So What? Intro to Knowledge Graphs

Well, hey everyone. Happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris and John. Howdy, fellas.

Hello. High fives all around. Well done. Uh, you know, it only took us how many years to get that one. And we mostly get it right.

And mostly, you know. Listen, we have a lot of expertise in other places, including what we're talking about today, which is the introduction to knowledge graphs. And so this was a topic that came up. Um, Chris was talking about this a couple of weeks ago, and we realized it was probably a good idea to bring it to our audience to really just sort of walk through what the heck is a knowledge graph. Uh, Chris, I remember when you showed me uh a couple of weeks ago, I was like, oh, this looks familiar.

It if anyone who's been with us long enough remembers the um graphs that we would show from events of who was talked about the most visually, it's very similar looking. It's a relational graph. It's how things are clustered and how they're connected. Uh, but without getting too far from ahead of myself, Chris, what the heck are knowledge graphs? What are we going to cover today?

Um, you uh you actually remember what reminded me to bring back this, uh bring up the old, old stuff. So and folks, if you you were welcome to actually, if you wanted to hit up the URL, you could. But this is so a knowledge graph is any entity relationship mapping graph where you have nodes, which are some form of entity, a person, an idea, a concept, a topic, a word, and then you have edges, the connections between them. So, for example, if there was a knowledge graph of trust insights, there'd be four nodes Katie, Chris, John, Kelsey, right? And if I were to take our Slack conversations, every time Katie mentioned Kelsey or Kelsey mentioned Katie, that would be an edge, a connection between the two.

And so if you were if you were to graph out just using our company Slack, you could see who talks to whom the most. Back in the day, and we are now talking, you know, oh holy crap, almost 10 years ago now. Um, we would use Twitter data because Twitter data was the most reliable and most open data format in the in the before times. Um, and what we would look at was who was the most talked about at an event? Because one of the things about you know influence measures at the time, 10 years ago, was it tools that measured social media influence had a bad tendency to measure who had the biggest mouth, uh, which is not super helpful.

Um the the it was the old 1980s financial services commercial that we the uh tagline was when EF Hutton talks, everyone listens, which is sort of the way that we think about our friend Mitch Joel you look uh said, which I think is a great summary of influence. It's not who you know, it's who knows you. And so these graphs represented a lot of that. Well, they've they came and went, you know, Twitter got bought and became X and new social networks are much less um generous with their data because all the social networks today are owned by AI companies that were keeping all of that data for themselves. Uh so it is very difficult to get this data.

However, the concept of a knowledge graph of saying what are the words and phrases or topics uh is more relevant than ever for one simple reason, and that is context window. Anything before I move on to that. Uh probably not useful, but Chris, isn't this just a word cloud? No, it's uh a word cloud is what is said the most. So going back to earlier, it's who has the biggest mouth, right?

If in fact, this kind of relates to what we were talking about before the show. If conferences chose speakers based on just who foamed at the mouth the most on LinkedIn, you'd end up with the most awful conference in the world. But I think it's an important distinction, um, because that was a question we got a lot when these network graphs were published. So first, people were thrilled when they saw their name show up as a bubble because it meant that they were being talked about a lot. They were, you know, being tech tagged in tweets, people were resharing their content.

Like it was a big deal. It was basically a you know, a point of pride to see your name show up as a prominent bubble, and then all the different connections. But we also had to make sure that we were clarifying, no, it's not just a word cloud, because to your point, a word cloud is the most mentioned person or is the most um the person who talks the most, you know. So I wouldn't necessarily show up because I'm not someone who admittedly posts enough about what I'm doing at a conference. But if I've done my job right, people are posting about what I'm talking about and tagging me.

And so at marketing profs B2B, you would expect someone like Ann Hanley to show up as a very large bubble as you know, the MC and the face of the uh event. If she doesn't, then she has not done a very good job as in that role because people aren't talking about her, or they're talking about her in a negative way of like, wow, she really didn't do a good job. So it's like it's really interesting to see how those bubbles came about. And it was helpful at the time for event planners to say, oh, who is being talked about the most? That's the person I want on stage.

Exactly. So this technology is still useful because today's AI tools, when you're processing data with AI, only have so much memory. They all they have a short-term working memory. It's called a context window. Now, the context window is how much um how much short-term working memory a model can remember before it just goes kaboom or has to compress.

