Name: TextQL
Uploaded: 2024-12-09T09:04:12.388Z
Duration: 2 min 33 s
Description: An AI-powered data analytics platform enabling users to interact with business data using natural language, automating workflows and generating insights.

textql

An AI-powered data analytics platform enabling users to interact with business data using natural language, automating workflows and generating insights.

TextQL

Hi I'm I'm Mark. I'm the co-founder and CTO of TextQL. Here with me is Kelen who is um the product manager and lead analytics engineer at the uh at the NBA. So um in this talk I want to talk about first like who we are TextQL, how do we view AI and how does the NBA view deploying Geni in production for data analysis. 

And this is a sponsor talk. They are a customer of ours but they are not a normal customer. In fact, uh the NBA has known us since we were only two employees and just in the first couple months of our journey. So a lot of the product features will show um a lot of a lot of our thinking around how to deploy Geni. How we interact with the data ecosystem is actually inspired by the NBA's specific needs and setup. Hopefully lots of the problems that the NBA faces rhyme with you all's problems. So hopefully this is useful to everyone. 

So yeah let's get started. So to um to start off uh Geni for automating developer and analyst workflows is probably one of the most hyped use cases out there. Um Amazon CEO Andy Jassy says he saved 4,500 developer hours with Geni and there are tons of different startups raising a ton of funding around this use case. 

So if we can make this, if this is so promising for software developers, where are the agents for data analysis? When I talk to people, lots of people say oh data analysis is just some SQL in Python. It must be so much easier to automate. But in practice we found this is actually the opposite. 

And our theory for why that is is how spread out the data ecosystem is for an analyst's day task. Think about how many different tools you use every single day in your work as a data analyst. Whereas a conventional software engineer might use a couple tools, maybe VS Code, some other IDE, a terminal, interact with the application running on their computer, the job for a data analyst involves way more tools and a much more complicated setup. 

They might have to check Tableau dashboards, go back and look into the warehouse to check the health of a table. They might have to go look in their data. 

Please share the AI Agent Store with friends.

Documentation look at wherever their source for for governance is. When you frame it in terms of how many places do I have to touch to answer a single question, the job of data analysis becomes much harder. And that's what we're trying to solve with TextQL. Our goal is that, well, we believe a true autopilot can first answer any kind of question from broad to specific. 

So what do we do next? That's very vague. Or how do we define X? So not just getting data but also getting definitions. Or what's the definition between X and U? Can you build a model for this? And how would this look if I change some of my assumptions? 

The second pillar is a true autopilot should be able to use any tool or at least as many as possible. Instead of being isolated to say a single warehouse or a single BI tool, a true analyst autopilot should be able to use anything that a human analyst would use. And lastly, automatically and with little setup, the tool should not require a ton of configuration. It should not be a walled garden. Instead, it should take as much as it can from what you have existing and use it just as, again, a real analyst would look at existing docs, existing procedures, and take advantage of those. 

And that's what we built with TextQL. By connecting to every single tool, or at least as many as we can, our agent can search across your data documentation, your data governance, any BI tools you have, Tableau, Power BI dashboards, any semantic layers you have, and use that to answer questions. 

So we can answer how do you define CAC based on definitions to ulation? Is there a dashboard in Tableau that forecasts it? And could you build a model that predicts how will this change if demand went up by 14%? 

And the way we do this is through something we call deontology, which is the engine that powers the search. So your data warehouse, your documentation, your BI, we index all of it through ontology and let our agent analyst Anna search through it just as a real analyst would. 

And just an example, I'm going to pass it off to Kila now. This is how.

Anna would use DBT. So if you have a semantic layer defined, this is one of our newer features. We'd be able to use this to query metrics. Um and in addition, using using using lineage to find joins using the catalog to understand the context behind columns much as a human would in their day-to-day workflow. 

So now can pass it off to Kin and talk about uh the NBA and some of our work with them. Cool thanks Mark. I'll come over here because I can't quite see the thing. Uh yeah, so to set the foundation here, I think a lot of what Mark mentioned, um a lot of what I think about with this talk is that all these like automatic AI agents and whatnot, um often come with this expectation of the way you'll need to model things, the way you'll need to set up set things up. 

