Video: Simplify Your Path from Question to Knowledge | Duration: 2884s | Summary: Simplify Your Path from Question to Knowledge | Chapters: Welcome and Introduction (4.56s), Housekeeping and Q&A (130.565s), Team Introductions (161.96501s), Unified Data Workspace (214.05501s), Contract Analysis Pipeline (760.64s), Data Catalog Search (1522.97s), Joining and Visualizing Data (1630.0901s), Automating Data Insights (1859.98s), Scheduling and Governance (2045.155s), Conclusion and Thanks (2809.265s)
Transcript for "Simplify Your Path from Question to Knowledge": Alright. Alright. Hello, everyone. Welcome to another Dataiku live session. While we're here, before we get started into the actual content, I always love to find out where people are watching from, where people are, you know, listening from. So if you see that little chat window, it should be on the right hand side of your screen. Just go ahead and type, where you are watching from. So we have a few over in Europe. We have Ireland and UK, a few different places in The UK being represented. We have some Canada, some Belgium, some Germany. We do have Michigan and Texas, so The United States is represented here. So that's awesome. We have some Philadelphia. I'm not too far from Philly, maybe about an hour and a half to two hours away. So, yeah, it's good to see everybody coming in. Please continue to comment where you're joining from and watching from. It's gonna be a very fun webinar. And today, for this Dataiku live webinar, we'll be talking about the modern analyst path from question to insight. So we know that things have changed, exponentially over the past few years in the world of data and AI, and it's really changed how everyone has gone about doing their work from data scientists to data engineers to the data analysts. So today, we're gonna be looking at the work of a data analyst and really how it's changed, how AI has enabled acceleration, how AI has enabled, quicker and more confident insights, but also in the ways that, we need to think about how we use AI before we just, you know, start opening up ChatGPT or Claude or Gemini and and start typing away and and trying to find insights that way. So as I said before, continue to leave, where you're from in the chat. We always love to see it. Few housekeeping items before we introduce ourselves. There is also a q and a tab. You'll see that right there in the chat window as well. You can click over there. If you have any questions, please leave them. We'll be answering all questions during the session as well as, spending a little bit of time after the session answering any questions. So do not be afraid to ask any questions. There there's no such thing as a, dumb or wrong question, so please ask away. So in terms of introductions, my name is Chad Covin. I am a technical excellence manager here at Dataiku. I've been with Dataiku for about three and a half years now, and technical excellence here at Dataiku is essentially all of our technical customer facing teams. I'm part of the team that helps enable, and make them as confident and as efficient as possible. And I'm joined today with Chris. And, Chris, I'll hand it over to you for introductions. Hey, everybody. Chris Helmus, principal sales engineer based out of Denver, Colorado. Been with Dataiku about four and a half years and have a pretty long history of helping folks, get access to data, democratize what they can do with it, do a lot more self-service. So excited to talk to you about this modern analyst view. Yeah. Appreciate that, Chris. So, we will be walking you through that that modern analyst view as he said, and, I always think it's important to to really start with setting the scene. Right? We can't really understand what's the solution if we don't figure out what's happening today and what some of the challenges might be. So let's let's picture this. Right? It's it's Monday morning. You know, in this case, technically, it's Tuesday morning, but it's Monday morning, and marketing is asking us a simple question. Which suppliers are driving the most customer complaints, and are there any patterns? And you know that the answer is somewhere in your data, but first, you need to pull transactions from the data warehouse, feedback from Salesforce, supplier info from the ERP, and unstructured text from support tickets. Okay. That's that's easy enough. And if you are having any connection issues, I would say, try leaving and rejoining. So you open your SQL client. Right? And then you open your Python notebook, and you export a CSV, and then another one, and three hours later, you finally have clean data. Maybe. Right? Maybe you're second guessing yourself. Did you miss something? Should you check that other dashboard that another team made that shows slightly different numbers? And this is the reality for a ton of analysts. Many analysts spend nearly half their week just hunting and prepping data, and when do they actually get to do real analysis? That usually gets squeezed into whatever time is left. So why does this keep happening? Why does the life of an analyst have to be this way? And that's really because we've been stuck in this best in class trap. So each department pick their own tools. Marketing has their BI platform. Finance built something in SaaS. Operations uses Qlik, and sales lives in its own data warehouse. And IT is trying to make make each accessible and trying to connect each of these systems together. Right? And for you, you, the analyst, or you, the analytics, leader, you're sitting right here. You're usually stuck right in the middle trying to manually stitch all these