If you ever used, for example, clawed code or clawed co work, from time to time to say, hey, I'm compacting my memory so that I can keep working. What it's basically doing is making a summary. Now, if you have a lot of data that you're working with, and it's all valuable, it's all relevant. That's a bad thing because what happens is as things get summarized, you lose a lot of details. So wouldn't it be great if we had a way to provide access to that memory to the source documents in a way that didn't overwhelm its memory?

Claude code, Claude Cowork, Open Work, OpenAI Codex, Google Anti-Gravity, all the agentic systems of today can read files off of your disk. And that alone is a big step. However, even that can get overwhelming because sometimes you just have a lot of information, and you need some way to almost have like a library system to navigate it. Now, there have been a lot of conversations in the last six months about how do we implement this. Um, Andre Carpathi, former co-founder of OpenAI now, who just moved to Anthropic this week, which was a shock to everybody.

Um, has proposed wikis, right? We're we're we're throwing it back 25 years to to wikis. I was like, wow, there's a new concept. Uh-huh. Um, and those are good uh situationally, however, a lot of data it takes a while for a language model to traverse wikis and it it can be very processor and and memory intensive still.

There are other uh systems out there, for example, in the last uh month, uh Resident Evil and Fifth Element uh co-star Mila Javovich and and her partner released uh MemPalace, which is their a similar concept system using compressed text to to help a model remember things, but that was not on my bingo card for this year. I feel like you just said a bunch of words. Like I got the resident evil reference, I know who the actress is, but what is it that they released? They released a similar system, a kind of compressed wiki to try and help AI tools remember things. Uh and from what users have said, it seems to work reasonably well.

But we're gonna throw back to much earlier and go look. And knowledge graphs are probably the best way to do that in an AI world because a knowledge graph gives you the ability to search across a network of concepts without having to read every document. And that's what because that's what kills memory in an AI is it has to every time you say, Hey, I want you to you know help me write a blog post about this topic, and you give it all these supporting materials, it's like great. You can see Claude, for example, do this. I'm gonna go read all those documents, and then you watch its memory blow up because you're like, Well, there's you know, one of those documents was 30,000 words.

That's that's not super helpful. If you have a network, uh a knowledge graph, which is a network graph, of all the concepts, a tool like Claude can go look at the graph and say, Well, just looking at this graph. Katie has asked me to talk about enterprise AI. What do I know about this? What I what's available?

And it looks and defines Enterprise AI and it finds all the concepts linked to it, and in a good knowledge graph, it has little pointers, little indexes to just those documents. So it no longer has to read everything all at once. Now it can read just the just the relevant ones, piece together things and come up with uh useful answers. So that's why knowledge graphs are suddenly hot again, because it's a way to preserve AI memory and point it only at the things that it that really matter for the given conversation. Is it if we think about it in like a more old school sense, is it basically like the index in a very large book?

If you're just looking for, you know, if you're looking at a cookbook, for example, and you just want to know the recipes related to chicken, then you go to the index and you find all the different versions of chicken recipes, figure out which one you want, and then flip just to that page versus spending your time going through page by page by page. Is it that is that kind of the idea? It's kind of the idea. Now imagine because that's a good analogy. Imagine you have that index, you tap on the word chicken because you're also looking for uh onions, and so all the chicken recipes, then you look at that chicken note.

Oh, look, there's onions. You tap on onions, then you and you can and then start to see sub-relationships, which is a super helpful way to examine the data. Um Brian has a question. Uh, what are your thoughts on obsidian and its knowledge graphs, especially related to Carpathi's second brain idea? Thank you, Ryan.

Yes, a lot of people are using Obsidian as that. Um, you don't have to. Uh their implementation is not terribly unique, and that's fine. Uh Obsidian is a uh piece of proprietary software, sort of a personal knowledge base that links up different data and allows you to create sort of a uh a master index. I use a version of uh a system, I use two different systems.

One is called Joplin, which is just a collection of notebooks, very similar to Evernote. And for a knowledge graph, uh, there's a completely free and not vendor locked in uh tool called log seek that does the same thing. You can give it a folder of documents and it can create graphs around it. It's one of the things that I'm going to show. Let's look at that.

There's there's two super powerful uses of these things. The tool that I use these days the most for this is a tool called Graphify. Free, open source, completely private, doesn't require an LLM, but works best with an LLM. And what you would do is you install it on the command line uh in a given directory, and you say, make a graph of what's in here. Now, if it's if there's code, it doesn't need an LLM at all because code automatically has references in it.

