Skip to main content

Democratizing Pharma Insights: Dashboards and AI Agents in Databricks

Summary

  • Novo Nordisk built a unified solution on Databricks combining dashboards, Genie, and knowledge assistants to enable non-technical clinicians to self-service insights from complex clinical trial data without requiring SQL or programming expertise.
  • Metric views serve as a semantic layer that standardizes clinical calculations and dimensions, while Genie is fine-tuned with industry-specific skills for SDTM and clinical protocols to ensure accurate natural-language query results.
  • A new integration between Genie and knowledge assistants on the Databricks Data and AI platform enables users to ask questions spanning both structured clinical datasets and unstructured protocol documents in a single governed workflow.

Democratizing Pharma Insights: Dashboards and AI Agents in Databricks

Watch: Democratizing Pharma Insights: Dashboards and AI Agents in Databricks
Pharma data is complex and fragmented across source systems, files, and documents. Clinical trial data spans regulated data elements with intricate dependencies. Traditional dashboards grow unwieldy and rigid; chatbots hallucinate and lack context. Non-technical users struggle to explore data independently. Novo Nordisk faced this challenge: enabling clinicians to gain insights from complex clinical trial data without technical expertise.
this video shows how Novo Nordisk built a unified solution combining dashboards, Genie, and knowledge assistants. Metric views serve as a semantic layer standardizing calculations and dimensions. Genie gains industry knowledge through fine-tuned skills for SDTM and clinical protocols. Knowledge assistants enable natural-language document querying. A new integration lets Genie answer questions spanning structured data and documents, enabling non-technical users to self-service insights with full governance.
🤝

Chapters

FAQs

What is SDTM and why does it matter for clinical data analytics?

SDTM (Study Data Tabulation Model) is an industry standard defined by CDISC for organizing and formatting clinical trial data submitted to regulatory authorities. In this video, Novo Nordisk demonstrates how Genie is given fine-tuned skills for SDTM and clinical protocols to bridge the gap between clinical industry standards and natural-language analytics for business users.

How does Databricks Genie work with pharma clinical trial data?

Genie is configured with industry-specific skills for SDTM and clinical protocols, enabling it to translate natural-language questions into accurate SQL queries over clinical trial datasets. Novo Nordisk uses metric views as a semantic layer to standardize clinical calculations and dimensions, giving Genie the context it needs to produce reliable, governed answers.

What is a knowledge assistant in Databricks and how does Novo Nordisk use it?

A knowledge assistant on the Databricks Data and AI platform enables natural-language querying of unstructured documents such as clinical protocol files. Novo Nordisk uses knowledge assistants alongside Genie so that users can ask questions spanning both structured clinical datasets and document content in a single interface with full governance.

How does Novo Nordisk enable non-technical users to explore clinical trial data?

Novo Nordisk built a unified solution combining governed dashboards, Genie for natural-language SQL queries, and knowledge assistants for document queries, removing the need for technical expertise to access insights. Metric views standardize clinical calculations so that clinicians and business users receive consistent, governed results without writing code.