And the reality is that's just like not how people have done work before, right? Especially at a company like the NBA. Um we use three or four different places that people document things. If anyone came to my talk yesterday, I talked about the last two years just getting 35 people to start using DBT, right? 

So um when we think about trying to set up an AI chatbot experience, of course we want to move to someplace where we have everything standardized in a semantic layer. Transformations built in DBT and whatnot, documentation all in the exact same place. But to start, we don't want it to be a blocker for us that we can't just like use something that can live as a thin layer almost of like discoverability. 

If any if nothing else over our all of our information, over all the things that we already have built. Um added context like like Mark mentioned, uh text ql was part of our Launchpad program this year. I actually met TextQL last April, I think, um when they were like kind of figuring out what they wanted to do in the market. 

It was just Mark and Ethan and uh we were talking about a DBT documentation co-pilot at the time. And at the time I was like I don't have enough models in DBT for it for me to need AI to document it. Like I can just do it by hand right now. A few months later they came back around with more of what they were thinking.

Through and and we brought them into Launchpad which is our program every year that we run um we run a few different tracks kind of figuring out a few different vertical that the MBA is interested in like investing in Innovation and we pick a few companies to um you know do an extended POC with and and take an equity stake in and um you know try and have something bigger than just like a customer client relationship um and really you know spend that time giving deeper product feedback. 

Yeah just to chime in we found the Launchpad program to be extremely helpful in building out our product road map our vision and how do we think about the space being able to spend our early days as a startup embedded inside a real data organization especially one as strong as the NBAs was it was completely invaluable to us and yeah absolutely. 

And if anyone in the audience is interested in learning more about Launchpad obviously please come up and ask more questions after the talk um so before and after I think a lot of people have seen slides like this right um one thing I want to I want to mention is you know the two days to two weeks of of an analysis the this is I think the reality of a lot of net new questions and we were able to bring down some of those to two minutes. 

I think the next phase is US wanting to be able to have text ql bring down the two days to two weeks for all net new questions what built out in this in this um process was 113 different questions a sort of Benchmark of how things we had we knew the answer to but we wanted to kind of see if TextQL would be able to without too much of our like leading the witness um answer those. 

Um it was really impressive I mean especially with the fact that we don't have a semantic layer um the accuracy was um very strong I would I think it's 100% coverage not 100% accuracy I'm not sure if that was an updated slide or not but 113 questions and was able to answer them. 

One of the most important things about our installment with TextQL I'm not sure if it's on another slide but we weren't we didn't want it to be something that automatically.

Went into text to SQL generation that was the like one of three pillars of reaction to a question to us. We have built so much work in Tableau. We had built so many things in Confluence. I wanted something that when I ask a question if I don't have to try and do the like moonshot of writing text to SQL just go get the answer from the dashboard and pull that up and manipulate it how you want. 

That was one of the coolest things I think about working with TQL is explaining this thing where it's like yes everyone wants to nail the text to SQL thing but we've already done a lot of the work. I want to be able to just like help someone in the senior leadership team like find the right answer. Sometimes it's hard to get them to you know open Tableau or open our NBA reports page. If we could just have something that they could just ask a question even if it's just to get them to the right place. 

Um, and that was one of the I think my favorite parts of working with Texas QL was how open they were. I think I mentioned like we have a ton of our product analytics documentation like what events mean in Confluence and there was no way we were moving all of that over to Eli or to DBT. I asked if they could set up a Confluence connector and like in two days I think it was like a scrappily built POC for a Confluence connector which we were able to leverage. 

Um, yeah so like I said key lessons here we wanted it to be something that felt easy the interactive sort of UI. Um, but the most important thing to me was that while we are working towards some like beautiful world where everything can be answered in text to SQL on top of a DBT semantic layer I didn't want that to be a blocker to getting started with AI. Right with getting answers to something. 

And so build being able to build something that was that thin layer that didn't have too much configurability even if we want to like move to that direction it's something we could actually iterate on instead of being like we can't use this until we have a beautiful modeling layer until we have a perfect semantic. 

Please share the AI Agent Store with friends.

