Selector AI Copilot: Talking to Your Infrastructure

So talking to your infrastructure, co-pilots are becoming the only way to develop software, right? And typically software is a little bit ahead of hardware, for the most part, in tooling, not anymore, not with selector, right? Think of questions.

What would you ask your network here amongst the delegates, right? What, what kind of questions in your own language would you ask your network if you could? Connectivity questions, routing questions, streaming questions, capacity, security, SLA, questions.

So we're gonna see this in action with a quick video. And let me make this full screen and we're gonna spend a little bit of time on this video. So in the copilot, I'm asking to please show my Juniper device inventory.

Now what's gonna happen is we get two responses. We have a visual representation in this case, a tabular view, and then we have the natural language answer at the bottom. So you get it in both forms.

A visual representation and the natural language. Now that table works in this copilot search filter, find sort, whatever. Now you're gonna see me filter down on a single device, and I can copy this to my dashboard.

So dynamic dashboarding, I'm not overwhelming you with some default dashboards from your natural language. You get the image. And then I can send it to my main dashboard here in just a second.

I'm gonna expose the APIs. So under the info button, because it's a platform and we encourage automation and development, we show both paths, the natural language query as an API and the SQL query as an API that you could use in your own automation to represent this dashboard or the natural language. It truly is an open platform.

And that we expose our API system, is it playing? And you'll see here in a second, the body of that API comes up here that's in your copy paste buffer. Now you could put that in Postman or Curl or Python.

I'm gonna scroll down a little bit and we're gonna see the natural language answer, which is completely accurate. 67 devices from Juniper. Six spine 61 Leafs.

The model breakdown, the operating system, the software version, the hardware version from one question. Device inventory, enter. That's it.

So now I'm asking for device health. Device health is similar in that I'm gonna get the visual view and you can see that there's, you know, six or seven devices that are unhealthy. I can click on any of those and I will in a little bit in the video.

And then we get the natural language breakdown of the health of our devices. I'm gonna send that widget to my main dashboard and build a story in my dashboard as I go. All driven by natural language.

And so now I want to drill down, show me just the unhealthy devices right away. It's gonna filter down to just the five unhealthy devices. Again, I could send that out to my dashboard.

And we're just waiting for the natural language response. And now I can actually ask for interface health. This is my phrase, interface health for S3.

So give me the s threes interface health. And now it has the, I have a single interface that has some problems and it even tells me in the words, right, you have one interface that's in a failing status that marked as violating. So now let's go further into that interface, which takes you to the per interface dashboards.

And we have the health traffic in traffic out. And if we scroll down, we're gonna see that this interface is, is flapping. We have an interface that's flapping here, up and down, up and down, up and down, up and down.

I'm gonna send this back to my main dashboard. So now I have the full story in one dashboard view, all driven by natural language. I'm not fighting with a GUI looking for what menu things are under.

I don't have to memorize a bunch of CLI commands. I don't have to come up with a complex rest post. I just ask it.

So we're gonna, any questions? Cool. Demo.

Uh, uh, this also brings to mind, has, uh, prompt poisoning protection come up in your discussions? Yeah. So we are very conscious of our guardrails.

If I ask this, what's the weather in San Diego? I cannot answer weather questions. Please focus on network questions, right?

So guardrails are very important. Naturally. Yeah, naturally.

So then I'm gonna show you, uh, at a full screen here. This works from anywhere. So my next slide, move ahead here and maximize, this is from my mobile phone.

So I've shared my phone to my desktop and did the recording. I go into a room from my phone slash select, invoke the bot device health three in the morning in the car with the kids at the soccer field, wherever you are now, we're gonna get, now we're gonna get, there's our visual view. If I was on the VPN or the local network, I could click the link and drive me to the portal.

Now if I want the natural language, 'cause it's a different API slash select pound copilot. Invoke the copilot. And now I'm gonna ask for device health for S3.

Since S3 was one of those red devices up above, I just keep drilling down with natural language, but I want the words, I don't want a dashboard. So I invoke the copilot and there's the copilot. Now let me pause this very quickly.

Talking about RCAA, did I ask anything about root cause or what's going on? I just wanted the device held for S3. The data indicates both S3 and L 1 21 devices are experiencing issues with interface traffic in both directions.

