a clear understanding as why. I know this topic is not really taught in academic programs. I was actually just talking to, I've been getting more and more involved with a student organization at my alma mater, Georgia Tech, the LibreTech Collective. It's actually a bunch of students who are advocating for free and open source software. Really serious about it too. And I was talking to some of their members and he was saying that as far as he knew, Georgia Tech didn't even offer a class like on information theory or graphs in the College of Computing. Not to say that they don't, but it's, yeah, it's not a common subject. Matter of fact, the only way I learned this stuff is from the TrustGraph co-creator, Mark, Mark Adams. Mark is an expert in graph systems, truly world-class expert, truly. I learned all of this from him. So take that for better or worse, and he'll probably say, I've, I've twist— actually, he'll probably say, don't, he's not gonna take credit for that 'cause he'll probably say, I've twisted all this into my own way of explaining it. He's like, it is not mine anymore. I'm not taking credit for it. And I found it very hard, very hard to learn about knowledge graphs. There's not a good, a lot of good information. That I could find on the internet through Google searches. The stuff I found was often contradictory. And he ended up making guides for me. We actually have this guide on how to build a knowledge graph using RDF structures that he used using his cat Fred. I keep saying I'm gonna make Fred famous. And he took 3 simple statements. I probably still remember them. Fred is a cat. Fred lives with Hope. And Fred has 4 legs. 3 really simple statements, and he went through the entire process of how that becomes an RDF structured knowledge graph. And it's, it's like pages and pages of explanation of how this becomes a structure. It's quite complex. And without taking like a really just absurdly simple example and seeing how it becomes these structures, I never would've figured it out on my own. Honestly, I, I wouldn't have. And I still had to refer to that guide for quite a long time. And I still do occasionally, actually, if, 'cause I don't, if I don't work with the structures often enough, I sometimes forget like, wait a minute, what does a class mean? And what's, what do these things mean in RDF again? Hold on, I need to see it in front of my face again. And I actually asked Mark, you know, how did he learn about knowledge graphs? And it was kind of the same thing. It was very tribal. He had, kind of people that taught him as well. So it's, it's a field of study that if you don't have somebody that can really work with you and mentor you and teach you, it's a really hard topic to learn about on your own.

Is context graph the same concept as a knowledge graph then?

In my opinion, yes.

Okay.

I have taken the position that a context graph is simply a graph structure that is optimized for usage with AI technologies.

So it's kind of like a rebranding of the knowledge graph, right?

In my opinion, yes.

So you're, are you seeing this as just a natural evolve or rebranding of the same thing?

It's a little bit of both, but the ontology point is really interesting because that's one that, that's really a semantic web concept. Ontologies, which come outta the whole RDF ecosystem. Are you familiar with kind of the promise of the semantic web?

Right.

Yeah.

Yeah.

'Cause it goes back to the '90s, so it's not something that everybody's familiar with. So basically, as people saw the internet emerging as being this transformative thing in our lives, they thought, you know, how are we gonna be able to exchange information from websites and services? Because this is, you know, during the frothy internet days when everybody thought a website was gonna be the end-all be-all, just kind of like how people think AI agents are going to be the be-all end-all. Wait a minute, have we done this before? Anyway, they were looking at ways of structuring this information to be able to exchange, and they— that's where kind of most of the ontology work began with RDF, OWL, a lot of those kind of standards. Well, OWL fits under RDF. RDF's complicated, and it's not as— it's kind of like an umbrella ecosystem, and you have lots of things underneath it, like OWL, like even SCOs. There's so many ontologies, hundreds of thousands of ontologies. It's, it's truly astonishing how many there are. But I also, I often have people ask, you know, what really is an ontology? And I break it down into this four related concepts here. There's vocabularies, there's taxonomies, schemas, and ontologies. And we kind of understand vocabularies, you know, we have words and they have definition. That's a, you know, that's a vocabulary. Taxonomies are when you create these hierarchies of them. Like for instance, you mentioned the medical field so that you have terms that have specific meanings in potentially, uh, no neurosurgery, you know, neuroscience, whereas, you know, doctors that are, you know, deal with the heart, you know, cardiovascular surgeons, et cetera, they have a different taxonomy. And the, Maybe the same word used slightly different way. So you have this hierarchy of definitions of vocabularies and other kinds of taxonomy. And those are all designed to be human readable for us. Then you have schemas and ontologies. You know, schemas are machine readable in a way that you can kind of structure key-value pairs. And they're really designed for, I wanna store information, I wanna be able to retrieve it. Ontologies are kind of when you take all of this You take the taxonomies, the vocabulary schemas, and then you package it in a way that's designed to be exchanged. Now, ontologies and schemas are, in a lot of ways, very similar and have like slight differences. So a lot of times you'll see those terms kind of used interchangeably, but there are some kind of slight nuance to it. And that's why I would say that, yes, there is kind of a little bit of rebranding old ideas here with the term context graph. Because most of, so most of the planning that we're doing now on a lot of the features and capability that we're gonna be building over the next year, 5 years, et cetera, 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 in this area that has just kind of been forgotten. And it was, it was really, really, really ahead of its time. And interestingly enough, was not continued. It kind of ends in the '90s. And in the 2000s, people started focusing much more on the mechanics of building these things, taking a kind of software-first approach to it. And then the deep learning camp with, and then machine learning and it became very mechanical. More data, more compute, we'll be able to batch training jobs, make the training pipelines more efficient, more data, more training. That's gonna solve everything. And that was kind of the, the Geoffrey Hinton approach with deep learning. And that approach won out for a very, for a while and got us kind of to where we are now. And this reemergence of the term context graph or reemergence of the ideas and then kind of branding it. Brand.