Layer on top of that um future plans uh yeah I mean it's it's rolling this out in more places at the MBA. Um so far we've kind of kept it pretty tight and that's mainly because it's a lot hard a lot easier to like lose people's trust. Um and so we wanted to kind of keep the questions that can be answered and the Tableau dashboards and the Confluence spaces to a tighter group of things that we knew were going to be right. Um but you know we want to keep pushing that envelope and and figuring out where to go next. 

Um yeah and I think that's the end do you do you have any more? Um no I mean that's that's about it. So one more thing might be the DBT semantic layer. Yeah and how how um you guys are moving to that and hopefully when you can when you roll that out we can be right there and start being able to answer your question exactly. 

Yeah I mean I think that's that's exactly the idea is that as we expand the usefulness of TextQL we're also improving like the underlying uh Foundation that it's working off of all the time. Um and that ability to kind of iterate and and and work back and forth I think is an exciting next part of the journey. 

Yeah I agree um I think we have time for questions probably have plenty of time for questions. Yeah have like uh yeah have a lot of time for questions so happy. I think we have a mic person. Yeah Alexis is running around with a mic if anybody has questions. 

Um yeah wait for the mic. I think we'll just need to um this is really interesting I want to know like what is the end user how do they interact with TextQL like is it through a slackbot is a UI or um so right now the the main way we've been having people use TQL is through the is through the web app. 

So web UI um we have worked on slack we have worked on like a teams integration that's sort of rolling out but we see it as sort of a little bit worse than than the web app because what you can provide in slack just in terms of like what charts can you provide like can you give any like detailed tables and so on that's a lot lot harder to achieve in slack something we're working on. 

So so right now most

Of it's done through the through the web app just like we have you know so many different ways that people do their work to set up what this is like built on foundationally. There are lots of different personas in terms of like how they ask questions and so you want multiple different ways to interact. 

And while there is the web UI, I think like the chat bot is perfect for people who don't want to deep dive one thing for further. They just want to get the number as soon as possible. Um and so it's exciting to be able to build out a few different ways to interact with this as an interface. 

How you mentioned, oh sorry, you mentioned personas. I was curious how adoption has looked like across the different personas like say a senior exec who doesn't have a data background to like the analyst. Um so I think for the past couple months, and correct me if I'm wrong, Kein, the roll out and the main personas have been sort of technical to semi-technical people. 

So I'd say the upper limit being people who can like read a Tableau dashboard fluently but not necessarily someone who could like make that Tableau dashboard from scratch. Um the senior leadership roll out is something that we've been taking a lot slower. Yeah, I would say like we want to make sure that it is like there's nothing that can go wrong before you know someone in senior leadership is asking a question. 

They’re still happy to rely on analysts and there are people that are focused on that and that's fine for now. This that is like the beautiful world we want to get to but um the layers below that, the people kind of in the hub and spoke of like data that want to be able to look into these things and get questions quickly. Um those are the people that we're most focused on from like the personas you mentioned. 

Hey, um are you able to answer questions without the semantic layer set up? And uh if so, were you able to measure uh the difference in accuracy between um you know answering questions without the semantic layer or with? Uh they actually don't have the DBT semantic layer set up right now so we. 

Please share the AI Agent Store with friends.

Are doing everything without at the moment. It's just something we're looking forward to once they move on to it. And that's partially why like we we build a lot of metrics cubes at the NBA without like short of the semantic layer. 

So um we wanted something that could option between. I know this is in a tableau dashboard. I know this is an a metrics Cube. If neither then I'm going to go try and like build it from scratch essentially. 

Um but I think there's a lot of of like academic papers out there that prove out like how much of a difference a semantic layer or like a yeah an ontology that is database specific improves accuracy. So like we know that that is going to make a difference. 

Um we just haven't set it up yet great. And with any other customers did TextQL get some kind of measure that confirms what this intuition? 

No unfortunately so we don't have any customers who use the DBT semantic layer. Again it's just something we're excited to move to when it becomes the case. 

All right cool cool thank you guys. 

Share the AI Agent Store with friends.

Note, that there may be slight differences in text and video. Watch video for the original content.

TextQL Video