Name: What’s New in Dataiku 14.4: Operationalize Agentic AI With Confidence
Uploaded: 2026-03-11T12:09:14.763Z
Duration: 35 min 17 s
Description: What’s New in Dataiku 14.4: Operationalize Agentic AI With Confidence

Transcript for "What’s New in Dataiku 14.4: Operationalize Agentic AI With Confidence": Because we have a huge amount of updates to share. So we can't wait too long for everyone. But if someone misses the beginning, we have the recording ready. Okay. Two more minutes. So in the meantime, if you already can answer the poll that will help me to to decide on the fly where to focus most of our time today. K. One more minute until we start. Just before we just before we jump into that sorry, Dominik. I'm just going to also mention very quickly. I will also briefly mention this at the end is I'm sending this within the chat. If you are very interested in GenAI and how to build these, we are also offering a hands on workshop. Hello, So I'm. sending the link in the chat, and we will also bring this up at the end. So don't worry. The the links will be there. Hi, Anyway, continue. everyone. Okay. Done. A few more seconds. Give everyone time to join. Perfect. Okay. So now it's free after. So hello to our Dataiku release in a box. So that's our first webinar in English to show you the the latest updates of our 14.4 release. This is a new webinar series. about that, So. We would just of the webinar, backstage. we will ask for feedback because depending waiting. on your feedback, we will decide if we can redo this for our next releases in the future. I we just got moved, off. I will walk you through our data re the Dataiku release in the box document. So you will can also find this on our home page. But in addition to only the PDF slides, I can also share some additional insights and also show you. live the examples so it's easier to understand what these updates really mean. We are also sharing me? the slides. after the session, so we will send out an email with all the information that you are seeing today. And then let's jump into our first overview. So that's the thing that we already have seen in the poll. We have lots of updates today. Okay. So that's one of our largest updates. that we. had since I joined joined Dataiku. So it is mainly about GenAI and agents. So at the moment, many developments are developments. are done in this area. link. It worked for me. to that, we also have updates on, machine learning, same error. AI analytics and. insights, Okay. data preparation, AI governance, Let me and maybe AI ecosystem, and also some Dataiku solutions. So we will go through all of them, but the most. Aurora again? today will be on the GenAI and agents side. So far, I can't see the results of the poll. Maybe Claudio can send me the results so that I know, okay, if there is another topic where most of the audience is really interested in so that I can spend more time and deep dive into this topic. In the meantime, I will already start with our first one. That's the Gen AI. Yep. Okay. Perfect. I just got the results from Claudio, so. I know seems to work. So when they open. on a private window, is a big surprise, Jena. and Agents. Reset resets important information. the cache. So if you have any questions, please write them in the chat. If it's something that I will show, then I will There we go. Cool. Okay. I'm going on stage. What happened? Some okay. I I hope everyone is still in the session. Let me check this. So according to the Goldcast, I can still see people in this room. Okay. Strange. Then something happened to our our webinar. But as long as everything is running again, that's the important part. No video. No audio. Video stopped. Okay. One moment. I will start the screen sharing again. Okay. Who can see my screen again, and who can hear my voice? Okay. Although I can k. So it doesn't work okay. Christophe can't. That's strange. Did you already try rejoining? Anyone else having issues seeing the hearing everything and seeing my slides? I will check with our team in the back end. Okay. So passive also can't join. K. One moment. Already Okay. Perfect. So Harry found a solution. Okay. Perfect. So never had this before with Goldcast. That's something new again. Okay. As I I was just mentioning, so if you have any questions, send them in the chat. And if it's something that I that is close to the things that I'm going to show, I can answer them directly. If it's something more complicated we don't have that much time today, especially now that we lost three more minutes due to this Goldcast hiccup. My colleague Harry is is able to answer them, or we if if it's something way more complicated, we can come back to you via email. Okay. Then let's start with the first agent update, and this is about structured visual agents. So there's now another type of agents coming to Dataiku. The reason to introduce that kind of agent is that sometimes you want to have something more deterministic. So if you have worked with agents before, you might have seen already that you can answer ask the same questions three times in a row, and every time you might have a slightly different answer. But there are cases where there are steps that are re really needs to be deterministic, and therefore we have the structured visual agent with which is built by blocks, adding blocks in a row, and then deciding the the flow. So we can we can decide in which area we give the agent more freedom and which area we really say, okay. Here, you have to to do this specific thing and get way more deterministic results. In addition to that, there's also one small thing mentioned here as the third item, the define human approval requirements. So the human in the loop was a feature that was requested by many customers, and now it is here. And this allows us to to add a human in the loop feature to all our tools, which means when when we send a request to the agent, it shows first what it is going to do, and then the human has to say, okay. Either I approve it, I reject it, or I change the request and then go forward. There are many use cases