Full transcript

[00:08] Hello everyone. Good afternoon. Uh thanks for coming to our session today. My name is Half. I'm from Databricks. And today Cena and I will talk about something that we're both passionate about. Uh democratizing access to insights. Um and today the
[00:24] the focus of of the presentation is in the pharmaceutical industry. Um and we're going to use the clinical documents in the clinical trial data as an example. But uh we believe that you came here because you're somehow familiar with that. But we do believe also that if if
[00:40] you're not, the examples can be easily translated to to other industries, right? So the business context then um clinical study data and protocols. Um a bit of myself, I'm Half Scheidel. And I'm a delivery solution architect at
[00:58] Databricks. I've been working with Novo Nordisk now for 2 years helping them accelerate the use cases. And uh today joining me here is Cena from Novo. Cena, a bit about yourself. Yeah, uh I've worked at Novo Nordisk for
[01:14] 30 plus years and most of the time in clinical development and being involved in numerous of our regulatory submissions through all the years including Ozempic and Wegovy, which you might know. Um
[01:29] I'm currently in AI and digital innovation trying to bring, you can say, my pharma or my business knowledge into play in relation to doing data modeling. And also trying to spearhead, you can say, the adoption of Databricks among
[01:45] business users and trying to change the narrative from Databricks being perceived as a data engineering platform to more like a unifying platform where user people can come together and work with the data, govern the data, and
[02:03] obtain insights from the data. Good. Our presentation today is 40 minutes, four parts. We would like to describe why the this data and the data models in pharma are so uniquely hard to to work with. Um and
[02:21] where the dashboards and chatbots fell short up to today, right? Then the context, a little bit about this clinical trial data, what are the the data models that are used, and the one example of the documents. Um then the third part is to
[02:39] um describing what is what is it that you need to do to make this data business and AI-friendly. And finally, the fourth part, the most uh interesting, I hope, is the demo of everything we work here. Uh which is to show you the the how all of this comes
[02:55] together in in Databricks with both a dashboard, Genie, and then also um a knowledge assistant. Uh so, let's get started. Cena, what is the problem? Yeah, what's the problem? I'll start with stressing that data doesn't equal
[03:12] understanding. And more data doesn't equal more understanding. In contrast with what many people think, it really depends on the quality of the data and whether you can navigate it. More data might just be more ways to get completely lost. So, let's start with
[03:27] the challenges we have with pharma data. And I I think it goes for many other industries. I'll put here that pharma data is complex, messy, and disconnected. And what does do I mean with that? You can say we have everything from receptor chemistry modeling data to
[03:45] metrics, milestones for clinical studies, and yeah, blood samples, imagings, and so on. So, it's a lot of different types of data. It's not as such messy or at least not for each individual study. But, the the way we we want to use the data today where we
[04:02] bring data together from many different studies and we want to mix and match the different types of data. Then, it starts feeling messy. Firstly, because it's in so many source systems, it's in file shares, and it's in different data tables, formats, and it's
[04:19] in documents. A lot of the data is really in document and most of the context is in documents. Uh so, you can say cleaning is a must. And it's a must because precision really matter. You might have a mixture of blood samples taken with glucose values
[04:36] from the US in milligram per deciliter and European blood samples in millimole per liter. And if you mix it, the mean values doesn't make sense at all. Um and the context is super super important. It really depends on what questions you asked in the studies
[04:52] whether you collected something systematically or it was more uh you can say spontaneously reported. You need to know that. Um then come to the other problem. And and that's about dashboards. I really love dashboards. It's It's an amazing way to convert uh
[05:10] to uh communicate patterns in data. And we saw more many examples this morning of dashboards basically in all the demos. Um but, you can say they will only traditional dashboards will only cover what you had
[05:26] had already thought about addressing. So, they they they don't adjust to the user. And over time, I think most of us had seen it, the dashboards grow and grow and grow in complexity and get more and more of them. And then it just becomes a maintenance nightmare. And
[05:42] today everyone says, "Yeah, go for the chatbots. That's the way to go." But, then you go into this, you can say knowledge gap really. The chatbots might hallucinate and they are often not aware of the context. So, a really
[05:58] good looking answer might be completely wrong or wrong in this context. And the users actually often don't know have sufficient data knowledge to ask the questions in the right way. I see many examples where user comes to me and say, "What can I ask the chatbot about?"