different things together every time somebody ask a question. And those dotted lines, that's your daily reality. You're copying data between systems. You're reconciling numbers that don't necessarily match. You're spending about 40% of your time on integration instead of analysis, and stakeholders don't trust the dashboards because each system is showing slightly different results. So, ultimately, people are exporting to Excel or Google Sheets, and they're making decisions on their gut feelings instead of true data driven, analysis. And, really, not only is it it's it's frustrating, right, to start, but, really, it's expensive. And half of these companies are abandoning projects that simply, are just too difficult to do because they can't connect the data. So now you might be thinking, hey. This is this is where AI comes and fixes everything. Right? And it is. It's definitely helping. Right? So 97% of analysts are experimenting with AI tools, and some are saving, eight plus hours a week using AI to write SQL, to generate transformations, to automate analysis, etcetera. And when it works, it feels like magic. But the reality is AI can hallucinate. So it can give you answers that look right but aren't so. So you need to really be able to understand how it got to that answer. You also have governance concerns about what data your AI can access. Integration can sometimes be complex. These tools don't magically just plug into your existing tech stack. Trust is still an issue, so you need to make sure that end users and stakeholders actually trust your outputs. So you're still spending a lot of time validating everything manually. So the promise is real. The time savings are real, but most teams are still stuck in pilot mode. They're not getting to production because AI tools are just another disconnected layer on top of an already fragmented stack. So what would it look like to have everything in one place, right, everything where you and your AI and your data, and all of that and your space to analyze that data are unified? So what exactly does that look like when you bring everything together? It really comes down to three core capabilities that work together. So for one, your agents will be giving you actual answers. So not just chatbots that are pointing you to that to documentation, but we're talking about AI agents that can connect directly to your enterprise data and give you real insights. And when you need something specific, like, for instance, that supplier example that we were talking about, you know, maybe you need some sort of supplier analysis. You can build your own agent in minutes with already trusted connections to your data regardless of where that data lives. For two, you can find what you need when you need it. So this means easy templates, so you're not really starting from scratch every time. It means accessing golden datasets that already have the context and quality checks that are built in, and the data your organization actually trusts. So now you're not hunting across five different systems or second guessing whether or not you have the right version of that dataset. Now everyone is utilizing one single source of truth. And then third, you're building with confidence. And we can think of confidence in two ways. One is speed, and the other is control. So with speed, you have AI assistants that are there every step of the way during your workflow. They're helping you prepare data. They're helping you write these transformations, build visualizations. But the big key is you stay in control. You can trust the AI to accelerate your work, but you validate the outputs. So when these three things come together in one workspace, that's how we actually go from question to insights fast. And unsurprisingly, that unified workspace is Dataiku. It's exactly what we described. So here's what it means for you as an analyst. There's no more jumping through tools, data prep analysis, visualizations, AI agents. It all happens in one place. So that Monday morning supplier question, you can now go from that raw data to those insights without leaving the platform, and you can get there fast. And then for those who are team leaders, now everyone is speaking the same language. Your analysts are able to collaborate with data engineers and data scientists through working in the same environment, and they can share work, they can reuse templates, and build on each other's insights instead of recreating the wheel with different tools. And then last but not least, the governance piece is very critical. Right? You get that central tower of control and trust without slowing your team down. And because Dataiku connects to your existing infrastructure, whether your cloud platforms, your data sources, and the I, the AI tools that you're already using, you're not asking for any sort of massive rip and replace. You're just simply connecting and using, what you already have. So that's that's great. Let's make this real though. And Chris is going to walk us through the role of a supply chain analyst, trying to understand our our world of suppliers, delayed shipment shipments and risk. So let's see how this works when everything from tools to systems and processes are connected. So, Chris, go ahead and and you can walk us through our demo. Excellent. Well, thank you, Chad, for setting the stage as always. What my goal is going to be is to take you through a little bit of that day in the life and bring this all to life. Alright. Great. So, yeah, I will take on the role of that supply chain analyst, and we're gonna be starting in an area of Dataiku call AgentHub. AgentHub is my