that will benefit from this update and okay. So now let's jump live into the demo. So here, we see a new agent model, which is now called the structured visual agent. On the right hand side, we can already see how this might look like. So here we have a scheduler, a maintenance scheduler agent, which is looking for a root cause error, and depending on what it finds, it will act or do different things. So it first figures out, okay, what is the reactor that it is looking at? It tries to figure out which is the root cause, and depending on the root cause that it identifies, it either executes a Python script, it asks it requests assistance from from a human, or if it doesn't find anything, it will trigger another root cause agent. So as you can see here, depending on the on the feedback that we get from the agent, you can trigger different things. And in order to do so, we have different different things available. So on the left hand side, you can see all the blocks that are here in the right hand side. If we want to change anything, there's the option to say, okay. Let's create a new block, and then we get many more or a lot of different options. So here are all the core AI options where you say, okay. Let's do an LLM request. Let's trigger another agent or force the agent to use one specific tool at that point of time. Then here, there's more about the logic. So for example, you can route and say depending on the feedback that the agent gives us, do different things, or we can run things in parallel or do a for loop until something happens. And also very important, you we can still use LLMs as a judge. The three green items are about memory so that you can save some important things for the agent to use in a later stage. We can also create outputs and also, as seen in the workflow, run specific Python codes. So that's the only thing where where you would need coding skills for everything else. Similar to our visual agent, there is no coding skill required. So let's look into one of these examples. What happens here is, for example, the first one, the LLM request. So similar to our visual agent, you can decide what is the LLM that we are looking at and then write the dis write the instructions, and then you have additional things that you can choose. But that's how it is built. So you you always have these blocks, and then you can fill out the the requirements for the block, and you decide, okay, what is the next block in in that specific case? And that's how we can build a whole workflow. And now let's jump into one specific example to see how this works. I will trigger this again from our agent hub. So the agent hub, our front end, to communicate with these agents. I've selected our scheduler agent and asked them, okay. What's happening on Reactor 2? It should help me to to schedule the maintenance. K. And I think out of time reasons, let's jump a little faster, to be honest, because it will take two minutes to to go through this. It's it's a complex thing complex thing here. One thing that I k. I still need to show it live. Otherwise, I'm I'm missing missing one important thing. It will ask me for for feedback. K. It's interesting. It's not running. I have to go to the to the already demo. So one thing that is usually happening, it will show me a window and say, okay. Hey. I didn't find anything. Shall I trigger the external agent MCP to to to identify the root cause? There I said, okay. Yes. You are allowed to trigger it, and then it came up with an answer. And now let's jump into the this this tool, and then I can show you how you can set up this human in the loop feature that I mentioned earlier. So let's find here's the agent. And for every tool, now you have the option to go to the settings page here in the top right, and there is a block about human approval. So it will ask for if if it has to stop to to get approval from a human, and you also have the option to allow editing. So if if the input that the agent wants to use is not correct, that you as a human in the loop have the option to change something so that everything works as intended. Okay. So that's our first big update. So another way to create agents. Now the second one, again, about agents, but this is focused on or you need to advance LLMesh to use it. It is about agent review. So everyone who already brought agents into production, they will know you have to you have to look at the results from time to time. So we have the agent evaluation option, but now we also have the option to use the knowledge from our subject matter experts to to make sure that the agent is doing the right job. I myself, as a developer, I'm not an expert in in in specific subject. I'm an expert in creating the agent, but for me, some of the answers, they sound they sound good, but then subject matter expert might say otherwise. And that's why we created the agent review. So we have the option to to give the agent a list of of questions, give them also a list of answers and expected behavior, and then an LLM of as a judge and a human will decide if the agent is doing the right thing. And here, again, let's jump into a live demo so it's easy to understand what's happening. So we go to our a GenAI tab on agent reviews. You can see I already added some questions in here. Let's jump into one of them, and that's how this information is added. So a subject matter expert needs to come up with a test question, also have to provide the, sorry, provide the reference answer, and also mention what is the expectation. How should the agent come up with the answer? So here, we say it has to call it has to call the knowledge bank information to retrieve insights about the refund and credit policy. What else do I have to set up? In addition to the question and the answer, we have to set up some traits. Therefore, we can go to the setting. We need to decide what traits do we want to add. For example, the reference, so is the the answer correct? Expectations, so is the agent calling the right tools? But we can also add some more things. For example, tone. Is it giving any