So why do you think it matters more now?

I think it's because bluntly, 2025 was the year that AI POCs failed. And we're hearing that time and time again. I've had so many people reach out. 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 because if you look at the timelines, This term context graph really exploded around like November or December. And what's probably happening, all of these enterprises are doing their year-end reviews and they're looking at all these AI POCs and going, well, none of this worked. Why? And it seems the answer has converged onto context is all you need. And then it becomes, well, how do we structure the context? And people going, Oh yeah, that whole neurosymbolic AI thing. I think they were right. And here we are. And it's quite a few. We, we kind of have like 3. Well, we actually have a long list of reasons that we see. And, and one of the top ones that comes to mind is that LLMs are very confident in their responses, but they can be wrong. And I actually published an article, um, just yesterday. I, I tried something really different. I tried to take a really different topic and show this in action. So, I actually took a topic from NHL hockey, something that seemed like totally mundane, and asked it, I asked all the major AI services a really simple question. Who are former NHL head coaches who are not currently coaching a team? And I thought, you know, that's simple enough. That should work, right? And every single major AI service made mistakes. They either gave, gave me names that were not correct, that were, you know, some people that were actually— they had been fired in the past but are now have been rehired and are current team NHL coaches. Or it gave me people who were assistant coaches, and that's not what I asked for. I said not coaching in the NHL. Well, an assistant coach is a coach, and you could say, well, that's splitting semantic hairs, but that's exactly what we're doing right now. You know, that's exactly what we're doing. We're diving deeply into the meaning of these words and trying to— what was I actually asking? What was I hoping to achieve? And I came up with this kind of rubric on how I graded the responses. And as I was going through it, I even myself was kind of thinking, wow, this is actually harder than I thought it was going to be. I can make an argument for giving either any of these really bad scores or really good scores. I even have a section in my article saying You know, there's an argument to be made that all of these deserve Fs, because then you start getting into the— you start getting into this really dicey situation of where do you trust somebody? Like, if you ask somebody to give you a list of 10 facts, like, for, for instance, Perplexity gave me a list of 13 head coaches for my list. 11— 2 of them were wrong, just dead wrong. So it was 11 out of 13. Well, if somebody claimed to be an expert and they gave you a list of 13 points and 2 of them were wrong, like, do you still consider them an expert? Some of the other services gave like 6 responses and like 1 of them would be wrong. Like, that's not very many. Like 1 out of 6. Like, do I still trust you? Like, where is that point?

Yeah.

