Sri Desikan, Elastic & Vrashank Jain, Dell | Dell AI Data Platform Event

Hi, I'm John Furry, host of the cube. Here in our cube, Palo Alto studios. Of course, we have our New York Stock Exchange Cube studios connecting Silicon Valley to Wall Street. We're here for the Dell AI data platform with Elastic program. Mashank Jane is here, director of product at Dell and Shri Desakhan, VP of product management, elastic. This is the Dell AI data platform with Elastic. Welcome to the the program. Thanks for coming on. >> Thank you, John. Thanks. Great to be here. So, you know, data is the hottest thing on the planet as you know. You're in both in the data business, but AI's brought in a whole another user behavior paradigm shift, but the underlying data structures, they're in there. You have search, which you guys have pioneered at Elastic. We've covered that, you know, at at a long time. But now there's a search paradigm, there's a discovery paradigm that agents and AI take. It's not as simply same as getting that first answer fastest. There's a lot of reasoning and intelligence now in the data discovery. So I guess the first question is what's the biggest change in the enterprise that companies are seeing the moving from a search paradigm to an AI native and agentic because the same problem I still want to get stuff fast. [laughter] What's the biggest change for enterprises? >> Yeah, great question. I I think I would start with the fact that um with AI and agent intake workflows you can now operate on both structured and unstructured data right that's the biggest thing and so what does that really mean it means an enterprise which has traditionally used a database and search separately now has a unified data platform where they can look at both structured unstructured data both used by humans as well as by agents and so the underlying shapes of the data now will have to evolve. Um so you can not only do full text search on it but you can also do hybrid search semantic search all as part of one single engine in in one single database right and that's what elastic brings to the table >> but does search really change three because it's still searching >> it's just there's other things going on is that a new dynamics models behave differently they're still doing search you got to find things >> exactly I think there are many dimensions to this so if you look at human search you search you get a list of answers back like if we do Google searches set up blue links right and and the human has the context to decide which of those blue links is appropriate agents don't have that context so the search has to evolve not only to retrieve data and have that list but also rank that list based on signals that it needs and so search has grown from a I would say a human curated effort to an agent curated effort and it requires a mechanism behind the scenes and a lot of what Elastic has built over the decades is now really appropriate and fits in really well into the agentic workflows. Things like vector search and hybrid search and uh re-ranking and concepts like that are super important as part of the agentic workflow. >> We're talk about the Dell piece because there's a lot going on too as the surface area is more horizontal. You still got domain specific intelligence with AI, >> right? You still want contextual greatness. >> Yes. >> And but it's behaving. What's your what's your perspective? >> Yeah. I I think um maybe the the the one phrase I would use is we're shifting from a really predictable way to search things to a really unpredictable way of reasoning over loops essentially is what's going on. So as agents start to basically loops through the question and try to answer the the question in multiple ways, it's going to try and reach out to as much data as it can find, which means we cannot be satisfied with a predictable answer to a question. Like we're going to have to be able to prepare much more data than before. Which also means we know we now have to shift our focus to not just the the documents, the images, the tables that we knew were really good quality, but we have to now start expanding to the stuff that's lying around in our enterprise, but we know it's just really hard to wrangle, but it holds tremendous amount of value. which means are because the agents are unpredictable in which data they'll go go after at any given point. We'll have to make sure that we can prepare a lot more data >> that they can you know uh cycle through. You know, the thing that I see and I want to get into is how this changes some of the architecture to be AI ready. But yeah, one change I would like to get your guys' reaction to is in the old old days a couple years ago, you have a data platform semantic layer, harmonization layer, we've seen that, but with the models coming in and you're seeing general intelligence of the front frontier models, now specialized intelligence in these more domain specific, >> you're seeing that come in and then what what we're seeing is enterprises saying, hey, I can split the models out from the data platform, right? and then figure out this new context layer >> that's new. >> How does that change? Because now does that change pipelining and prep and because this do you first is that h do you see that happening and what's the impact? We we do see it happening although I wouldn't call it a necessarily a change. I would say >> evolution >> the evolution of exactly the way that we've defined context layers like in the past we said context layers are essentially a semantic layer which is a bunch of definitions with a knowledge graph that was really based on our human capabilities of being able to connect data sets which arguably were very limited. Now what we're doing is we're pointing models and domain specific models to our data sets to say hey you go figure out all the semantic information that's buried in here and obviously it generates 10 times more than we had before. The graph is now 10 times more accurate than it was before. So I think the context layer has just evolved into something much more relevant because it's being fed by generative AI rather than having humans curate this on an ongoing basis. >> Yeah, I think that's a great point that Brashank brought up. We um we've been working on context engineering and having products and solutions and as part of what we call as context engineering is uh there's data in an enterprise scattered in many different applications right traditionally you have to run ETL store it somewhere and >> humans have to curate it and you build materialized views for those who have used databases >> so what's happening now if you want to put agents against five different backend data sources uh and you're just relying on MCP and tools and the traditional way a model does it. There's a lot of token costs. You know, it's