toxic answers, or is the tone as we as we want to have it? Or also look for hallucinations. So you can add your own traits. You you have to describe the criteria. That is used for the LLM as a judge. And, also, you have to decide for every trait what LLM do you want to use. So there might be some LLMs that for example, here, one specific task where we say we need a more powerful LLM than, for example, for the hallucination part. And then as soon as everything is set up, let's have a look at the results. So we can see my agent is not running as intended. So we have only three passes, nine fails. That is decided by as soon as every trade is is marked green, it automatically say, okay. That's passed. As soon as one of them is red, it says, okay. It's failed, and it also mentions what is the reason why it failed. And here you can see there's something different in this first case because we see the reference is red, but a human so my colleague Steven said, hey. I'm a subject matter expert. I would say the answer is correct. So the LLM as a judge was not right. Let's override this information, and I can say, okay. Everything fine? What can I do now with this information? I'm now able to compare different versions of my agents, for example, and see, maybe my last run, which had nine failures, is worse than my version three agent. So here I can see I had eight I had eight correct answered questions and only three fails. So I can see maybe my last developments were not going in the right direction. Let's also have a look at the risk history. So here we have only the summary, but we can see the summary overall runs and see if we are developing into the right direction or if we need to go back to to another version and start all over again in order to to improve results. So overall, the review is here to to make your life easier, to to also include your subject matter experts in figuring out if your agent is doing the right job. And it also helps to convince your team that that the agent is is verified. So because as soon as subject matter experts, they already have contributed to developing the agent, they have seen the answers. They have reviewed the answers. So getting such an agent into production is, I think, way more smooth because there's no no people who say, hey. My gut feeling says otherwise. So a very cool tool to bring agents into production. Then let's jump back into the slides. So we are talking about another tool for agents, the semantic models for agents. What is this about? Usually, agents are dumb because they start with an LLM, which has no information about your data or your business. And that means, for example, if if the agent tries to write SQL code, it has to work with the information that you give the agent. And to be honest, I think everyone already has seen SQL tables where you can't can't figure out by the name what's in the table. And, also, when you look at the different columns, if the column name is not not not good, you you sometimes don't understand what is in the column. And how should the LLM know based on this information what you what it needs to do when you ask it a specific question? And therefore, we have now the option to to set up entities and relationships and also add a glossary to to enable the agents to to answer specific tasks that are also using your business your specific business language. Again, we have a demo for this, and let's jump into one example. Already I already ran through it to save some time. Yes. So the question there was, one which is a request. had lowest you could show us, in 2025, the c details and download, any vehicles with they both do volume? in the previous case? And here we already see Cool. one issue. The agent, Nice. Very nice. normal agent, wouldn't have any clue what does PMA mean. Even I, at the moment, to be honest, I I don't have a clue what what is the PMA, and that's something that we have to to train the agent on. How can we do this? We are setting up a semantic tool, and the semantic model allows us to to add some entities. For example, we give this agent access to four different datasets. And for each dataset, I'm starting an entity. Here, I can enter the description. So what will the agent find in this dataset? It doesn't have to be a dataset from this specific project. It can also reside in another project. And also add some primary keys and the attributes. Here for the attributes, you can give information in the description field. So that is needed to needed to show or this information is given to the LLM so that it understands what is in the name field. And also, there is the resolve option. What does the resolve option do? We so we have one case, for example, we say we are expecting, for example, countries like France, Brazil, Germany, but in the column, it is it is only abbreviations. So there's one for example, f l one, which stands for France. So here in the description, we can tell the agent to say, hey. Here are abbreviations of countries. So then it can with the resolve option available, it can automatically understand that f r one stands for France. Also, thing, here is the the magic. What is the PMA for? So PMA stands for profit margin analysis. Here, we have the definition for net profit margin. So we already give the information how this specific KPI is calculated. But at this point, the agents still have no clue that this is what we understand on a PMA. That comes next. So what else can I what else can I give just to my agent so I can see many questions in the meantime? Harry, I hope you you are covering all of them. You got that. Perfect. Perfect. I will come back to this after I've explained the the semantic model. Okay. Good. So what else is important? Usually, when we have lots of different datasets, we have to tell them how how do they connect to each other. So here, we can say how how can the the agent join different datasets. And here in the glossary, here's where the magic happens. For example, the net profit margin, we already gave the definition. Here, we can give a description. And here, we see a synonym