central place to start digging in to questions. When I say start digging into questions, I mean to start interacting with agents that my organization has made available to me so I can start digging just through natural language like you'd hope, but in a way that's connected when we talk about real answers to actual data, actual applications, the enterprise assets that I need. So for example, I might just start and be like, what's available to me, you know, based on me, Chris, the supply chain analyst. And I see I've got supplier management agent, some things related to manufacturing, and a sort of Dataiku specific agent, which we potentially work with. And when I look at this, you know, right away with these supply chain questions, you probably guessed it, I could talk to all three, but I could also just say, you know what? Let's focus in a little bit on that particular agent related to supply chain. And so I could start asking, you know, the things that are relevant to me as an analyst on this Monday. So something like, you know, can you show me delay rates for all suppliers? So I can kinda get an idea of what's going on. And I think the key is as this starts to run and work, you can start to see the chain of thought reasoning as it goes. And that's something I'm gonna be able to look back on later so I can understand, like Chad said, that trust out of the equation. You know, what it is being called, how is it thinking, what tools is it working through so that when it produces a result, I actually know something, kind of understand what I'm working with. So now I've got some delay rates by suppliers. Wonderful. I can see, you know, some doing great, some less great, just like you sort of expect. And maybe if I wanted to dive in and understand, well, maybe it's, you know, something like tariffs are being involved, and I wanna know how much we're spending with these folks so I can put in something about suppliers impacted by tariffs and total transaction costs. And while that's running, you can see again, I can give direct feedback. This goes back to my enterprise teams that created the agents, letting them understand whether the answers make sense to me so they can incorporate more context, incorporate fine tuning other methods, new datasets to make this helpful. And I can also see, for example, again, that's a that risk letting me understand how the agent did what it did. Same idea here. Kind of seeing that SpeedLine, in particular, a lot of transaction costs, especially those relating to tariffs. So, you know, when I think about this, the next thing that pops up in my head is I really need to understand some of the contracts for these suppliers. And so I might ask the agent, do you have any contracts? And it's gonna look and say, like, hey. You know, I don't have them. And so this is a really interesting part of this hub when I'm trying to answer my own questions before I get into the nitty gritty is just because there's a bunch of enterprise agents doesn't mean it's gonna be able to answer every question. If they could, you know, I would need to be here asking the questions myself, right, and trying to find these insights. And so Agent Hub with Dataiku provides a really unique experience where I can actually create my own agents down here in the bottom left. So here's, like, my enterprise agents. I can create my own. I've started one up, so we can just take a look. But, really, what this lets me do is it lets me define what my agent is gonna do either from a bunch of ready made prompts. So this was kind of my own I put in that I want to leverage contract documents to understand grace periods. You know, grace periods around, you know, if someone was five days late, but in the contract it says they have a week grace period, then that means they're not gonna incur penalties, for example. I can pick models that are approved to me for my organization. So it could be in this case, I'm using OpenAI, but it could be Bedrock. It could be, Google Gemini. It could be any open source model the organization is using data. I could connect and leverage it based on, again, you know, who I am and, the cost controls in place and what I can call. And you can see that I've actually uploaded, just as an example, one contract from one of our suppliers called AutoCore. And I've actually used Dataiku to autofill a lot of these descriptions and example queries to get started. And then to kind of where things get a little agentic and interesting, right, is instead of just giving it documents, I can also add in approved tools. In this case, I have a pretty simple Google search and, like, sending messages to myself, but this can be any tool the organization builds that I can select and I'm approved to leverage, you know, to inter to interact with maybe Salesforce or something like a ticketing system or to, talk to a BI tool, let's say. And what's great is right over here, I can test this. I can ask for my new little agent to analyze the contract and tell me the grace period, and it can tell me, hey. This contract has a five day grace period, and I can actually take a look and it can show me where it pulled that text from the contract and why it's saying that on that trust side. And then I can ask it something like, you know, can you search for news or can you send me a message of this conversation? And you can see, it actually talking to that send message tool that now lands in my inbox, let's say, or in my team's message. So I have these enterprise agents starting to ask questions, those quick questions, those ad hoc questions I get asked all the time on a Monday morning. If I