um that's the biggest uh concern for a lot of enterprises today. >> And secondly, what are you getting after spending all those tokens, right? Is it accurate? Are you able to move from pilot to production? These are the challenges that we see every enterprise facing. So our approach to that has been um like Rashank mentioned, how can you pre-build this context in a way that minimizes token costs and maximizes accuracy. You're not going to be as deterministic as you know just a query on a database because that's not the way generative AI works. But any answer to a question should be accurate, factual, and to the extent you possible be consistent right from person to person. So this notion of a context layer pre-building the context uh is becoming more and more important and that's an investment that you know >> that's like the connective tissue between the models whatever they may be. >> Yeah. >> And the data underlying data >> right >> and I like that um you didn't use the word curated but you said orchestrated. I don't know what word you used but you said accuracy. A lot of people blow in tokens a lot of tokens and it's not accurate >> right. Why does that happen? Yeah, I mean traditionally what has happened is um you know when you ask a question of a model there's a context window right and the model uses that context window to understand what the intent of the question is now over time this context window has grown larger >> and the general thesis has been the longer the context window is the easier the model can reason but it's not turned out to be true >> so uh because as the context window gets larger think about humans right how much can we keep in memory. How can we uh you know it's called context rot and you know uh things like that >> your brain gets full >> your brain gets full pass out >> and um I'm sure the models are improving but at the same time the goal of this context layer is to feed the context at the the right context at the right time. That's the way to think. >> Okay. So why is data prep a make or break piece of getting it searchable and AI ready? >> What's what's why is that why is that such a now more important than it was before? Yeah, I I think it goes back to the the point that we're now starting to uh looking to ingest much more data into the context than before. Now, it turns out that I think most enterprises have already tapped the easiest data sources that were available. Those have tended to be, you know, uh blog posts that we know are true to be online. We know that our warehouse already has good data because we've spent so much time in data engineering making it really good. But we're not satisfied with that. Which means we now have to start going after our sharepoints, our one drives, our our legacy data sources, our SAS applications that are generating more data, but we're not going to wait until it actually makes its way into the warehouse. So the more we're try to go further and further deeper into the enterprise, the more we're going to face data that is not ready, it's not structured, it's not tagged, it's not labeled, and it's not transformed. Which means the the problem that has always existed which is we haven't done enough data prep is now such is now a 10 times bigger problem because agents are hungry for that right context but that PDF is actually not searchable because it's all handwritten from 1985. >> Um so as Rashant mentioned data prep is all the more important in an enterprise in the age of agents because there is structured data and there's unstructured data and context exists between these two. Right? So if you take an example of a salesperson wanting to answer a question on uh how is this account you know what can I do for expansion in this account and it the history is there in documents the actual data is in Salesforce >> and there's previous interactions from the other sales reps asking the same question as part of agentic traces >> all this has to be prepped in a certain context layer right so I think data prep becomes a different uh beast with with respect to AI now >> and and why is that important Because you know we the fear is AI needs data the better the data the better the AI that's a well understood concept but if they can't get it >> and some security practice will say hey you know what I want to feed the AI data set that I want it to run with right >> so they essentially pre-prep data sets >> is that kind of conflicting or is that just situational >> I think that's a race you can never win meaning you can never get to a point where you feel like you have now prepped enough data for this agent that will never happen. There's no such point. Which means you kind of have to run continuous pipelines on curating new data, old data, and you keep have to push the frontier to say, can I also now go after a data set that I couldn't go after a year ago because I didn't have the right model. Well, now the models are getting better and better, right? Like 2 years ago, it was impossible to start analyzing hours and hours of video footage. Well, now it can. >> Yeah, cuz the GPUs have gotten better. The vision models have gotten better. So, what's what's going to be next? It's it's we're just going to keep moving further and further down the line >> and they have to multi-step reasoning kind of implies that it's the agents are learning. So it's like it needs more data. >> Exactly. >> Not going to just recycle stale data. All right. [clears throat] So how do you guys bring it together because elastic and Dell as a partner you guys have a partnership >> the world doesn't want another point solution integrated and co-designed systems. You mentioned MCP that's great. It's kind of like where APIs we saw greatness come from APIs. Now you got APIs and MCP. there's still a lot of tooling around it. So those things are just you know interoperability points but the software coming in the market has to be integrated into these systems which is crossing disparit boundaries different data platforms so you have a lot more >> crosspollination I don't know what word you call it but how are you guys bringing that together what's the what's the solution >> yeah I I can go first and and you can definitely add to it um uh I think we're trying to solve the the problem for both IT and data That's the first thing we need to understand because we have a lot of solutions in the market. The data market is frothing with companies and a lot of them solve really point solutions in a in a very grand sort of design of the data platform. And the first thing we wanted to do was bring the best of breed capabilities into a singular platform that a that organizations go to one place to buy, to deploy, to support, to upgrade, to manage. We wanted to make sure that the the components are in that platform are best to breed. meaning it's the