Like you have functions that call each other, libraries, and so it can make what's called an abstract syntax tree, which is again like that cookbook index with all sub-indexes automatically. And that holy crap, will that save you a lot of time in a tool like Claude Code? Like it will just it will blow your blow your way about how much faster Claude Code becomes and how fewer mistakes it makes, which is a lovely thing. It can index all the documents in a folder and provide that exact same knowledge graph. Questions before I my bra my brain is already spinning.

I'm like, okay, I already under I got it. I understand how this would be applicable to me. So for example, um, the way that I organize all of my projects that I'm working on with Claude Desktop, is I literally have a folder on my desktop that says Claude Organized Files, Claude Coworker Organized Files. And under that, within that directory, are all of the different projects. And a lot of times, um, because of the nature of the work that I do, a lot of the work is relational.

And so, like it could be something I'm working on for um an event where I'm working on a talk, but the talk is based on my knowledge of what we do at Trust Insights and the strategy, and then there's a whole folder on the strategy of trust insights, and then there's you know, folders of content around that, like, and so they're all related. And so a lot of times what I'm doing, and this is probably burning more usage than it needs to, is I'm giving Claude, I'm like, here's the master folder. There's stuff inside every single one of those projects that is applicable to this new thing that I want you to help me work on. And so it sounds like if I'm understanding correctly, if I had a knowledge graph of the contents within each of these subfolders that gave sort of like connected all of the topics and what lived where, I could probably save some usage within my Claude desktop instance. So, because then I wouldn't be saying like search everything and find what you need, because that's what it does now.

Right, exactly right. So I win, I quit. So let's look at an example of what a knowledge graph might look like. I'm gonna show you a code one first. So uh the trust Insights tool, if you go to trusted sites.ai slash AIView, this is our uh AI view tool that allows you to do GEO assessments uh for the third phase of GEO relevance, which by the way, if you haven't taken our our not free but not terribly expensive course, uh at Trusted Sites.ai slash GEO101.

Um, this is a tool that will allow you to either to inspect a single page or compare two different pages to see how they stack up. On the back end of this is a lot of code, a ton of code. And so this is what the knowledge graph of the AI view code base looks like. And we can see there are some uh big clusters here. Uh there are things like BM25, which is a search algorithm.

Uh there are the BM25 test library here. There's the HubSpot client test library over there. And you can see a lot of there are some of these like like rate limiter tests there. There are a lot of dense nodes that have a lot of connections around them. So when Claude, when I say hey to Claude, hey, I don't think this thing is working right, you know, maybe it's the rate limiter um that prevents you from just using the tool all day, every day.

Um, Claude can say, okay, you have a question about the rate limiter. I'm gonna I'm gonna look at the graph and say, oh, there's a rate limit. And these are the files and functions involved in rate limiting. I don't have to go and like you were just saying, Katie, read the entirety of the code base. I can go and see what specific files or functions are part of the rate limiter, and then I can go look at that that just that code and say, Yep, this is broken or no, this is working fine, the user's an idiot.

Um question is, and maybe I'm jumping ahead, so tell me if I am. Uh, is the success of a knowledge graph dependent on the governance and consistency and naming conventions of your files themselves? Or is it really just like a, you know, find a keyword and list all the places that you found it? Um it can be both. One of the useful things with a knowledge graph is you see a lot of these little dots around the edge that don't that are kind of stragglers.

Those are indicators of things like dead code, bad governance, right? Uh naming conventions that don't make sense. And the more of them you have, the more likely it is that you've got some technical debt that's added up. So this is a useful diagnostic tool as well to have when you say to Claude or the coding tool of your choice, hey, go and simplify my code base, look at the graph, look at the stragglers, and figure out why are these little hanging stragglers hanging up there? Or hey, there's a big chunk that like this this big chunk here from this method.

Why isn't this connected to anything else? Like, what is that thing? And do we need to be concerned that it's just hanging out there? It appears to have a lot of things connected to it, but it doesn't connect to the main section. We might have dead code, we might have features that we didn't implement properly or never wired up.

That happens a lot. Uh, and so it's is both diagnostic and it does benefit from good coding standards. Like if your coding standards are good, your naming conventions are good, um, you have clear expectations, then this graph will automatically look better than if you just kind of vibing your way through it. I want to take a moment to acknowledge what you just mentioned. And so that a knowledge graph is a really efficient way to start cleaning up technical debt.

For those who are unaware, can you at a high level explain what technical debt is and why you should be cleaning it up in the first place? It is duct tape and chewing gum. So a user reports an error and say, I'll patch that, and they write a fix. And then they say, I'll patch that and I'll patch that. And suddenly, like, you know, like those cars that you see where like the registration sticker is just like an inch wider than the license split involved in the license plate because they keep putting the sticker on over and over and over every single year until like it's it's sticking away from the plate.

