Explainable AI with a Context Graph | TrustGraph

[00:00:00]and visualizations.

[00:00:01]So, we have some data.

[00:00:03]So, same thing.

[00:00:05]Why don't we go back and we can see again the same thing, agent query, tell me about Voyager 2. Ah.

[00:00:16]Well, we'll get back to that second.

[00:00:18]Let's just see. I want to do this.

[00:00:20]What model are we doing? Oh, I forgot to change the flow. So, this is actually using uh Quinn one of the flashes. I actually don't remember which one. So, we'll see the process that it goes through.

[00:00:33]And I actually haven't used this model in a while. So, again, another mistake that we're going to work through together and see how it does this. I don't >> [snorts] >> um Yeah, it doesn't seem to have captured quite as much that cuz this is one of the flash models.

[00:00:48]Which, again, this is why we have these explainability features to see how the model worked through this.

[00:00:55]So, what did we get here?

[00:00:57]It did a pretty good job, though. We're actually seeing the stuff that's really in the graph.

[00:01:03]And one of this is about cause of the improved semantic filtering algorithm. So, if we go back and we go to Voyager 2, we can see this is pretty much the information that it retrieved. There's this information on these flyby events.

[00:01:18]Yeah, we go back. That's pretty much what's here.

[00:01:22]And we could go through and do it again, but let's do it again at a different place.

[00:01:28]So, if we go here, what's a good time now to talk about, you know, some of the features that we have here. So, we have the ability to ingest your own data.

[00:01:37]My context graph explorer, which you have to have ontologies loaded to get this and you get this and you go, "Woo, isn't that cool?" It's like, "Well, yeah, I guess." But, what are you going to do with this? I don't know. Tell us in the comments down below if you have an idea of how to use this. We don't typically use these for knowledge exploration, but we know some people do.

[00:01:56]That's just not We don't tend to do that.

[00:01:58]When I say we, I mean me and Mark. So, I guess I should say that.

[00:02:03]We is pretty singular as in two people.

[00:02:06]So, let's look. So, this is the agent query.

[00:02:10]And we have This is the graph data navigator, which is looking at from a more traditional graph perspective, um which actually I I do like to use because of the natural language search.

[00:02:20]You know, we can we can I can go to the stuff about Pioneer missions.

[00:02:24]You can get a little bit more information about that. That's a pretty nice way of doing that.

[00:02:29]But when you're designing agents, and you'll notice that we haven't really done a lot of detailed design here. And we've done that on purpose so that you can play around with this yourself and see how the system changes. One of the really nice features that I like is this ability of testing. So, if we go here, we can actually debug this agent. So, let's um let's change the model. Let's see what GLM does. This is GLM 5.2.

[00:02:54]So, tell me about Voyager 1. I want to know about Voyager 1 this time.

[00:03:01]And this is going to give us Again, if I'm on this panel, we'll still see the answer. We're not going to see the actual graph. What do we actually want to see? We're going to see the tools that the agent is using. We're actually going to see how many tokens went in, how many tokens came out, latencies on the LLM calls. Why would we want to do this? Because if you're designing agentic systems now, you know how pardon context efficiency is and token efficiency is. So, these are things that are really, really important now.