need to go a step farther, I have the ability to make these and not just make my agents, but actually publish them and share them to folks on my team so that they can leverage it, understand that they like it. We don't have to be making a thousand different agents through this sort of analysis. Now when I think about this, this is great. I'm getting some answers. I'm finding things, and this seems useful. But this is kind of sparking the next question for me as an analyst where it's great I can put a couple of documents in here and talk to it and find some answers, but I really need to go a little farther. I need to actually analyze at scale where all our contracts live, extract information, and try to figure out, you know, how we're doing with these suppliers in comparison to the grace periods in their contracts. And so I can actually start a new conversation, and maybe I'll use my find my template agent this time and say something like, you know, help me find a template or a display chain. I'm reading documents. Right? I'm I need to kinda know where the data's at. I also need to know you know, I'm not sure exactly how to make this sort of pipeline that's going to be able to analyze these contracts at scale. And this agent specifically pointed at my actual data, like, coincidence, so it can actually find these templates that I wanna use. And so if I hop into this template, I can actually see as an area to pulling from contracts, extracting images, and extracting data out of those contracts. And this actually looks exactly like what I'm looking for, something very similar. And in Dataiku, it's very easy for me to select what I want here, and I can actually just say, you know what? I wanna grab this. Just like I might grab an Excel template or, a template from Git, I can actually say I wanna copy this over into a project I already have in in Dataiku where I can start using this. So I'm gonna bring it into my target project where I'm investigating delay, let's say, and I'll copy it in. And this is how easy it is in Dataiku. You know, I'm searching. I'm finding things that are useful. I now have a template that I can understand in review that's visual and understandable. And I can actually now say this template that's doing something, this was the kind of template contracts in that example. But if I wanna plug in my own, it's as simple as coming to this step and plugging in my own contracts for my suppliers. And now I've got this map. And while we execute this little part of the of the template, let's talk about what we're seeing here. So I've got my contracts pointed at my enterprise contract store, you know, maybe in a cloud storage bucket like s three, maybe in an application, maybe in SharePoint. It really depends. We're then taking those contracts. We're processing them so we can get the images ready for an LLM. And I've actually got some LLM processing going on, which we'll investigate, and then some kind of normal regular data prep processing to really get exactly the data I need to continue working. And so if I take a look at what's going on here, I can always, in Dataiku, just double click to get all the detail I need. And if I wanna actually see, you know, does this prompt make sense? Is this does the prompt that's in the template really work for me? I can choose this option in Dataiku called prompt studio. So right in the same place, I've got my enterprise level data. I can now start this prompt. So I can now actually test out, like, a simple prompt, like extract grace period, and I could say, let's go look at that across maybe two of my contracts. Wonderful. And I can see, okay, it kind of gets there, but it's a bunch of other stuff. I don't need, like, that much stuff. Maybe I can try a more specific prompt, and I'm just kind of accelerating things a little bit so we can go, where I'm much more exact than what I want. I want the grace period in the supplier. I have some critical instructions, and we can run that. And so I get this kind of iterative thing where I can look at the contract, understand what's going on. I have my prompt, and then I have my result. Perfect. Pulling up the grace period, pulling up the supplier. And so I think we had that agent that I kinda just uploaded a document, and it was able to pull out a certain grace period. But now I have an actual pipeline that's taking my enterprise contracts and feeding it through an LLM assisted way to pull out really nice structured information. And so I went from a bunch of contracts to now some structured information that JSON output of that LLM that's telling me, you know, what the grace period is. But that's not this kind of weird JSON thing is not really where I wanna stop. Right? I wanna process this. I wanna get it into a nice structured format so that we can apply it in our use case. And this is where Dataiku, besides just the AI to help you answer quick questions and besides the AI just to, help you find your templates, it's also the fact that when I go into a repeatable visual thing in Dataiku, and you can see here I have some steps that are processing my data. So I get down to a nice sort of path of the file. It's kinda messy. We'll take a look at that. I get the grace period extracted, and I get the supplier right out. And I can actually see exactly how Dataiku did that by just clicking the eyeball. You can see how it actually processes out that JSON into the two columns. But let's say I'm not that familiar with some of the things I wanna do next, and I actually want Dataiku to help me. And so if I wanted to maybe let's say, maybe let's just extract out a little bit of this