[06:14] Yeah, you can just try it out. How do I ask in the right way? So, so that that's really the problem. And you also had some considerations. Yes. So, even though I am from Databricks, I do recognize that there is a in this context a challenge, right? So, even though you have all the
[06:31] governance you need to catalog, but still if you have tens, hundreds of tables, it's still challenging to connect, right? And then as Sinan mentioned, the the logic then that resides how to connect these data sets and and how to query them sits in
[06:47] multiple places. You have ad hoc queries in notebooks, you have scripts, you have dashboards. And that's really difficult to then to source. Business definitions then also living in in in code in different places and in many times that people's heads.
[07:03] Which is linking to very interesting to the tacit knowledge that was mentioned this morning, right? Lineage is great. It works great. It's there already. But it tells just where the data is coming from and not so much of what it means. Dashboards are similar problem. So, if
[07:19] you have logic or metrics that are defined across different widgets or even across dashboards, it's very easy for them to go out of sync and then the numbers will drift. It's very difficult to then keep these calculations in sync if they're spread
[07:36] out. And about the chatbots, right? That the Sinan mentioned the there's a challenge to make chatbots uh relevant and they need context, they need synonyms. Next. Um, they need also
[07:51] instructions, right? And again, almost right in in pharma is never acceptable. I'm pretty sure in your industries and your work, too, that we want the answers always to be correct. And then, um, what to do then? Yes, what to do.
[08:07] The key is really to create structure, guidance, and boundaries. And of course, also uh ensure governance. Governance is a big deal with pharma data. We're not talking about it here. It's a major topic in itself. So,
[08:23] tidy the data, that's where everything really starts. I can't stress it enough. It's super time-consuming, but it is a must. Then, in the what we have been experimenting with this is really then at the fundamental there in the bottom
[08:40] to make the data more you could say business-friendly, AI-friendly, dashboard-friendly by using metric views to to capture measures, dimensions, descriptions, and so on. And then, you can build build dashboards that kind of cover maybe 50-70%
[08:58] of of the interesting stuff in the data to to to guide the users in what they can see in the data. And then, a chatbot can supplement, you can say, with further questions. But, make sure that it's locked in and only have access to the
[09:15] data that you decide it has access to. And then, of course, train the users, train the the chatbot, and feed back any learning. So, before we go on and and show more concretely what we have done, I just want to take you into a little tour
[09:32] about working with clinical study data. So, it all starts with the protocol. It's this the all the clinical study protocol we have heard about many protocols, but here it's the clinical study protocol. It's really this
[09:47] very, very detailed recipe for how to conduct the study and and why you conduct the study. Then, lots of planning. So, we run global studies. So, it's endless coordination about times across
[10:02] time zones, zones, cultures, countries for all the many activities that needs to take place in the right order. The conduct of studies really in the hands of the investigators at the hospitals and clinics.
[10:17] They recruit the patients. They inform the They get the informed consent. They administer the study medication. They do all the assessments. They enter the data into all the study nurses enter the data into the
[10:33] clinical electronic data capturing systems. Then, we map the data in Pharma to to structures, do the analysis, and finally wrap it all up into the clinical study report, which covers all that aspects. And
[10:49] around all of this, lots and lots of regulations, and it has to be like that because ultimately it's about human safety and well-being. So, it's it's a just there for a very good reason. Let's then look at the underlying data
[11:05] because this is about data, and and you can say each element creates data and depends on data. And I've just illustrated some data domains here. So, you have study design, you have study participants, investigational medicinal products. They are considered those as
[11:21] separate data domains. And then at the top, I think I'll study milestones, study sites, investigators. And here, just I won't go into details, but just some examples of what the data could be. This is just a tiny little fraction of the data.
[11:37] If you model all of this, you can connect it and you can go on endless data journeys and get a lot of insights. The great thing in relation to modeling is that within pharma you have very very comprehensive, well-documented external data standards. So, you have the CDISC
[11:54] Unified Study Data Definition Model for study design. You have the ISO IDMP model for medicinal products and you have for example the CDISC SDTM model for patient data. And then on the top, that's more the clinical operational
[12:09] data and that's no sort of predefined external model for that. Uh but before we go on, I'll just take take a little bit deeper dive dive specifically into the SDTM implementation guide