Pointing at the link between S3 and L 1 21. It wants to tell you what's going on, even though I didn't ask for any specifics. Right now, we live in a multi global society, global village staff everywhere.

Customers everywhere. I'm gonna ask for device inventory in Juniper in Spanish. Okay, that's it.

I just placed it in in Spanish. I'm gonna get the dashboard view, but my copilot answer, it's going to be in Spanish. We've tried Greek, Hindi, German, Italian, it, it knows Sanskrit.

We've tried crazy things. There's the answer in Spanish. Now, if you pasted that back to Google it, it's a valid answer in Spanish.

Now, really quickly, if I can pull this off, I'm gonna do one live for everybody just to show that this is not staged in any way. If I could get my phone to behave, so I'm mirroring my phone and if I go into my Slack and find the room where I've enabled my selector, I can say slash select and I'm gonna invoke the copilot and I'm gonna say device health for S3. Please answer like Yoda, just have some fun.

I'm sure it's been a long couple days for everyone. So the, the bot's been invoked. I've sent the co-pilot, we're just waiting for the S3 health back as Yoda.

And hmm, based on the data struggling, the device S3 is failing its statuses. Hmm. Violating rules.

It seems to be concern. This is, but then it actually gives me, you know, the legitimate answer there. That was just for fun.

But that question could be, could you please make, um, could you please send an email report or, or generate a, an executive summary about the P one that I'm experiencing? You can add context to the co-pilot, right? Anywhere, any language.

ai. There's no login, there's no signup, there's no fees. Our ceo, EO Conan wanted to give back to the community and really was interested in the ability to talk to P caps.

So just to show you one demo of this very quickly, um, this is what the page looks like. There are sample packets here. If you don't want to use your own packet, just download one of our packets.

We're gonna browse and grab this capture and say, what can you tell me about this packet? And we're gonna hit send. Now what you're gonna notice is it's fetching some geolocation and fetching some network intelligence for public IP addresses.

This originally was a rag system where the PCAP becomes JSON and we put it in a vector store and we do similarity searches with the ai. But in the age of agents, we realized if we detected a public ip, we could invoke agents to go out and do things like curl and paying an Ennis lookup. So as you can see, not do I only do I get the port number source and destination IP and different information.

It tells me where the packet is. This is from San Jose and it's from WebEx. And here are the prefixes with the a SN number.

They belong to the best packet analyzing tool out there today. I will be speaking at Wireshark, by the way, to highlight this tool. And this tool is getting a bit of an overhaul next week where we're gonna incorporate model context protocol, MCP And John.

Uh, there are, there can be sensitive data in pcap. Yes. How do you handle that?

What do you, do you recommend any scrub your ips before feeding it into Great point. So it is public, so I would, I would recommend using lab P caps or obfuscated P caps, like you said, there could be sensitive data in there. Now, just some things that we're doing.

The A cap in JSON, the data payload is typically just a bunch of hex garbage. Mm-hmm. We'd remove that.

We're not looking at the data, we're looking at the headers more or less. However, if you don't want your public IP to be scrubbed or checked or anything, be very careful here. The other thing is, this is not a replacement for wire shark.

It's an augmentation to wire shark. It's garbage in, garbage out. If I just do a wire shark capture on traffic randomly and send it to this thing, your mileage will vary.

But if I put some filters in my wire shark and, and do a highly curated P cap, then it works much better. Okay. John, quick question for you.

Yeah. Getter. Um, are there any plans to make this locally accessible In the tool?

Yeah. Or, or let for the public? Yeah, for the public to be able to download Yeah.

Copilot, run it maybe locally on their Mac or if they have GPUs to run it in-house to, to, you know, avoid those sorts of issues with, with private data. So we're, we're in, I'm in discussions of open sourcing and it likely will become an open source tool that people can use on-prem in their own enterprise at home. Okay.

Uh, it, it will be coming soon. Okay, great. Yeah, great question.

The other thing is, if you own selector, we're also looking at integrating it in the selector platform. So if you're a selector customer through the co-pilot, you might be able to send P caps and get analysis from the network language model. It's, it's all in the roadmap.

Okay. Cool.