data. Like, maybe I wanna do something like extract the file name ending in dot PDF from the path. And this is because I hate regex. I hate data processing. I just want Dataiku to take my text and turn it into a visual thing that I can quickly validate. And so now it makes the step. I actually see that it has the prompt attached, but it becomes a step. When I click the eyeball, I see the results. Now I get just that clean file name. And I think this is a really key part of working with AI assistants and Dataiku and how you can move really quickly is that you're always going to have that validation every step along the way. You're not getting, giant blocks of code that are hard to work with, though you can, of course, have Dataiku help you generate code as well. You're going to get this visual representation that makes it really easy to understand the results and say go ahead, execute. So what have we done so far? We have found a good template. We plugged in our enterprise level contracts. We've kinda tested out our prompt that makes sense to kind of pull out grace period supplier from these visual contracts, and we've even used a little assistant to help us clean that results up and get it into a nice table. But, really, all we have so far is we have our suppliers and their grace periods. My goal with this was to apply that to our supplier information or transactions to figure out which suppliers are not just delayed like we saw on the initial agent, but delayed over their grace period. So what I really need to find is some more structured enterprise level data. And, of course, in Dataiku, I could just write a SQL query. I could search through my enterprise system, but I wanna see if there's some things already set out there for me, some things Dataiku can do to help me find the right data. And so when I go up here to add some new data, instead of just connecting to a database, I'm gonna use our data catalog where I can search through things like popular datasets, datasets that are used in a lot of projects that I have access to. That could be interesting. But I can also look at very curated data collections. So in this case, we actually have a supply chain data collection. And this all kinds of data could be in here. It could be data from Snowflake, data from Databricks, data from Redshift, data from wherever. It's just pointers at that enterprise data. But now to help me understand if this is the right data for me, I have a lot of information. I have some data quality checks, green. I know we have full metadata completeness, so things like the description of what this is about and actual descriptions for the columns so I can make sure I understand, you know, what does PO confirmed quantity means. And I can even do things like get a preview of the data or understand where it came from upstream. Kind of see where that project is coming from, what was used to create it, maybe even look read only at that upstream, template and maybe even contact the person who created that project. So I'm feeling pretty good that this is gonna have information I wanna use, and all I need to do is click use. And it is now inside my project, a pointer at it, and I can easily start scanning across or saying something like, let's see if our supplier names make sense. Same suppliers we were seeing before. Wonderful. And I know I have these suppliers in this dataset. I know I have similar suppliers in this dataset, but let's say I'm not actually very familiar with how to bring data together in Dataiku. And so one of the things I love is that Dataiku doesn't just let you make steps inside of a recipe to manipulate your data. If I'm not really sure, especially let's say, maybe this is my first or second Dataiku project. I'm sure many of you out there have, are still early in your journey or considering Dataiku. I can kind of just pick. Here's my data on suppliers, my transactions, and how delayed they are, And here is my key information I extracted from my contract. And I can just actually ask Dataiku. I'm not really sure what to do. I just want you to generate something. And maybe I'll say, you know, join these datasets across supplier name. I don't even remember exactly what I called it in this dataset, but we'll see if Dataiku can sort of figure this out. And what I really like about this process is two things. One, one is that Dataiku is going to give me reasoning about why it's doing what it's doing. So it kind of reformulates into exactly what the datasets are in the flow, and then it does some reasoning to explain what it's doing. That's one. I like that, and I can iterate right here the prompt. But the second thing is when I hit create recipe, again, just like the prepare, in this join, I have the map of what happened. And if I don't know if it's wire name and supplier are right, I can just click the equals, get a little preview, and see, yep, like, auto supply auto supply. This looks like the right column. I'm ready to go. And I'll just change it to an inner join to make sure I'm getting you know, just matches. I just wanna have where I have the grace period present in my contracts through this analysis. And just like everything else in Dataiku, I get a very visual option where I can say, you know, let me keep all these columns, and I really just need grace period here. And all of these check marks I'm making, the type of join, what columns I want, all these things that are working, actually Dataiku in the background is still writing SQL, which means I get, as an analyst, text to visual interface. So I can feel confident I'm getting the result that I want, But data who is still