for the profit for the net profit margin is PMA, and that's how the agent knows what PMA stands for, how to calculate it. You can see this was a manual entry. Here comes something else. For example, if let's assume a CEO sends an e a new email where it says, okay, it's talking about the big three. The big three means it's France, Germany, and Brazil, and it also talks about growth engines and mentions India, Turkey, and Morocco. Now many colleagues will ask these questions to my agent, and my agent doesn't have any clue what big what what they are talking about. So either I have to go to my agent and now add this information manually, which is fine if it's only one or two information, but you can see there's multiple information here. And to make my life easier, I can import these rules from from from documents. So if I want to add a new term, just click add terms, say extract it from a document. I upload my file, and what you can see here is, for example let's figure out here's the big three. It was manually extracted, and it gives me the information. Okay. Big three is the term. It already came up on synonyms on their own, and it tells me, okay. Here's the definition. Here's the examples of the countries. And that's all that my agent needs to know or the tool needs to know. And in addition to the terms, we have to link these terms to the specific columns. For example, with the big three, we have to tell them it has to look for the attribute for the name of the market so that it can filter on the right on on the on the right country information. And that's it. So let's go back to our agent, figure out what what did it do. And when we go to see details, here we can see where the magic happens. So it was able to identify the the definition of of PMA. So that's the definition that we have given in in the semantic model information. In addition to that, it is filtering to 2025, and it also understood us that it is only filtering for car sales, which has a positive sales volume. So it filters out the non non zero sales by the by using a having clause, and it also used information how to join different datasets. So that's where where the magic happens, and here you can see the results of the of the of the query, and it gives us only it gives us the lowest PMA as feedback, which is 18.85%. Okay. So I think it's a very powerful feature. It takes some time to set it up, but if you are working with some specific SQL tables, which might be cryptic for some LLMs to use, that's a very good way to boost the result quality of your agent. Okay. So a few more slides on GenAI, but these are covered pretty really fast, so then we can jump to the next topic. So so far, I've shown you always agent hub to consume to consume agents. So this is one way. The other way is you have the agent directly in your free flow, and it is working with your data, and you are not starting a conversation at all. And the third option is now you can also use the tool that you use on a daily basis. So if you have Slack, you can connect to our agents and and directly ping them from from a channel. So it's it's another way to consume the agents in the future. We already had questions for Teams. So when is Teams coming? Our team is also working on a Teams integration, which will come in one of the next updates. So far, it's it's Slack that we started with Microsoft Teams is is in the future road map. In addition to that, we have four slides. We have a similar topic, and this is about how can we manage and orchestrate agents that are outside of Dataiku. So you can do this now with with agents that are built in Snowflake, Databricks, AWS, and and Google. So we have all the all the tools available to to to also bring these agents into our system. Why is this a benefit? It helps us because you have the same testing as on Dataiku build agents. You can also do data quality and monitoring on these. And in addition to that, you can also use, for example, the the Cortex Search and Cortex Analyst tools as a tool when you build your agents in Dataiku. So you can use very powerful tools if you are if your data resides, for example, on a Snowflake system. In addition to that, similar for Databricks. So despite using Cortex, you can use Genie AI. And the same is for AWS for AWS Petrol agents and for Google Vertex AI agents. So, again, pretty powerful. You can have now one platform where you can can monitor all your agents. Okay. So half of the time is already over, and I I mentioned the JAI block was the largest one. So most of the updates updates on on in this area. Now let's jump into the data preparation for AI. And, again, we have some JNI magic working in the background. And here, our flow assistant, that's that's evolution of our generate next step assistant. So by describing in plain English what you or in any language as you like, you can describe or you just describe what you plan to do, and it automatically automatically suggests you which recipes to use. And as soon as you say, okay. Hey. This looks good. It automatically creates the whole workflow for you. So it is a tool that makes your life easier if you are, for example, new to the to Dataiku or you want to speed up some specific building processes or if you are stuck and you don't know what to do next, just ask the flow assistant, and it might come up for with the right answer. One important tip, don't write a huge essay where you describe your whole project. It makes more sense to have small tasks. So if you are building a very complex flow, it makes the life of the flow system way easier if you if you cut it down into some smaller smaller scripts, similar as as you build the the tool yourself. So if you build a flow, you you also start step by step with different, for example, different flow zones, and that helped a lot with with getting the right answers. Let me show you live how this feature works. So let's jump into the flow assistant. It works the following well way. So I just select the the datasets that I want to work with, click on flow assistant, and then the selected datasets, they