Like, where is that point? And I mean, that even, I think I I have a great anecdote. I don't know if you want to hear it or not. It's, it's a bit out there. I call it when 9 out of 10 experts were wrong, man. So, so it actually is. So this goes back to my DOD days, and without trundling this into the ground with a lot of stuff that people are going to go, what? There's— people get really hot and bothered about classified information. Well, what they don't realize is their entire other system called CUI, Controlled Unclassified Information. And if you're really smart, well, if you're really smart, that's where you can hide stuff in plain sight. So there's an entire set of markings that's called STINFO. It's a science technology something, something, something. And it has different labels about how something can be— you see it all the time. If you see like something that's former, like a DOD document, if you look at the bottom, there's almost always a distribution statement. And if you're looking at it, it's always gonna say Distribution Statement A. Well, they go all the way to F, wealth. They're supposed to go to F. I was running a program, very, very dubious. We'll leave it at that. And I was, it was being handed over to me by another organization. And I had all these documents that were labeled Distribution Statement X. And I'm going, what is this? And I've always prided myself in knowing the rules and regulations really, really well. So I mean, I knew the STINFO rules really well. And I'm like, what is this? Is this like a typo? So I started asking people and I asked people, 30 years of experience, you know, classifications, you know, out the wazoo, you know, very, very, very, very, very high level of access. And they all said, I've never seen that before. That's gotta be, that's gotta be a typo. And I got to the end and I messaged a buddy of mine who's a former EW guy. That's electronic warfare, by the way. And he went, Ho ho ho ho, you got one of those. I was like, what have I got here? Please. He's like, well, he said, basically you've gotten something. He said, it is real. This is a special exemption that this was a black program that went white and they're giving themselves an out to take it black again. And he kind of showed me like, if you look here, you can find some information about it. So the moral of that story was I asked a lot of people. A lot of people I trusted, a lot of people that if you'd have put them on CNN tonight, you, we would agree that person's an expert. They were wrong. It was the one guy that happened to have worked in some very niche things that had seen it before. And what's really funny is you can actually, if you Google it, uh, there's information about it now, which is very, very odd, but back then there wasn't. So it's, it's funny how these things change over time, but that's one of my favorite stories for, uh, when 9 out of 10 experts are actually wrong. AI, LLMs are very confident about their responses. They don't go like, well, maybe I'm not so sure about this. And unless you know the subject very well, you may not be able to spot, hey, wait a minute, I don't think that's right because I asked it something I knew very well. I kind of already knew the answers. And so when I saw answers that weren't right, I immediately went, well, that's wrong. Well, that's wrong and that's wrong. But is that how you use AI? I think most people are using AI on questions they don't know the answer to. I mean, that's kind of the whole point, right? Yeah. Is that if I knew the answer, why, why would I ask AI? You're asking about something that you intrinsically have no frame of reference. For instance, I actually was talking to somebody in the Trust Graph community today about the article, and they said, he said, I didn't read it initially because I don't know anything about hockey. And he said, well, he said, if it had been about, you know, the English, uh, you know, Football League, you know, that would have been a different story. And I said, well, that's the exact point. If the article about the English Football League, I'd be in the same exact boat because I don't know that much about the English Football League. I wouldn't have any way to judge the responses the AI services gave me. I would probably just accept it on face value. And it legitimately scares me that I don't see people fact-checking the information that's being presented to them. This morning in our Slack, I said, I think software engineering is mostly just around writing tests. You know, the coding tools can actually write the code. It's all about does it achieve what you want it to achieve? Yeah, we're seeing a monumental transformation in how we work. For sure. But, you know, what I'm not seeing is that if you— these models still confidently give you the wrong answers without the right context.

It's—

yeah, this happened the other day. I— well, not— I, I post on Twitter a couple weeks ago. I, I— again, I like to ask AI silly things occasionally, and I asked it because a friend of mine and I were chatting about— because I met him at a Metallica concert here in San Francisco years ago, and we were chatting about you know, how, you know, what are the top 10 songs that they've played live over their careers? And I thought, well, this one's easy because they actually have a website that has all this data, but it's on separate pages. So surely I can give this to the AI services and they'll figure it out for me. Not a single one of them got it right. Matter of fact, Perplexity gave me a top 10 list that wasn't even in numerical order. Like, not only was it wrong, like, The one that it had for number 10 should have been number 4. Like it actually just had the, it couldn't give me a 10 list in the right numerical order from, from most to least. I'm like, these are just glaring.

Yeah. I was trying to do validation and it lead to my, uh, to a source, uh, which is a webpage or blog post. I clearly remember I looked at the data on the webpage. The AI give me the correct information. However, then I double-checked that data through Google and I want to verify, is this data on that webpage real, right? It turns out that data is wrong and that data, that blog post, whoever wrote it probably used the AI and generated wrong data. We are facing this compounding misinformation era.