That's what technical debt is. Technical debt is just patches on top of patches, band-aids on top of band-aids that because no one ever fixed the underlying problem. And over time that makes your code harder and harder to maintain because it's like you go to fix one thing and it breaks five other things because it was all like a house of cards barely being held together with duct tape, chewing gum and bailing wire. If, and this is something that that human coders hate doing, but machines are superb at if you have a great knowledge graph and you have a way for it to see the system as a whole, you can say this big file here, refactor this. You gotta you know you can remove the technical debt and basically almost reboot that file.

And maybe it's three files now because you've had added so many features and functions. Technical debt adds up when you just band-aid things all the time, instead of saying, Well, let's take a step back. Maybe we need to revisit the requirements, maybe we need to revisit the spec. Maybe we need to refactor or rebuild part of the system so that those features are integral to it and not just add bolt-ons. And I think that you know, part of that uh explanation includes it makes your code less efficient, it makes it heavier to you know use up more bandwidth to take up more space.

Um, you know, John, as someone who does a lot of repairs and construction physically, you know, I know you spend your weekends uh, you know, doing like home repairs and that kind of thing. Like, is this something that resonates with you, like in terms of that analogy? Like if I could, you know, redo all of the wiring and get it right the first time, I wouldn't have to, you know, like jerry-rig this, you know, box over here and hope it doesn't set on fire. Yeah, it definitely you know ties into that kind of example. But I I actually spent a number of years in software development with a system that did code management like this.

And so, yeah, this idea of cruft piling up, there's so many things that are in that bucket of, you know, okay, the last version of the product use this database system, and then we transition to this one. Well, who's gonna go back and pull out all those chunks of code that we're calling the old thing? And being able to run a diagram like that that just shows, like you said, you see those freestanding clusters that no longer call anything. You know, you know that those just need to be paired off. So it's yeah, in every system, cruft and entropy is a problem, right?

It doesn't matter whether it's yeah, you know, I have some plumbing uh in the bathroom over here that has three different types of connectors, you know, it's like metal to plastic to PVC, and you know, you do that, and yeah, you always reach this point where you're like, okay, I could repair this a fourth time, or maybe we just cut it all out and put one pipe in there. You know, maybe that's the better way to go. So yeah, but this is just really fascinating at a bunch of levels for me because I had transitioned. I'd like I really want to put a pin in uh well and call out the fact that these were hugely successful for us when we initially did these because these network graphs would have like five or six what we would call influencers now on these things when we posted them on LinkedIn, and they would run completely wild. Everybody would want to brag about how they were the most important person at this show.

It was literally throwing the crackpipe to LinkedIn, and you know, they just went completely crazy for this stuff. And it's funny now though that I'm being dragged back to oh, wait, this isn't just like a marketing stunt, like these do have very important functional implications. And the one that's blown my mind today, too, and I'll I'll you know get off the soapbox, but it's the idea of using a network graph as a dynamic index for an LLM. You know, that never drops it because you hit it. You've got like, okay, I've got all my prompts and skills in this huge pile, but if I just ran a network index of that, I could then feed it to the model, and now it won't have to churn all those every single time.

It will just go cherry pick what it needs, and you look you're basically showing it like here's the model, and it knows that oh, I only need this cluster to solve this problem. And so that's you know, goes right to the uh saving tokens. You know, you're just you're making everything run faster and easier, and you can dynamic dynamically update it any time. So let's take a look at a second knowledge graph and and look at a non-code example because a lot of folks are not coders, and that's fine. Um, this is a knowledge graph of my LinkedIn posts.

So one of the things I do is I write all my LinkedIn posts in in the system uh Joplin that I uh and they're all in markdown. And the same graphify software you can run in any folder of any content, you need a a language model to process it because it's not code. Uh so for example, in Claude Code, I will use uh Claude's haiku model, which is the super fast, lightweight cheap model to build the graph using the Graphify software. Again, graphify is part of it. Um, but what you can see here is this looks very different than software, but you still do see clusters of things like retiring old AI advice.

Uh you see this big cluster middle, of course, generative AI, which is something I talk about a lot. Um, you see over here the new LinkedIn guide. So the Trust Insights unofficial LinkedIn algorithm guide is here. And there's a lot of, I've done a lot of posts on LinkedIn about that. And you can start to traverse uh and then tap in on it and just you can see what the individual documents are within this corpus.