going to write SQL in the background, so we take full advantage of where your data is, the power behind it, and all those great warehouses that your company has been investing in. So no data movement. I just get that nice visual representation. This helps me. I'm off to the races. So now I now have this whole nice pipeline helping me process all of my unstructured data, all those contracts. I found some enterprise level data and brought it together with the help of an AI assistant. Let's take a look at the result. Again, in that theme of validation. We're now at the end, I have my grace period for my different contracts. So now to actually get my answer, maybe I just want a quick visual answer. I can actually write in the same interface, go to charts. I can say, let's look by supplier, and let's create a quick calculation. I can do that on the fly. Or I can just say, hey. You know, we can do this as, like, my grace period minus or my delay minus grace. So I have my delay. Kinda just type it out. Wonderful. Minus my grace period. Let's grab an average of that so you can kinda see our average by supplier. And just like you know, it's a very easy and nice kinda drag out working with visual, and now I can quickly see. Alright. On average, AutoCORE doesn't just have a large percentage like we saw in the very beginning of delayed transactions, but on average, they are hundreds of days late. This is something where we need to really consider if we need to be using the supplier. Obviously, we're on the flip side. You could send them like CarTech is actually on average early as it goes. And this is a pretty this is a simple chart and a relatively simple answer, but it's not a simple path to get here. Right? Finding the data, knowing how to process things, unstructured and structured, knowing how to navigate all the different kinds of things you can do in data and make sure that you can trust the result. The what's really, I think, unique at Dataiku is what happens after this. You know, I've gotten to an insight, which is interesting to me. It is now very easy for me to tag in a colleague to validate my results, but it's also easy for me to think about, you know, how can I start to do things like see if my project meets enterprise standards, automate this thing, and get my results somewhere other people can understand and share? And so, for example, some things I might do, before I even think about automating this, like, maybe it's worth running some standards on this. And Dataiku has that built in where enterprise standards established by my IT group, I can actually say check project standards. I'd say most things look good, but it's actually saying, hey. Your final dataset doesn't have all the documentation it's supposed to. As an analyst, I don't love documentation. I don't think anyone loves documentation. And so, fortunately, Dataiku does have another assistant where in this final result I produced, it does carry forward a lot of the information from our database, so we already had that information. But what it doesn't have is some metadata around the description of the dataset, and some of the new columns we added like grace period. So I can actually leverage this generate metadata assistant in Dataiku. Dataiku will go through, take its best guess of what we should put for those new columns and for the description so that, you know, if we're going to make this something we automate and then set we reuse, we have all the information we want. And, of course, normally, I would go through this and kinda validate what it's looking at, but these all look good to me, so we will approve. And now that information is attached to that dataset. Alright. So let's now make sure this stays up to date, that we're pulling the latest contracts as they come through this latest updated table so that we get our most updated results and that chart stays up to date. We can investigate and use this data and other processes. Fortunately, in Dataiku, that can be as easy as scheduling. Before I do that, you know, I'm feeling a little like, this part from my unstructured data from the LM, I wanna make sure I can trust it. And so before I automate my pipeline, I'm gonna put a quick data quality check-in place. So I'm going to just this result that comes out of my LLM, a really good place to start adding additional data quality data. It could give you a really great GUI to do this where I can create a new rule. And maybe I just wanna make sure, for example, that I'm not going to be getting empty values in a key column. I can say just I don't want them at all, or I can do things like, hey. Over time, can you make sure the empty values are relatively consistent? I really love this bottom option, but for now, let's just say for this use case, we wanna make sure supplier is never empty. Alright. If we're not pulling the supplier out of that option, I wanna make sure we get an actual forced error. And that's really helpful for, like, just me to visualize and understand, but it's also something I can incorporate into my scheduling, which is why I kinda wanted to hold off. So now that we've got a data quality rule right here, looking at that result, making sure we're able to parse and get the right information out of the contracts, now let's go ahead and schedule. I'll keep it simple. I wanna run this once a day. We'll run it at 09:30. And I can do all kinds of useful things, like set triggers based on time or even triggers on things like when new contracts hit that bucket. So a lot of flexibility here. Also easy for me to add things like a reporter, like an email that says, hey. If there's ever a