[12:24] or implementation the SDTM model. It's maintained by CDISC. It's mandatory that all your clinical data when you do regulatory applications, your clinical data has to be submitted to for example FDA and it has to be to be structured
[12:40] according to the CDISC SDTM model. It's a very good model actually for aggregation of data and analysis and for you can say cross-study stacking of data and all of that. And
[12:57] there's a very comprehensive SDTM implementation guide which we have actually used to model our data and semantically enrich the data. So, and now a little bit about the protocol also. Yeah, the document. So, Cena spoke about the structured data part where all the
[13:13] tables are mapping to to domains and and how is that part of the industry model. Another component that was mentioned was the clinical study protocol. It also follows a a template. There's also a specification for that an industry
[13:28] standard Uh uh and it's a required document so it's extremely important in this journey that Cena mentioned from going from a molecule to to a drug in the market and the contains the recipe and it's required for safety and and ethical
[13:44] reasons as well. So there is a template which means that you would expect this document to be something that you know it's easy to understand easy to find information. Um it sounds neat on paper but the reality is like many many documents is that
[14:00] it's very hard to use on a day-to-day basis. Uh why is that? Hundreds of pages uh amendments appendices figures and and data plus pros mixed in the same place concepts are repeating and spread out
[14:17] and the answers are scattered right so if if um um a researcher or a biostatistician that is seeking information there is going to be in different places. Different needs by different readers. Um yes you can search for content in that
[14:33] but that's not enough right so what happens is that um um going to study protocol making comparisons across studies is very very time consuming and it's a very error prone too. So the perfect problem for for better tooling perfect problem for
[14:50] for data bricks too. So what to do in practice then right so what are these suggestions then would look like in practice Cena? Yeah. So what have we done here? So firstly you can see first piece is to map your collected data into SDTM data.
[15:06] That was not part of our exercise that was luckily already done for us. But then in in in the middle basically what we did we used data brick matrix use to actually model the SDTM data. So the SDTM data is the data is in many,
[15:25] many different data sets you can have as could easily have 40 data sets from one clinical study, of course, depending on what you collect in the study. And all of this should be linked together. And we use the metric views to define the linking. We also use
[15:42] the metric views to add descriptions for columns. We use the metric views to define dimensions and and measures. And actually, we use the SDTM implementation guide for adding the the descriptions. And then we try to use more
[15:57] user-friendly column names and so on. And all of this then actually serves as a very thin semantic and analysis layer sitting between your SDTM data
[16:13] and the dashboard and Genie. And the good thing here is what we used we used dummy data or or synthetic so no Novo Nordisk studies, but I can just redirect all my metric views to a real Novo Nordisk study and reuse everything I've
[16:29] done. So, so it was a little tedious to do, but you can just reuse it then endlessly. And it make the dashboard build very, very easy. So, the flow just high level, the SDTM data mapped into metric views, and then the dashboard is
[16:46] built on top and the Genie's base is built on top of the metric views. That actually means I can fully control what the users can see. So, if they're not allowed to actually see data on subject level, I can just remove subject
[17:02] ID as a dimension and only for example allow them to see it by sex, by country, by treatment arm so forth. Just for those of you who have never tried metric views, that's actually a a pretty nice low code
[17:19] low code click and point UI in dashboard in data bridge where you basically define your joints, your measures, your dimensions, and you can go in here and add your expressions and so on and there's this magic little pen
[17:37] everywhere that you can use to chat your way into what you want to do. But actually we also use the Genie in another way, more sophisticated way. Yes, so what we did was Asena in the process, right? Was
[17:52] teaching me about the industry. It's something that she pointed me to to the implementation guideline, right? So I knew about the standard, but then I said how there is an implementation guideline which it basically shows this data models and contains the descriptions for for all
[18:09] these data sets. And then I it was very very clear that we need to bring all of the descriptions, right? So some were present now, but we need to bring all of that. So what I did then was just to package that as a skill and then
[18:24] fine-tune the skill to to work for this use cases, right? So if you want to apply all the descriptions from the implementation guideline, if you want to ask for deviations, does my model then have any column names that are different? That skill helped a lot and that it was very