Yeah, there's another component here. There's a component though. That's actually kind of been the white whale for graphs. I always wonder though, do, do my, does my Moby Dick reference still land? Did they make, did they make people read Moby Dick anymore? I have a feeling that nobody reads it anymore. So I, I, I always, I always make that reference and wonder if somebody, you know, you know, if somebody gets it, 'cause they'll immediately go, call me Ishmael. And not many people do that anymore. Anyway, there's a, there's a big problem here and I, I, You kind of touched on it a little bit. All these AI services really are just RAG services now, which I think is kind of funny, that they're going out and they're scraping the internet and finding sources, and they don't fact-check those sources. They're not looking at, you know, who the source is and have some score from it. A lot of times they're taking things from social media like Reddit. Reddit's a really common one they pull from, Wikipedia, etc. And public service announcement. If you think everything on Wikipedia is correct, do you know how Wikipedia actually works? Basically, the way the admins actually are trained to, to moderate Wikipedia is if two people make conflicting edits, you take the one who has the most edits, which means you have an entire class of people— this, this is the direct source from Wikipedia admins, okay? I know people that do this too. So what people have done is over the years, they get lots of edits, which builds their Wikipedia trust. They sell that service to people to edit Wikipedia articles, which means you can hire somebody to get whatever you want in a Wikipedia article. Happens every day.

You say you can hire somebody to write whatever you want them to write in a Wikipedia. Correct.

Correct. Wikipedia does not fact-check anything. Their criteria for what goes into an article is who made the edit, and if that person that made the edit is a high— what's the word— if they have— if they're a high—

not ranking, I don't want to say credible source, right?

Well, it's not credible. It goes down to if this person is a very, very, very frequent editor of Wikipedia with tens of thousands, hundreds of thousands of edits, Whatever they say goes. Remember, Wikipedia is a, is a nonprofit. These people are volunteers. These people are not professionals. The people that moderate Wikipedia are volunteers.

Oh my goodness. Yeah. But it's, it's a such a high, you know, highly weighted source on AI search engines.

I know.

On Google, right? It's like, you know, this is, this is the top results.

I know. And I would say, you know, the more, the more contested a topic is, it's, you're gonna have a lot of people fighting for it. Mm-hmm. So, but, but, but people have done this, you know, we have, you know, there've been cases where teachers would assign a topic for an essay, go and edit the Wikipedia article with incorrect information, and then they tell the students, you're not allowed to use Wikipedia. And they would see how many students would include that fake information they planted in an article in their essay because nobody's checking it. One of the music artists I used to work with, his Wikipedia page had a lot of stuff in it that was factually incorrect. I had no idea who had added it. Matter of fact, actually, you know what's funny? There's a, a YouTube channel called Loudwire who literally had a, a series called Wikipedia Fact or Fiction where they would sit down with artists and go through their Wikipedia and ask them, is this fact or fiction? And for some artists, it could be 50/50. You know, there was a lot of stuff, like most every time it was about 50/50 on stuff being just sometimes just like totally just fiction. Like they're like, I have no idea where that came from. I have no idea what this is. No truth to it whatsoever. That's one, that's another one that's, that's concerned me for years is how much stock people put in Wikipedia. Especially when you understand the process. And then you say like, hey, this is moderated by volunteers. Do you think they're really fact-checking all this? This is about facts, right? It's very easy to manipulate information. And this is kind of what you're talking about with context graphs, is there's a, there's a big element here. And the elephant in the room that gets ignored all the time is time. And it's the hardest to work with, and it's the most impactful and important.

So let's say if I were to build a context graph from scratch, right, based on my RAG knowledge and my AI knowledge, um, So let's say I'm a company and I'm gonna take my company's documents and data, right? And I will chunk them and run them through an embedding model, like, 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, because GraphRAG was very popular like last year or two years ago. So it's kind of like, you know, combining the RAG and GraphRAG approach to build this context graph. Like, am I, is my understanding correct? Is this the sequence? Of how AI is leveraging context well?

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.

What is, what is a graph?

Okay. So that pushes you in a different direction. And this is kind of, this is gonna be tricky to go through without visuals, but although the visuals don't necessarily help either. Property graphs are when you have 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.

Okay.

So I could add properties to that, like Mark told me. So that's like, you know, an attribution property. Mark told me on Friday, so I could add more, more properties to it. And those are the kind of things that are really— that property graphs do really, really, really well. RDF graphs focus more on much more complex structures in the TripleStore itself, in the hierarchies, in classes. That's where you get ontologies. Ontologies are much harder to work with in property graphs. That's not saying you can't do it, it's just harder to do. When you talk about ontologies, you're talking more about an RDF-style graph. And most of the well-known graph systems like the Neo4js and all the others are, tend to be property graphs because the dominant use case, I don't wanna say dominant, but it kind of is. Yeah, I'll go with dominant for these systems was fraud detection and KYC. So almost every company has like a fraud team, KYC, and where you're trying—

KYC is?