failure, I wanna make sure you email me. But the final thing I'm gonna do is in the actual logic of my automation. Right now, I'm just building that final result, but I wanna kinda add that intermediate step so when it hits where I'm running those checks, that if they fail, the whole thing will stop. And that's as simple as me just saying, I wanna make sure I also build my contract extracted where I built the data quality check. We'll put that as step one and say, you know, build, l m output first. And this last step says run only if no prior step failed, so that's all I have to do. If this fails and I don't get the right suppliers I'm supposed to, we won't run the rest of the process. We won't update the dashboard. We won't update that new dataset. So I can really have confidence even when I'm using some of these new AI tools that I'm gonna get exactly what I expect. Now to top this whole thing off, because I think I've built something really useful, I'm gonna do a couple of things. One, I'm gonna take this dataset, and I'm actually gonna publish it to that data collection so my other colleagues can understand it. And it's actually saying, great. My metadata is complete. Wonderful. Go ahead and add it to the collection. I'm also to kind of alert my central data teams that I think I've built something useful. Maybe they want to, leverage it, scale it across many other business units. I'm gonna go ahead and tag this as ready for review, and I'll tag in, you know, a data team. Can you review? Right? And add my data team. So we went today all the way from it's Monday morning. We have some supply chain questions. We started an agent hub. We asked some general questions about what agents are available, some questions about tariffs, about delays, which kind of sparked the idea that I wanted to get more information about contracts. Built myself a quick agent, which was very useful, but was really based on just a handful of contracts and some tools to go out to the Internet. So I went ahead and searched for a template, which led me into something that made it very easy for me to copy and paste the complicated data pipeline of processing unstructured data that I could verify and validate. And then we use Dataiku's AI assistance to build the rest of that process and find our final answer about KarTech being our most helpful supplier when it comes to delays. So I hope this made a lot of those concepts a little more real, kinda screwed you how you can go from agent to assistant to trusted data to get your answers. And I'd love to hear if there are any questions in the chat for us to review. Awesome. Thank you so much, Chad. I I really think you you nailed it and kinda walked us through exactly what that that Monday morning, question to answer process will look like. Alright. We do have a few questions in the q and a. If you do have any other questions, we have about twenty minutes left. Please go ahead to the q and a tab, leave whatever questions you might have, and we'll get to them. First question we have is when you built that supply chain agent, how did it know which parts of the unstructured data were relevant? Did you define that or did it infer automatically? Good question. So the the quick little one that I built in agent hub, when I upload that document, it's effectively going through some nice fancy embedding process in the background so that when you ask the question, it can do kind of similarity searches across the question to parts of the document. So sort of like auto kind of figuring the context out. Once you start getting to the much more complex questions I was asking the enterprise agent, you know, that involve structured data and unstructured data and more things, some of that's gonna be defined in the prompts and logic and tools that the central data team set up. So you can definitely get pretty far with what's kinda just built in and drag and drop, but as the questions get more complicated, there could be some more, context that needs to be added, both unstructured and structured context. Awesome. Awesome. So another question is, as an analyst, if I'm new to agent hub, what's the best type of first agent to build, slash something simple to get started? I can I'll take a quick stab at this, and then, Chris, I'll pass it over to you. So the great thing about AgentHub is when you are looking to build your, first, second, you know, twentieth, fiftieth, whatever, my agent, there is a specific section where you can get, templates for prompts on what quick agent to create. And that's really helpful because it gives a full context on how the agent should behave, its backstory, its context, and it gives you a great place to start. So, we really mean it when we say you go from, you know, nothing on your screen to having an agent up and running in just a few minutes. So, Chris, I don't know if you want to dive in a little deeper anywhere on that. I I just add that one of the nicest things I I kind of quickly showed that publish button to your to your team. But, you know, if maybe you are not it's maybe your first agent, but not potentially your team's first agent, so you can also get inspiration for some of the things they put together, kind of the the tools they've thrown together to get something done. Yeah. That's a great point. Alright. We have another question. How do you usually check that, what the agent gives back is accurate before sharing it? Yeah. So this is, like, the question of the century. Right? And I think some of the things that I really like in Dataiku and one of the things that I actually, didn't show deeply on the agent side is especially