Now, this tell this is an index essentially, like Katie was saying, like a book index of 272 LinkedIn posts that I've done. Um, and so I can look at this and look over time and like what are these little things out here that like oh the Trust Insights Learning Survey, uh I don't even remember what that is, AI and inequality. Uh we have Google Workspace Studio. So there's been little things I've covered here and there, but other things which are you know very, very interconnected would show that I've talked about that topic overall a lot. I could see, and I might have to write this down in our Slack uh for it to talk about later, is I can see using a tool like this, or companies asking to use a tool like this, which I'm just gonna throw throw this up here for a hot second um to vet whether or not their quote unquote experts are truly experts and what they are really talking about and if there is some sort of like a real narrative to what they are an expert in or if they're just you know jumping on the bandwagon and posting one viral post about generative AI and then everything else is about uh NFTs or something that's you know unrelated to the thing that they're being hired for like what is this person best known for talking about like I could see this as a way to sort of like start that process of vetting you know all of the different experts that are in the field right now that are you know suddenly experts in AI like well what does that look like are they really experts in it how often are they writing about it and then you can dig into okay they have you know 50 out of you know 80 posts that are about generative AI let's go ahead and see the content of those like it I can see this as like a starting place.

Mm-hmm. Where this would become super useful for me as a as a content creator is I could say you know what Claude or the tool of your choice I want to come up with a new book but I don't feel like I have the time to put one together but I've written close to 3000 words of content and a lot of it is around AI topics. Go to my knowledge graph. Look at the communities that are the biggest that I've talked about the most that are related to generative AI. Then start tracing your way through them and then start extracting my real words.

So don't, you know, we're not making stuff up here. Extract my real words in logical order based on the graph to help me assemble a new book about generative AI. And use my words as much as verbatim as much as possible, because you know we want to we want to preserve our real voices. And with this knowledge graph, now again, it doesn't have to remember all 300,000 words, it can't, but it could traverse the graph and go, oh, what about this? What about this?

I could pick a little from this, a little bit kind of like a chef in a kitchen when you know we have a five spice track, like 5,000 spices up on a little this, a little of this, a little of this to make something. And because the graph has mapped out the topology of my language, it will be able to assemble things that may and in patterns maybe I don't even see. Maybe I've never thought of, oh, you know what? I talked about this topic a lot, but in different LinkedIn posts spaced out by weeks or months. I've forgotten that I've actually covered this topic a lot.

That's the way that I could see someone like me using it is you know, I'm the first to say, I don't think I have anything to write about in a book. And then Chris virtually slaps me and says, Are you kidding? So it would be helpful for someone like me to see, like, oh, this is the clusters of things, and those become the chapters of the overall theme, which is probably something around I'm gonna be uh shocking. Let's see if I can find it. Uh the 5P framework by Trust Insights.

Uh pretty sure if you want to learn more, go to trustinsights.ai slash 5p hyp hub framework. I'm pretty sure that uh I write about it enough that I have uh you know enough not you know material for a book, but me as a human, I sit here, I'm like, what would I even say? Where would I even start? And see being able to see like not just the five P's, but the clusters of topics and those becoming the chapters would be super helpful. So let's go ahead and you know kick that project off behind the scenes.

And what this also does is this tells you if you know this, if you know what your knowledge graph looks like, if you were then to commission four or five deep research reports, say on enterprise AI governance, you could put them in a separate place, create a graph of that, and then give Claude access to both graphs and say, okay, what in the overall landscape of enterprise AI governance is in these reports that is not on my graph that I need to fill in some gaps. If you're sitting there going, uh, I don't know what to put on LinkedIn this week. I don't I want to say something intelligent, but I you know, I don't want to be a me too, I don't want to jump on the bandwagon. Oh hey, we need to talk about network topology within large corporations as it relates to AI because things like latency and your intranet, the physical wiring in your building can dictate which LLM software you're going to use for your company's internal AI hub. And you can say, I've never talked about that, but and I don't see anyone else talking about that.

And yet you go to a company, a big enterprise, and the CIO and the CTO are like, our network topology is looks like you know, duct tape and chewing gum because it are you know, we've grown organically, we've acquired 28 companies, and now we want to have this enterprise AI initiative, and we're like, we can't even get our network to work properly. I can also see a use case of this for, you know, and this is what you're mentioning, but to put it in concrete terms, companies you know, reaching out to use this for their content strategy. So, you know, a company might say, we like to think we talk about X topic a lot, but then when we actually take all of their content and put it through something like this, it could be a very tiny bubble, and then they have all these disconnected pieces. So I could very much see this being, you know, the starting point for a lot of companies where they want to tighten up their content strategy because they're not getting known for what they want to be known for, especially in a large language model. If they're not writing about the things they want to be known for, they think they are, but they're not, or there's an opportunity to, you know, they have these small little clusters that they can build on, that might be a great way to visualize oh, this is these are the topics.