Know your customer.

Okay.

So that falls under that, you know, AML, KYC, where AML is anti-money laundering. So, you know, if you're doing commerce, especially with consumers, this stuff comes into play and you can kind of offload it sometimes to some people, but, you know, the bigger you go and the more you start taking it on, you have to do it yourself. And, Anytime there's a transaction, you're querying a graph system to try to determine, is this a fraudulent transaction or not? And those tended to be labeled property graphs. On the other hand, you had an entire other industries that used more RDF style graphs for very, very large scale systems.

You mentioned, uh, can you explain, just, just add a quick explanation of what does RDF represent?

Well, RDF, again, that is the framework. It's actually Resource Description Framework. I always get it wrong. It describes a standard that essentially was the semantic web, or is the semantic web. The semantic web still exists. Matter of fact, Google SEO for years operated off of schema.org. That's an ontology. If you know Google engineers that work within Google Search, They know a lot about graphs and a lot about ontologies. Google Search is a graph system. Very, very, very big one. And it just tended to be different use cases, like the intelligence communities, you know, they use graph systems for determining whether or not they're going to trust the situation in a millisecond, or is this thing that I'm seeing over here that's strange, what does that mean? Oh, and the other field that graph systems get used a lot is cybersecurity for anomaly detection. You know, I see this thing popping up on my network, you know, that's, I've never seen it before. What does that mean? And you have very sophisticated algorithms that can, you know, in an instant, in milliseconds, score this as being, this is okay or this is not okay.

So, yeah, these are all the use cases. And my approach is just a very simple, like, a that, you know, I mean, if I don't have a lot of documents, it's not very hard to set it up that way.

Right, right.

So yeah, go ahead.

There's a transformation process. So if you have unstructured data, you have to transform that into graph structures, and there are a multitude of ways of doing that. So the way that we did that initially in TrustGraph is a very, very flat structure so that it's essentially just a triple store, which means it's just, you know, thousands, millions of triples. And somebody's probably saying, well, how do you retrieve from that? Well, 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. We called it TrustRAG for a while and nobody cared. You know, we— it is what it is. Now you can make an argument that a system like that in totality is a context graph. Again, I think it's about how the information is structured and structuring in a way that is optimized for AI usage. And then the big thing that was really talked about in Foundation Capital's article was what they were calling decision traces, where you were actually capturing how agents are behaving, which goes back to the very, very old idea of reification in graphs. And what's you're gonna have to bear with me on this. Great. So the best definition I've ever seen for reification is it's statements about statements. And you're probably thinking, well, that's really, yeah, exactly. Well, that's, that's the best definition I've seen. So re— remember how my example of Fred has 4 legs?

Yeah.

And then I said, well, Fred, Mark told me, and Mark told me on Friday. That's reification. Those are statements about statements. So the initial statement is Fred has 4 legs. Mark told me on Friday is a statement about that statement.

Okay.

So decision traces or how an agent behaves, when you capture that, you know, this is the context, this was the prompt, these are all the model settings, this is the model itself, this is the model was in AWS, this is how long it took the model to respond, this is then the next prompt, then the next context, and here's a chain of events that led to an output. If you take all that information and you then add it back to the initial context that was used, you know how I've reified that context. And interestingly, that the initial kind of, the initial article didn't really make that connection. And I can't take, we can't take full credit for this because a lot of other people like Kirk Markle from Graphwood and others have also said like, hey, you guys are kind of missing a piece here. There's the initial graph itself of context, and then there's how you reify, how you add this this new context, this behavioral context, connect these two and how this becomes a living system that evolves over time.

Why do they call it a decision tree? Is it, is it, is it just a different form of the graph? They're just saying, you know, like a decision tree, right?

I think it's my, because you have to call it something. Not everybody calls it that. I found I have found creating terminology for technology to be more challenging than ever before. You would kind of think it's gotten easier, but it's actually gotten harder because the technology's gotten so much more complicated. So you're always constantly searching for that elevator pitch, you know, how do I explain this to somebody in 15 seconds that has no frame of reference for that? So you tend to anchor on words that people have a strong sense of understanding. Like decision trace, people go, oh, I think I know what that is. So can remember we were talking about earlier about how knowledge graph never really permeated the zeitgeist and it seems like context graph has. It's because, world according to me, because people have a much stronger sense of what context is. And when you join that with graph, they're like, oh, I think I know what that is.