[18:41] helpful to then not spend a lot of time. As Asena mentioned, it can be tens of different data sets and having that skill was very useful. It can be used in other places, too. You can fine-tune that to to help you with synthetic data generation and so on. So So feel free to
[18:57] to try if you work in in the industry or in related fields, right? Then beyond Genie. So now we're going to back to the documents, right? So Genie then with the enhancements that we
[19:14] described, so Cena mentioned the importance of that foundation adding the descriptions, making sure that the model is well-documented, it's great. You can answer questions from various complexity levels going from basic counts to to to more
[19:31] specific, right? Requiring multiple places to be located. So, what is the mean change in body at week 12? Um However, there are questions that we need information from the protocols. So, we need to go back to that document
[19:47] and locate the section and infer the meaning there, etc. right? So, the knowledge assistant then is the the the specialized to the agent that is used for that. With that, you can ask about um um the
[20:03] specifics of of the the trial, for example, here asking about the treatment emergent and then the inclusion and exclusion criteria. And then with that, you can ask questions. The same way that you can ask questions to Genie, you ask questions to to a knowledge assistant to fetch data
[20:18] from documents. But in most problems or the more interesting ones, you would like to ask questions that span the structured data and your documents. And that's where the the supervisor agent
[20:34] comes to play, right? That you can have then a a in a brain that will then take your question and and decide if that should go to the to the structured data or to the documents. Sometimes even spanning multiple
[20:49] questions and combining that to to provide an answer. What is what is new? This is probably familiar to many of you, right? So, what is new is that in this context where the supervisor agent is primarily using a Genie and a
[21:05] knowledge assistant, Um, is an a more streamlined experience, right? So, as you know, the the supervisor agent could have multiple knowledge assistants, multiple Genies, and uh checking MCPs from other sources, etc. Um, but in this context, then there is a
[21:21] more streamlined version, which is this new uh integration between Genie and a knowledge assistant. Um, that that is something that in this in this scenario where we have just the a Genie, right? So, if your
[21:36] users already are familiar with your dashboard, you have a Genie space that is working well, and then in parallel you Oh, there's knowledge assistant. I have my documents. I created a knowledge assistant, and uh they're working well. You don't need to go back to the supervisor agent. You can simply now
[21:51] just add the knowledge assistant to to an existing Genie, and then um uh the questions that you ask from the Genie space will then consult documents and reply the the according to both the structured data and unstructured data. Um, that's what we're going to demo
[22:08] today. Um, so about the demo, Sina. Yeah, about the demo. And I I must admit with when health showed me this, I was kind of woah. That that's a major game change, and that you can really seamlessly integrate interacting with documents and data at
[22:25] the same time. That's a huge huge strain uh change, I think. The demo we are going to show is uh I want to stress it again, it's on synthetic data. It resembles It resembles patient data in an SDTM structure, and and it's also on
[22:41] made-up uh protocols. We reverse engineered the protocols based on the M11 template, and actually uh effects are based on the data. So, we also used AI for that. Yeah. So, we are trying a live demo. We did
[22:57] recordings, but now we're trying it out live, and we hope you can see it. Yes, it's there. It's there. Okay. So, this is just a very standard
[23:12] dashboard, and and first I just want you to Basically, I've used the data bricks in-built SQL dashboard, and I tried to max out on what I could make it to to do. It's For those of you who are from No knowledge, you will see the colors.
[23:29] So, I coded the our CVI colors so that everything looked sort of styled our way. Uh so, it's just with a number of tabs that you and I problem with the mouse. So,
[23:45] if it had been sort of like a full study, I would have more tabs. So, so I'm sort of on the tabs giving some like key numbers. And here's an excellent example of why pharma data sometimes is a little bit difficult. So, I have
[24:01] actually three numbers for number of subjects because I have number of screen, numbers randomized, and numbers completed. And if if you if you don't know that, then you might accidentally hit the wrong number. You can say and think that that's the only number.
[24:17] Then uh if I go back to study design here, I show another example of filter down on on one of the studies.
[24:37] I think I clicked it in. Yeah. So, so here again, I have two sets of numbers. I have subjects treated, 41, and subjects over here, that's 39. It's It's two different numbers because one is planned, the other one is actual. So, so you have so many numbers that you can
[24:53] get get lost in. Let me remove my local filter. Just go to adverse events. This shows a typical adverse event kind of just