Um without iterating, you know, a knowledge graph. It would there be a way to like marry SEO data to something like this that so that you could say, you know, here's what I want to be known for, here's what I'm writing about. What's the white like what's the Venn diagram of it? So you're getting into what will become the trust insights 30 GEO 301 course, because I would just finish recording 201. 301, it looks at advanced stuff like this, because one of the things that you can do is you could take the same algorithms that language models use for things like vectorization and embedding, um, and decompose it against the actual model embeddings.

We just did this when Google's new Gemini uh flash model came out this week. And you can find the semantic white space that is missing, or you can take what actually shows up in search listings after Google has done its own Googling, examine the results, and say, okay, well, where are you would create a knowledge graph, right? Which is what honestly what really embeddings are, right? An embedding space is just a really big knowledge graph of math, and then say, here's who we are and what we talk about. Here is what is in the semantic space inside of a language model for this topic.

Gosh, these don't look alike. And and I wonder why, and oh, maybe that's why we're not anywhere in any of the the results that are shown because we haven't said squat about what the the language model knows is important about this topic is high probability about this topic. Now, when we teach that in 301, that's going to be like, yeah, you better be good at Python, ladies and gents, because you cannot, there is no consumer-friendly, no tech way of doing this. We'll do our best, but you are getting that's why it's a 301 course. Or just throw it out there.

You can go talk to John. We have a guy. We have a guy, we know a guy. Um I mean, sorry, all like sales pitchers aside, I think it really is interesting. I can see it being incredibly useful.

Um, you know, I'll be the first to admit like I feel like my content is really disconnected. I know logically that's probably not true, but being able to see something like this to help me really understand the data would be great. Because I'd be like, oh, okay, this is really what it looks like in terms of maybe I'm overindexing on a certain topic within the five Ps, and I really need to pay attention to these other things over here to make it more well-rounded, or maybe it is well-rounded, but I don't know because I'm too close to it. And doing a content audit is something that a lot of people still do. It's still a very important skill, and companies get hired to do it.

But generative AI is sort of setting the expectation that you should be able to do a content audit more efficiently, but that's just eating up a lot of usage. And a lot of people are doing it haphazardly because there's no good systematic way like this to really put all the connections together. Exactly. So you look at, you know, when you look around in these different spaces in here, um, you start to see what those themes are, what those themes are going to be for like, you know, whatever the maybe I I'm looking at this myself, and I'm thinking, okay, well, I have my old 48 principles from last year's book. Maybe the knowledge graph goes through and looks and says, Well, Chris, you have introduced 12 new principles since last year, or you've retired six of them.

So in the 2026 edition of your book, you use the knowledge graph to extract that information out and say, okay, I'm we're we need to do some updates. This is we are not so it's funny, knowledge graphs date back to the 1970s. Uh, and they were really popular in AI back in the 1990s, in what was a branch of AI called symbolic AI, uh, where you have things like ontologies, it was all rule-based AI, kind of like you know, uh chatbots and stuff, you know, press one for this, press two for this. That was the popular form of AI back in the 90s, uh, you know, when hammer pants were still in style. That fell out of favor because it was too rigid, it was too inflexible, um, and it ran into a lot of problems.

Then in the 2000s, the 2010s, and this decade, neural AI, aka, you know, language models and things like that, neural networks, deep learning are all the rage. But what we've found, and you know, consumers have learned this the hard way, is even with good prompting, they can still generate wildly inconsistent results from time to time. So, what a lot of folks in the field have proposed um is those folks from the the 90s bringing back symbolic AI into what's called neuro symbolic AI, where you have a graph, an ontology, a framework, guardrails that are fixed and say this is what the layout of the book is going to be, and then neural network, you do your thing in generating stuff, but only in the confines that the symbolic network has dictated. So you can't go off and write about chicken curry, right? No, you have to be working in this section on copyright tests, and so the symbolic knowledge graph and ontology dictates what the model is and is not allowed to talk about.

And so you end up with much higher quality results from generative AI. In the same way that if you know it, for example, in Claude Cowork, you say these are the Trust Insights brand standards. You must use these colors. You are not allowed to use other colors. That set a guardrail says, okay, so I have to be creative within this constraint.

