Chatbots like ChatGPT and Copilot are remarkably skilled and articulate when it comes to subjects they are trained on. But they frequently struggle with domains outside their scope of learning.
One area that has stood out in that regard is network monitoring. The human-level fluency that open-source AI tools are loved for tend to fall apart when faced with questions about domain-specific information. For example, if you ask ChatGPT questions about SNMP data or syslog, it will not be able to provide helpful answers.
The reason is, it was never designed to do that. Yet it does not seem to dampen the desire of users to throw problems at it that it doesn’t know how to solve.
“There is a natural tendency to assume that we can take our infrastructure data and just connect up to ChatGPT and get a conversational interface,” says Nitin Kumar, co-founder and CTO of Selector AI, at the recent Networking Field Day event.
It’s a reasonable argument, except, it has a major flaw. The cloud platforms where these services run, are vastly different from the private infrastructures where the data resides. Connecting the two requires a medium.
“Not only is the distance a problem, even the semantics of that information that sits in these systems is incomplete.”
A Skill Problem
Hybrid cloud deployments have been on the rise in the past few years, a trend that has made networking woefully complex. Think of a retail network that connects hundreds of stores scattered across the globe. These stores connect to a cloud network where all their applications run.
If a cash register at one of the stores fails to connect up to the payment application in the cloud, it can grind things to a halt in a moment.
The network is the usual suspect of course, but to an ordinary IT guy, that tells nothing of the real root cause.
Increasingly, it is getting trickier for people without specialized knowledge to analyze network data. Data silos demand specific types of domain expertise, and only a small group of people can make sense of that information.
“People who are experts in these domains have access to that data. They understand what’s going on, but any other person is not able to figure out where the problem is,” Kumar points out.
Serving Data to a Wider Group of People
Data democratization can change that. “The ability to access and understand any kind of data across all of these different domains,” has been the key driver for the Selector Platform.
Selector provides two key capabilities – a conversational interface like ChatGPT where any IT persona can make natural language queries and get answers instantly, skill no bar, and an alert layer that puts together data points about errors and failures into comprehensive alerts. Together, Selector Copilot and Smart Alerting bring intelligible and consumable data to the fingertips of users across the board.
“You are no longer hostage to experts or vendors or a particular domain,” Kumar emphasizes.
Investigating Incidents with Conversation
In a demo, he showcased the two interfaces in different scenarios.
Selector Copilot welcomes users with a clean portal and a search bar for typing queries. One can ask any question in English, and it fetches results in seconds.
Kumar gave a couple examples to elaborate. In a scenario where retail stores in a particular region are encountering issues, Copilot can surface contexts – what could be causing the issue, what internet services are down in the area, and what applications are likely to get impacted.
Copilot puts together simple topology maps, color-coded visualizations and incident summaries to make complex investigations super-simple. “The manner in which this information is presented is very human-like,” he says.
The interface also offers historical perspectives by showing users before and after pictures.
Follow-up questions can be asked as one proceeds with the analysis. All outputs can be saved and organized in a custom canvas on the left-hand side of the UI for troubleshooting sessions.
Information found by interrogating Copilot can also be found pre-bundled under alerts at the alert layer. The biggest selling point of Smart Alerting is alert consolidation. Instead of pushing out “onesies and twosies – one event, one alert” that is distracting and counter-productive, Smart Alerting encapsulates information of tens of alerts into one notification for wider observability and low noise. The ready analytics help users dig deeper into an issue and track down the origin in the shortest time possible.
The alerts also have various workflows. “You can create an incident around it. You can create a PagerDuty incident or a ServiceNow incident, and teams downstream can then take care of that,” Kumar adds.
The Framework Below
Under the hood, a software layer acts as the bridge connecting the on-prem infrastructure with the cloud services.
Network data is collected from far and wide in the network through collectors deployed across regions.
“Collection of the data itself is a big problem,” Kumar notes. “You need to be able to know the devices that you want to go to. There’s the crawling aspect, mechanisms that discover all the devices in the subnet or a particular region, do the last mile connection, and start SNMP polling.”
Adding to the problem, this data has no one format. “There’s no single format that’s going to be applicable to all infrastructures, and we need to embrace the diversity, not run away from it,” Kumar exclaims.
Selector has a built-in layer that transforms all the different schemas into one universal format. This is a homegrown data compiler called Data Hypervisor.
“Just like compute hypervisors abstracted details of compute and storage from applications, the Data Hypervisor hides the formats of the underlying data and presents it in a uniform way.”
The data is saved in data stores which serve as a single storage layer for this mixed corpus. It is subsequently processed, and analyzed by the Selector LLM deployed close to on-prem.
Results are published via a variety of third-party interfaces that include, but not limited to Slack and Teams.
Don’t miss the Selector Platform demo and other deep-dive presentations by Selector from the recent Networking Field Day event.