[25:08] typical graphs. I'll go down here. Actually, in the beginning I didn't trust that much in Genie. I I tried her. For me it's a her, but it could also be a him. But I I tried it out on the directly on on on the SDTM data
[25:25] and it was really good at percentages and rates and so on. So I had to make a sequel function to calculate you can say a lot of numbers. So so so basically when you look at adverse events for example you
[25:40] it's not just counts of adverse events or counts of subjects with adverse events. Counts are always super super easy. But when you have to do comparisons you need proportions of subjects with adverse events out of those at risk or rates of adverse events if they where you take the exposure time
[25:57] into consideration. So did I kind of the sequel functions I can hit here. So let's change to sex. Then it recalculates. So now it's by sex and let
[26:13] me change to maybe body system organ class. And here these I didn't change because actually this is the SDTM data. So that's why you see these odd abbreviations. So now it's recalculating and it actually goes it calculates lots
[26:28] of numbers pretty fast and I also tried it on a stack of 300 studies and it's super fast. Later I realized okay I can actually also ask Genie. So so Genie can actually So let me ask
[26:43] Genie.
[27:03] And now we hope it performs. We just decided this last minute to do it live rather than recorded. It's always a big risk. So, it's it's uh thinking. It says I can skip to a faster answer. I'm more into uh
[27:18] solid answers than fast answers. So, so I won't do that and then we cross our fingers that it's not too slow. Um or that connections
[27:34] It's It's It's connecting. Shouldn't get there. But But you can see already in comparison uh in a way sometimes it's actually easier. It can be faster with the old way. So, what did I get here? I got that it was actually 85%
[27:50] of subjects that had adverse events. Let's uh Let me close this window. Okay. If you're an employee, you're probably sharing my pain now. Yes. Yeah, so let's let's check the number.
[28:06] What's it doing there? It's misbehaving. Let's try again. It just disconnected somehow. Uh
[28:28] That you can't plan your way out of. So, we apologize. It's It's nothing It's It's our computer so Yes, I think now it's back.
[28:44] Okay. Thank you. So, But if I click here? Yes. Thank you. Yeah, so if I click here, I can do all subjects. And I can do just all adverse events.
[29:03] Yeah. And then hopefully I got the same number 85.3. So So I've started to trust a little bit more in Genie after
[29:19] done lots of cross-checking. So basically you when you make the metric views, you stabilize a lot of things and you can predefine your unique subject counts and your unique assessment counts. So you
[29:35] can define things in there. I haven't precomputed all possible combinations. I've just made some basics that it can then actually combine in the right way. And one other nice thing and and I'll just
[29:50] This is the last thing I'll show in my demo because I really give want to give health time to to show the new stuff. So and maybe we have to deny print it in. We have 10 minutes. Okay. 10 minutes left, I think. Is that it?
[30:10] Yes. So Okay. It's blocked. I can guarantee you it actually does make some pretty nice PDF and it can make Genie can make graphs and tables and so on. So depending on what I asked it about, if I had asked it about the proportion of
[30:25] subjects with adverse events by section treatment group, it would likely make a graph. Could make it a little bit easier for me to to look at. And in the PDF reports, it actually inherits the styling that I put in my dashboard. So it's the same colors, same
[30:41] fonts and all of that. I think you will have to trust me regarding the PDF or watch my recording and then Thank you, Cena. Thank you. So, metric views is essential to to make a this kind of um
[30:57] disconnected model um with lots of tables to to give a accurate answers, right? Uh then the document, so going back to the documents, uh the knowledge assistant, uh it is a a simple type of um agent here so in Databricks today
[31:12] compared to a year ago where you had to stitch together these components, now you simply point to a folder, here are my documents, and it's going to be ready, right? So, it is really simple and then um you can see that for my knowledge assistant for clinical trials, I just pointed to the protocols. Um the
[31:30] protocols then uh were created also using some um magic with uh Databricks AI query and also using an SME, right? So, Cena was essentially in in providing um context for generation of these protocols. And then just to show, right?
[31:46] I talked about the inclusion criteria, which is uh is the example that I'm going to show now. So, so the the the protocol then in this example shows the inclusion criteria as a combination of numbers but also just text. Uh participants must must be male or
[32:03] female, etc. Age, then there's body mass. There is um um uh this index is like written with text. So, there's a lot of um understanding here, right? To for you to translate this and make comparisons. Uh it's not a um an easy thing, right?
[32:20] And then also exclusion criteria. So, these are my documents that were then um included in the knowledge assistant. And in the knowledge assistant, I could already start to to ask questions, right? So, here for example, the let's do that.