I can't use colors other than these. How can I be creative with the colors that I've been given? Question. Um something that is still a very useful tool, but was always a little bit of an oddball tool compared to the other large language models was Notebook LM. So Notebook LM is part of the Google ecosystem, still very much a tool that people use today.

One of the things that stuck with me with Notebook is that you could give it a bunch of resources, and it became like a wiki where you could only search within those resources. So you would create literal notebooks. You know, is it creating its own network graph on the back to say, here's all the things that are connected, so I can answer any given question about this because I'm constrained to just this data? To a degree, yes. It's not a semantic knowledge graph like what we just showed.

It is what is called retrieval augmented generation or RAG, which is a term you've heard a lot. Yes. This is a knowledge graph that Notebook LM builds. It transforms all the documents into the same kinds of tokens and weights that the big language model uses. And it essentially says to Gemini, which is the back end model, you must query this network of weights first.

So it is a knowledge graph. It is, but is a knowledge, it is not a semantic one of concepts, is instead the same tokens that make up the big language model. But uh Notebook LM is uh what we call a hybrid system. So it has the knowledge graph of its uh weights and tokens, and it also has the structured data because when you ask it a question, um, you can say what source did you get this from, it can go and hunt down that source. A lot of AI companies these days use databases like Postgres SQL that has vectors and embeddings in it so that you have a knowledge graph in the database, and then you have structured data, rows and columns that we're all used to, so that it allows AI to say, I'm looking for this concept.

It pulls up the rows, it pulls up the the vectors and then the vectors map to rows and says, okay, now I'm going to read the verbatim stuff that I'm going to do because there are a lot of cases where the verbatim text is mandatory. It's one of the reasons why graphs, uh, you know, uh AI graphs, rag, failed spectacularly with coding, because you can't have concepts in code that you need to have exact verbatim text because one misplaced semicolon and the whole thing blows up. Right. But for a tool like Notebook LM, yes, it is a hybrid tool of both a knowledge graph that you can't see that is a a rag system plus the structured data. Okay.

That makes sense. I mean, you know, when you were describing it, my first thought was this sounds a lot like what something like Notebook LM is capable of producing. Not at the same, no, you know, obviously, to your point, not exactly the same, but um, it produces these mind maps, which Chris, you're showing on the screen of all of the nested topics, related topics, um, which is super helpful. Um, question in terms of use cases could someone like John use a knowledge graph to understand all of the uh new business calls that we've gotten over the past year to figure out here's what people are asking for. Here's like the main cluster.

So we should probably create a campaign around this. And here's like the three edge cases that we should probably discourage people from asking us about. Yes, you could do that. You would probably want to build it into a system. So the graph will tell you the relationship of concepts to each other, but a graph needs some steering.

So what you probably want to do is you probably want to have a like all of the call transcripts and the notes and the emails of clients you won or deals that went very far versus deals that didn't go anywhere and stuff, and be able to look at the two knowledge graphs side by side and say, what's the difference? Why is this one different than this one? I think there's definitely that, but more of what I was asking was what I guess what are the pain points that people are describing. So, yes, I want to know which deals closed and which ones didn't, so that we could do more of that, but I guess I'm more curious about what pain points are people talking about, and are they you know, how are they related so that we can really understand what services like I'm thinking it of it in terms of like if John were to go through and audit all of the services we offer and say, all right, for 2027, we need to ditch these three because nobody cares, and then we need to double down on these four up here because here's all the data that says we address these pain points, but here's how it's talked about, so that we can update our web page with that that language of the pain points. But I feel like right now it's still a little murky because someone might say, you know, I'm really far along in my AI adoption journey, but I just need help with implementation, and someone else might say, you know, everybody's on board and using it.

What do I do next? It's the same problem, said two different ways. And so I'm wondering if using these knowledge graphs to help bring together these clusters of topics, uh semantically, is a way that we could be looking at that information. Not only would that be a good way, we could actually do that probably before the end of the day today. Uh because yes, we have so we use an AI transcription system.

We have a decent naming system for our our calls, and years ago, we wrote in a connection to the APIs and that Mondays and Fridays, when I run the company update, it downloads all the transcripts from that week. So that and it stores them in a folder. So we literally have a trans a folder of all the transcripts for you know, uh at least two years worth in markdown format, which is what AI and knowledge graphs love. That just run graphify on the folder and say, okay, show me. Yeah, I love it.