when well, I showed one. When the when the agent's pulling from unstructured data, you can get the actual text that it's pulling from. That's a really nice way to validate. Same thing for, like, the Google search. It'll show you the website it's pulling from. You can go check that website. So the unstructured stuff, I think it's pretty easy to understand whether it's making sense, trusted source. When it comes to structured data, one of the things I didn't show that Dataiku's implemented is it will show you, the SQL and the tables it's pulling from. So if you know the data as an analyst, that's a really nice way to validate. You can even take that SQL, throw it into a SQL notebook in Dataiku and kind of keep exploring and understand. But when you think about that enterprise agent being set up, the huge thing is it's going to be feedback from SMEs like yourselves, other folks that are then being fed back into the pipeline to improve that agent. So there's there's tools right in Dataiku to help you see what it's pulling from and to validate. And then over time, like, whether it's doing the right thing or not are gonna be fed back into the agentic system to improve it. Yeah. I think that's a great point. Even to add further on that, something that I personally love in Dataiku, when it comes to making sure that my agent is actually doing what I expect, or just generally being accurate and pulling from the right places, is to look at the traces. Right? And and an easy way to do that is through Trace Explorer. And what I love to do is to actually go into a specific event node, and you can see within the traces, kind of the rationale that the LLM is kind of thinking. Right? So you can actually trace it from, question to ultimately the output, and really understand, okay, how did the agent actually think through, this issue? And if there's anything specific that you see that you're like, hey, I don't I don't think it should be reasoning that way, that's, where you can then go and add more additional context back to your enterprise agent. So that's something that I love. We have another question here from David. How skilled and staffed does the corporate governance function need to be to enable this type of enterprise wide, ready to be used data? So I think the the answer is it's it's always ever everyone's always ever maturing. Right? So it's not like you finally got to this point and then suddenly you can, like, flip on this sort of self-service world. I do think that governance and understanding how we're making sure we're not duplicating things, getting the wrong answer is something that has to be considered. I think one thing there's a whole area in this, you know, there's people process technology. Right? We're talking about the technology here. But there is an area of Dataiku called Dataiku Govern, which is specifically designed to help approve workflows before they go into production to ensure they're using the right data, to ensure that the right folks are involved in the process, to make sure a business and a risk stakeholder, especially when AI is involved, are saying the boxes are checked. And if not, like, making suggestions on things need to be improved. So I think I think what I mean by that answer is we have lots of technology to assist, but we definitely, definitely, definitely encourage our customers to include governance in the very beginning. And, commonly, I see things like, you know, analyst team as a hub, or as a spoke in a hub that includes central data practitioners, but also central data governance folks. Mhmm. Yeah. And great point. And even to take it, you know, to that broader example is Dataiku thinks about governance through everything that you do. It's embedded in every single process. Everything is logged. Everything is monitored. It's easy to refer back to when when audits come. And that's what's really great about, the platform. Right? It's kind of your second hand when it comes to, the world of governance. Alright. So we do have a little bit of time left. If anybody else has any questions, please go ahead and add them. I see a few I see one more question coming in from Partha. Can an agent be created to automate marketing activities in the social media, as an advertising agent? I can answer that one. Yeah. Definitely. Definitely. So you you can definitely create, an agent to sort of automate a lot of your marketing activities. You know, agents are created to automate copy, generation to, at this point generate marketing on brand images. So you definitely could and that's something that you could do in Dataiku as well. I'm seeing another question. Is there a recording that'll be shared to attendees? Yes. Renata, thank you for answering that one. Any other questions? Or if not, you guys can always contact us, via our website. We are both on LinkedIn as well, so you guys feel free to connect with us. If you have any questions there, you can, ask us. But with that being said, I think we are going to wrap up. We wanna thank you guys so much for joining this Dataiku Live. We really enjoyed, this specific presentation. I think it was a lot of fun creating it. Chris, you can keep me honest on that one, but I I think this was a was a really cool one. I was I was an I was an analyst back in the day. So, yeah, it feels nice. There you go. To to get back to your roots. That's awesome. That's awesome. But, yeah, thank you once again, everyone. The recording will be sent out via email. So feel free to rewatch, and pause and slow down during certain parts of the demo to see exactly what Chris did. But thank you guys once again, and till next time.