[32:36] The oops. Um which studies are present in the protocol documents? And then within this knowledge assistant context right would already search the
[32:52] documents and then provide the answer. So it's going to say, "Well, based on the documents, these are the study IDs and their their titles." I could also ask a more a bit more complicated question right a bit more sophisticated. Compare the
[33:08] randomization approaches across studies and highlight the differences. So remember we're still in the documents. This could be other types of documents. It could be multiple sources. My knowledge assistant has only one source. But you could add multiple right? You could have different types here.
[33:24] And then here so then now it needs to go into the documents, understand where the the what is randomization, um bring that back and compare right? So the main differences include number of arms etc. And always uh showing the the sources.
[33:40] Very good. So now we have the documents. We know it works. It's great. What about then the combination right? So here's where a a question on the Genie space then uh combining with the with the knowledge assistant
[33:55] comes handy. So I just want to show you also in the UI. So if you were to add this as suppose you already had a a Genie space. It's now in data right? If I go to data, you're able to add here. So you can see I have all the metric fields that's Genie created. The knowledge assistant becomes a new source
[34:12] here for my Genie space. Um and this is it becomes like this. Great. You can add multiple too. I just have one. So an example here right? What could be a question then that just to validate that I I still have that here. So within Genie
[34:28] um Yeah, which was body weight part of the primary outcome. Then this we know that um uh now I'm the genius space. I have the metric views. We know this is in the protocols. Then uh just to validate that
[34:44] it's doing that, right? And then a more um relevant question here. Can we try something that requires uh information from both? And that is the question here on the Oh, sorry. Oh, this one.
[35:07] Yeah. I practiced reading all of this, but that's let's let's see if I can get it right. So, for each study, do any subjects fail to meet the protocol defined inclusion criteria due to clinically significant abnormalities in vital signs?
[35:23] So, this is a very long one and it's a but it's the same time you can Yeah, I understand what this means, right? So, are the subjects uh in their um vital signs according to the inclusion criteria? Um but to you cannot translate
[35:38] this to a single query, right? If the information of what inclusion criteria is sits in a document, uh then the agent will need to go there first, right? And then parse that text and translate that to queries, translate what are the numbers that the um
[35:53] um mean the vital signs for inclusion criteria, and then go back to the metric views to query that then. Um due to time, I might just show you the Yeah, so just to show you the these steps because then this already gives
[36:09] you some idea how it was working, right? So, you went to the documents to get the the criteria, and then now it's trying to, you know, based on that text, creating multiple queries for for the the vital signs. So, blood pressure um
[36:25] and then statistics etc. etc. Then translating the numbers, and then we'll provide and then of course if you click in each of these, it will show you the intermediate steps and intermediate data sets, and eventually then we'll combine the answer and then um um show you that uh I think in this case
[36:43] none um which is good. Um but this is the level now questions that we know can trigger uh the questions uh over documents. You can have multiple sources and then back to uh to the structured data queries. Good. I am just going to switch back for
[36:59] the wrap up? Yes. And then hopefully there's time for one question. Yeah. We will also be available afterwards and if you want to see more, then that's
[37:15] also So, basically to to wrap up, if you have ensured that you have a nice, well-structured, well-governed, tidy data foundation, then I didn't know about this ontology layer
[37:33] until today. Maybe it can help a little bit later on, but for now I I would say metric views, use that as a very thin semantic layer where you at your descriptions, definitions, nice column names,
[37:48] key calculations, dimensions, and so on. If you have done that, you can really build dashboards in no time. And because you can actually in Databricks chat your way to dashboards, if you the users can also do that it
[38:03] themselves. Personally, I'm more into dragging and dropping. I'm not very good at typing fast, so for me it's faster basically to drag and drop the stuff I want to show into the canvas. But that's a taste thing. It also makes Genie so much more
[38:21] trustworthy. Actually, it turns her into something of use for for chatting with the data for real. Actually, this helps shows about that you can connect data with document document with data. That's a game
[38:38] changer, I think. And overall, the combination of all these things adds the flexibility that allows, you can say, non-technical users without a lot of data insights to to self-service insights into the data.
[38:53] So, that's how, you can say, bringing it all together. That's it. Thank you so much.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.