But I think what's interesting is like what I'm trying to think of is so for you know, our audience and those watching is like, what are the different use cases for using this kind of technology? Because you know, you started off talking about you could use it on your code to help, you know, understand like where you might have technical debt. We kind of moved into you know, understanding what all the files are that you have so that you can not waste usage in your LLM system. But it's like I want to make sure we're thinking through those other concrete examples, and that to me could be like a really powerful one for companies that are struggling with what is it that we do? What do we sell?

How do people hire us? How do we speak the language of our customers and really meet them where they are? You can also use it with named entities. So this is something that uh has been done for years. Microsoft did this.

Um, if you were to take, say, all the papers from NERIPS, the one of the big conferences, all 5,000 of them, and have AI instead of using topics, use named people. You could see the relationship of who is cited the most in academic papers because they are the influencers in the academic space, they are the people who are quoted the most. Um, somebody did this with the Epstein files. They took all 3.5 million files, digested the names out of them, and said and built a knowledge graph of here is who is most talked about in this huge corpus that otherwise is so vast you can't process it. You know, anything where you have a very large quantity of data, you can use knowledge graph to understand the structure of that data.

So I mean, that to be, you know, sort of the meta use case. If you've got a lot of data and you want to understand the semantic or the logical structure of the data, not the files and how the name, but like what are the concepts within a big pool of data, like all our sales calls? A knowledge graph is the application to use for it. And I think that that's a really good point because I feel like right now the default is just give like Gemini all of the transcripts and say summarize this for me. And so you're gonna get like six big key topics, but no real good insights.

And I feel like this is sort of like that next level. Summarization isn't good enough. This is better. Yep. If you have a long-running client, you know, a big client, you take all that client's calls and transcripts and stuff like that and say, what are the themes that keep showing up with this client over and over again?

Right. Uh, you have a call center. What are the things that people complain about the most? Because as you're doing the language processing within LLM, you can also add features to the knowledge graph to say, you know, uh sentiment count, show you know, uh show edges that are negative conversations in red, show positive conversations in green. And suddenly you can see not only the graph overall, but you can see what are the real pain points.

Yeah, that's I mean, a lot of that is, you know, when we're using the LLM, it's data mining. You know, we know what we're trying to get, we're digging down to get a single solution, but this is really data exploration. Like you kind of don't know what's out there. You need something that's uh uh it's a whole different UI and it's a whole different experience of raising, you know, all these different pockets to you and letting you know what's in there and be able to ask questions of it instead of like you said, just getting the summarized list, which is not giving you the picture. Yep.

Even even strategy. Um, if you were to, because we've done this, if you were to take all the open job positions of a competitor and build a knowledge graph of the concepts, the key skill areas, you could see what a competitor is hiring for at in the big picture and go, Wow, I know exactly what it is that that they're going to be doing for their corporate strategy because they're hiring for these eight things as much as possible, and it keeps showing up over and over again. Katie is having ideas. Katie's jotting down notes. Sorry, not sorry.

Uh-huh. Yeah, sorry, not sorry. You're both about to be very busy. Yes, that's an intro to knowledge graphs. Now, in terms of what software you should use, that depends.

Um if your company already pays for co-pilot premium, you have Microsoft knowledge graph on the back end. You will have to talk to your IT department or how to get access to it. Um, but that is available. Uh at for AI specifically, the the free software uh grapify is excellent. We use it.

Uh and I do recommend it. If you are doing it a large scale project, the gold standard is a system called uh Neo4J, that is a system you it's a big server that serves up, you know, billion node knowledge graphs, and they're used everywhere. So LinkedIn is a knowledge graph. In their technical papers, they show that they have a six petabyte in memory knowledge graph that spans the planet across six major data centers that is updated every few seconds. That's how big the LinkedIn knowledge graph is.

But there's so the the reason why LinkedIn runs on a knowledge graph is because they're effective and they work. Any final parting thoughts. Will you drop those links into our free Slack community analytics for marketers for those who are interested? We certainly can do that, yes. And we're gonna have some additional benefits coming soon for the analytics for marketers community.

Just have to uh finish building them. I gotta get my surprises in where I can. I gave you that one free pass and you didn't use it. Yeah, well, I'm using it now. Uh I'll tell you about it later.

Uh that's gonna do it for this week's show, folks. Thanks for tuning in. We will talk to you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you're watching it.

For more resources and to learn more, check out the Trust Insights Podcast at TrustInsights.ai slash TI podcast at our weekly email newsletter at TrustInsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free analytics for marketers Slack group at TrustInsights.ai slash analytics for marketers. See you next time.