That makes sense. Yeah.

But the word knowledge. If we ask 5 people, we're gonna get 5 definitions, different definitions of what knowledge is. So when you put knowledge and graph together, people went, I'm not sure I know what that is. And I, so what is old will become new again. The underlying principles. And again, if you start looking back at some of the papers that we're looking at from the '60s, '70s, '80s, and '90s, this is really not new. There were people going all the way back as far as the '70s looking into reifying this stuff into graphs, literally going back 50 years ago and taking very sophisticated approaches to doing it as well. I mean, I've said many times on podcasts that the most— the best AI papers I've ever read actually were from the 1990s and the DARPA MUC conferences. You see, it's hard to find this stuff, but there was— they were doing incredible work. Actually, with automated graph building back in the '90s, they were taking documents and building pipelines and ingesting them into structures with really— like, they were scoring it. I forget what their metric was, but they were getting above 90% efficacy in their pipelines. And this was in the '90s. And then weirdly, a lot of this research just kind of— it kind of tailed off in the 2000s. Now neurosym— now people are talking about neurosymbolic AI. It's funny. I've actually, since the context graph explosion, I've all of a sudden seen that people giving talks, talking about neurosymbolic.

The whole process is automated, right? There's a lot of things is about maintenance, right? This is, this is the key with like vector database, I think, you know, because your data do change after you embed everything, things still, still can change, right? There is still data shift that could happen.

Absolutely. And you're getting exactly into all the things that we're thinking about and about to launch with Trust Graph 2, the subtitle, Context Comes Alive. I already have all the imagery planned out that we're going to use for this. It looks, it's going to be a reference to something that probably nobody gets but me, but it amuses me anyway. What you're getting at is what happens when the data changes or doesn't change, and that's what's important. So I have this little theory— well, I, I shouldn't say it's mine, I'm sure somebody's come up with this before. I'm not a big believer in fact or fiction, black and white. I, I think everything is kind of more in the middle, somewhere in the gray areas. 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, something that always was and always will be. That's about as close to something being factual as it can possibly be. So to your point, you have data, your data updates, well, you re-ingest that data and you could do a lot of things. You could create a new collection for that. You could create new context cores, but what you really want to do is for— you'll have new stuff, but for the stuff that's not new, this is where the reification comes in. So you have that initial context: Fred has 4 legs. Well, does Fred have 5 legs now? Does Fred have 4 legs still? So now you start going, here was the initial observation, here's the second observation. Has it changed? No. And then you start— every time you do this, you start cataloging that. When was it observed? Were there 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? Because what if all of a sudden Mark told me Fred had 5 legs and he was just making a joke? He lied to me. Matter of fact, this is something that Vicki and I got into with the— our world model debate is that the act of human lying seems to imply a lot about the integrity of this concept of, of world models so that 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. 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. What does this mean?

You're saying you're accumulating data on top of the existing, the first data snapshot. And so you're capturing the change over time, which gives you history data and reasons of like context and background of why changes happened. You're saying TrustGraph is, is taking a graph video of your existing data and monitors it over time.

Absolutely. And I have an example and I got called, I got criticized for this in my What is a Context Graph video. I said that you look crazy. Somebody said I look crazy for talking about this, but I think it's relevant. And that's information in the UFO/UAP topic.

Okay.

Me. Without going into too much detail, it's a topic I've, um, researched most my life and not something I've talked about a lot. Uh, it's actually part of the reason I actually did leave USAF, believe it or not. I don't talk about that too much. But anyway, if you look at the data that we have now and you look at the discourse around the topic over the— essentially we're looking at about 70 years now, 70-ish years that people have been researching this topic and coming forward with data. Oh, for the first 50 to 60 years, all that data was considered misinformation. Now we have, you know, Marco Rubio, Secretary of State, was in a documentary where he said, yeah, there's absolutely stuff going on. You had active government officials collaborate— I can't say that word today— validating these claims that are 60 years old. 70 years old, some cases. So if we looked at that progression over time, like with a credibility score, if you looked at, you know, the 1950s, '60s, '70s, '80s, '90s, the public would have given this a very, very, very, very, very low credibility score. But now you have people that seemingly are credible people that are giving veracity and validation to these same topics, the same data that's 70 years old. So what that's telling us is that just because something is old does not mean it's still not the best source of information. Now, at the same time, it's the same time, things also change. We do get new perspectives. Well, the thing I've actually seen some examples of this, and, and I I wish I could think of one off the top of my head, but I've seen some rewriting of history and I start, you know what, it starts, I start having like this, this kind of like moment of terror thinking, you know, what happens when you're the only one left that remembers what happened? And again, it goes back to our initial, you know, my 9 out of 10 experts were wrong anecdote. You know, what are you more likely to believe? 9 out of 10 experts or 1 out of 10 experts? So if I'm the only one, right, if I'm the only one that kind of was there and remembers and there's nobody left to, you know, say like, yeah, I was there. That's exactly what happened. What is that information? There's, I'm the only one that can validate it. So how does somebody else perceive that when you're the only source? And it really concerns me that I don't hear people from the AI Frontier Labs talk about these topics more about the ambiguity. In meaning, in what is fact and fiction. These are really, really important topics in this day and age. And we seem to be getting farther from that thinking rather than, than closer. It seems that fact now is whoever can get the most likes on social media, whatever permeates the zeitgeist. Well, that's faction. One of my favorite movies is from the '90s. It's called Wag the Dog. It's a very, very dark comedy. I find it hilarious, but some people might not find it so funny in today's times because it's actually about— wow, I just realized, talk about relevant— it's about a president creating a fake war to cover up for a scandal. And Robert De Niro's character in the film has one of my favorite lines, um, because somebody was saying, you can't create a fake war. It's fake. It doesn't exist. And he went, what are you talking about? I saw it on TV. It's real. I saw it on TV. It's real. I saw it on Twitter. You know, it had, you know, 4 million views. It's real.

Such a profound comment. You know, I do think this is part of the algorithm. It's baked into it. More likes, more upvotes, you have more credible source of information.

I really hope people talk about this topic more. I really do. I'm, I'm really to the point, I'm really worried. I'm really worried about some of this stuff because I'm flabbergasted myself when I ask some of these AI services simple questions and it just still gets it wrong.

Are you fundamentally still different from Neo4j, what they're building? Or how are you different?

Yeah, quite a bit now. So TrustGraph 1.8 actually supported Neo4j. You could use Neo4j as your graph store. TrustGraph 2.0 does not.

Why?

We have, we are now designing our future roadmap around RDF 1.2 and reifying data with that strategy. And also we've kind of used Cassandra as our storage system since day one as our primary, and it was our highest performance configuration and now We started to optimize for it. So in some ways you could actually view TrustGraph as a complete alternative to Neo4j as a graph storage system that just so happens to do quite a bit more. That's why we're, we're testing out calling it a context backend now. I think the, I think maybe the easiest way to think about it is we're kind of more like Supabase for AI context. Oh, 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 it's, it kind of follows into that database abstraction category. Perhaps.

Yeah, I see. Okay, that makes more sense. So, so there is, you, you, you are basically enhanced version of, uh, of Super, Supabase.

Where you have graph structure. Not just graph structures. We actually, we store tabular data as well, because essentially, like I said, it is Cassandra. Cassandra is—

So basically structured data, unstructured data, and graph data, and embedding vector data store.

So yes, all of that. Now for the vector embeddings, we're still storing that in Qdrant for now. Cassandra does have that functionality. We just— Qdrant is still higher performance from what we've seen. 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 and unconnect with me on LinkedIn after this.

Their CTO, Philip, I sent him—

I talk to him all the time.

I'm inviting him to my podcast as well, so I wonder what he's gonna have a very different take, I think, on all of this. Yeah.

So he's, he's actually been the big champion about the term knowledge graph. So, you know, it'd be interesting to ask him his perspective on that because he's, You know, that was, you know, he really, he, you know, he really banged that drum for a long time. And here we are now, people, you know, context graph just immediately clicked with people. And so it's just immediately, like literally just overnight, you know, on December, I forget when the Foundation Capital article was posted, December 22nd, I think maybe, you know, on December 21st, he was talking about context graphs by January 1st. That's, you know, people were complaining on Twitter. That's all they were seeing.

Do you have a doomsday prediction for AI with, with all the things we talked about?

I mean, I think a lot of people already feel like it's playing out. You know, I, you know, there was a popular meme going around a couple of days ago about saying that, you know, it did a Venn diagram of the movies Idiocracy, The Matrix, and Terminator. And they, you know, it said in the middle was 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? Are they really?

Yeah, that doesn't—

there was this, you know, well, it becomes like a, it, it goes from being a comedy to a dark comedy to being a documentary. I feel like the movie Brazil is very much that. Uh, if you have not watched Brazil in a while, I mean, uh, mid-'80s, probably Terry Gilliam's masterpiece. Another film, comedy, it was always a dark comedy. Terry Gilliam, even in recent years, has actually said it went from being a dark comedy to a documentary. And I think for me, that's, 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, who has the most social media followers. You know, if this person has 5 million social media followers, what they say must be right. And at the same time, you know, I see government power being consolidated in the hands of the few, you know, making, being able to make decisions just unilaterally with no oversight, no checks and balances. And, you know, when those things start, you know, converging, all of a sudden you become, That will occur, you know, what happens when those things, how close are we to those things converging and how do you disentangle them? How do you get people to start going, you know, you gotta like fact check this stuff. You can't just listen to who has, just because this has, you know, 5 million views doesn't mean it's correct. Then you start getting into what is correctness and that's not an easy topic to tackle either.

I have a traditional closing question. If you could recommend a book to, let's say, to the you when you were 20 years old, yourself today, and maybe yourself 20 years later.

Ooh, that's a tough one because, you know, the books that come to mind, I don't, they're kind of all about how you would frame the world or frame understanding and Ooh, which one would I actually read and recommend to people? I think, I guess the book that I recommend the most to people is if I had the one leadership example that I really hold on, on my, is my number one pedestal, it's Dwight Eisenhower, which some people kind of look at me funny about. And admittedly, he had very, very problematic views of, of kind of a woman's role. Let's just be honest. He, you know, didn't think women should wear pants, for instance. Seriously, some, some archaic views. Yes, I get it. There's a book that's actually written about his presidencies though, called Ike's Bluff. And I, the name of the author just escapes me, is a historian, but it's about appearance versus reality. The American public when they looked at Eisenhower during his presidency, they saw this guy that played golf a lot, that didn't seem to be very engaged, that didn't seem to be active. And during the second administration, the Kennedys actually used that against Eisenhower a lot in JFK's presidential campaign. You know, look at Eisenhower, he's out playing golf all the time. You know, look at JFK, he's young, he's energetic, he's He knows the people, he's gonna be the per— you know, the, the harbinger of change. And, and what they didn't know is that that was all somewhat by design by Eisenhower. He intentionally kind of misdirected what the public knew about because what was really happening was grim. We came to the point of annihilation. About 3 times during his presidency where people's fingers were on the nuclear button and they were like, they were pushing it in. Like, it was that close. And these were just totally, totally unknown to the American public, totally unknown even to historians for many, many, many years until people started really digging. And appearance versus reality is my favorite theme in literature, and I think it permeates our day-to-day life. And I think that's a really important lesson to learn early on, is that things cannot— things are more often than not what they seem on the surface. And you— here's an example of somebody who intentionally, intentionally created a narrative for the American public, basically to give them a false sense of security, because it was false. Things were not going well. At all. But that to me is, you know, real, real leadership. He sacrificed his personal well-being to take on that burden and try to solve those problems in a way that you didn't have people, you know, worrying like, is today the day when the bombs drop? They had no idea, which also goes back to, you know, kind of one of my core principles in life is that there's no limit to what you can achieve as long as you're not worried about who gets the credit for it.

Well, that's really deep. Thank you so much, Daniel. I really appreciate you spending the time with me today. It's a fun conversation. You shared so many examples. I mean, I think context graph is really hyped up right now, and you are actually being— you've been working on it before all these hype happening. Right. At the end of last year.

Yeah. It took us by surprise too. It's like, wait, we've been working on context graphs? Yes, we have.

Yeah. But I think you're truly working on something that's trying to, you know, help people make AI trustworthy enough to deploy, which is really, really important. And your view about, you know, the reality and fiction and showing up on all these examples that we talked about is truly, you know, revealing and inspirational. I think people should pay attention to this, right? This is happening around us every day. And if you're not doing your own fact check and you just believe what you're thinking, right, and follow the reviews, I think this may be a new skill we should develop for all of us in the AI era, which is, you know, more judgment. More critical thinking of what is truly the fact.

I feel like you need to insert the, uh, the Robert Downey Jr., you know, preach, you know, meme right now.

Preaching meme. Yeah, I will, I will be right back. Yeah, thank you so much. It's really a fun conversation. Don't hang up yet. I just need