best vector database and search database that we could find in the market. the the best SQL engine we could find in the market uh which is starburst the the um you know and bringing uh the best unstructured data management capabilities but when we built that and our own IT team said this to ourselves which was don't give it to me as a bunch of point solution that now I have to manage as an IT team we're like every IT team is a cost center >> um they're all getting pressured to reduce the spend that they have while managing a larger more complicated environment Which means we wanted to give IT teams an experience to have a data platform that's easy to manage, deploy, scale, upgrade and it lives within their four walls because frankly the data we're trying to go after is the usually the most restricted stuff. >> Yeah. And I would say that the the IT is now infused in every department. So the opportunity for it is the B >> revenue contribution. Yeah. They could actually contribute to revenue. Yeah. So they're viewed as hey let's get our cost reined in. But the opportunity with AI is >> Exactly. >> They're now enabling. They're the liberators. >> Yeah. Absolutely. >> And the data is the key. >> Yeah. >> The data is the key. And from our part, you know, we're we're, you know, Dell is a great partner. We bring the software side of it, right? And uh when we think about, you know, what Elastic brings to the table in Dell, there's a lot of customers who need airgapped environments, compliance, you know, but this market is pretty big. >> Think of banks and insurance and so on. So for these customers who are who have a lot of you know compliance requirements this provides a package solution right. So we have elastic as a software. We also have Gina models which are first party models for taking unstructured data and creating embeddings. We have reancher models. We have the ability to parse unstructured data into text. Getting all of this in a single packet solution is a huge advantage because otherwise they have to stitch together so many different options. >> Right. So that's a big ROI. >> Let's make this real, put it into action. Can you share some use cases that are in production because we're starting to see >> pilot to production migration or not progressions, not migrations but progressions. Certain a lot of migrations going on in the data world which is a good thing. But yeah, you know, you're starting to see an acceleration because of the coding >> open the door up for and ages. So coding too. So starting to rain in the governance, it's starting to rain in security, right? As as a first principle. >> Yeah. >> Now production's coming fast. What use cases can you can show the elastic Dell combination? >> Yeah, I I I can give a few, but um I'll focus in on on one that's that's uh sort of super recent. Uh this is a a large um multinational bank, and they're looking at about 25 million contract documents uh spanning years or decades rather. um and their organization obviously wants to be able to search through that, but uh get those results both fast and to a high degree of accuracy. Now, this would have been easy if it was just a bunch of documents created in the last 5 years that were all digital, perfectly, you know, organized, etc. But we're talking decades, which means we're looking at documents that are complex. They include handwriting, they include charts, they include uh incomplete documents. There's also no sense of what is a good document, what is not a good document, what should even be in the repo and what shouldn't. And so the the Dell elastic sort of joint solution here works on both sides of the pipeline. Meaning on the left hand side of the pipeline is the pre-preparation of figuring out which PDFs or which documents to parse, how to parse them. Having a a multi-threaded pipeline to say some documents deserve heavier treatment because we need to have a model really scan it. Other PDFs are super easy because they're already you know uh uh in a format that you can create a markdown and you're that's the sort of highway fast lane but eventually all of this data fed you know gets fed into embedding model the vector databases is now 50 billion vectors plus this is a very very large use case and the goal is to hit uh search latencies that are subsecond at a 95 plus accuracy. So this is a infrastructure data prep vector database model problem all combined into one. >> Yeah. From a software perspective, it's a great fit for elastic, right? >> Um and we we earlier we talked about how unstructured data is changing the landscape of search. Uh because the structured data is still there. So when you as a customer of a bank, you're asking the question, you know, I you know, I want to know more about my account details or a particular loan servicing etc. you know the search has to know your your specific account number which is a full text search or or your name but it also needs to know the the NLP English language translation or whatever language you may be asking the question in >> related docs and bring that whether it's from PDFs or other sources right and so that's the part that's not easy and that elastic has solved over the years we can call it hybrid search semantic search whatever uh but the end output like in case this large bank they required recall and precision which are the technical terms the the goal is to say you know I need to I need the search to match all the documents that are possibly uh irrelevant relevant but it also needs to show me the right you the precision which is are these the right documents once you know it's able to search so I think the this combination is what elastic calls is the agentic search and and and that's what we bring to the table in this case >> yeah I mean search and AI discovery might word. I don't know what else to call it because search is search >> but now you have this new AI discovery and so you see a lot of demos out there. Yeah. >> So I guess the question is you're smiley production grade is different than giving a good demo. >> Yes. >> As it moves into production you got to have the discovery low latency. You got to feed the engines. Yeah. >> Feed the math. Feed the GPUs and CPUs and XPUs. >> What is the difference between a fancy demo? How can you squint through? How can a customer saying okay I like that's real that's not what's the tell and how do agents change all that? >> Yeah I think the there a couple of tells I think one is if a demo is showing you a really good result look at the source data how complex was it to begin with usually demos work on really well curated data sets that are easy to parse uh the minute you throw a a document that just doesn't behave in in the same way uh demos tend to fail. So like one of the big pieces to go into production is the effort that you again you have to put into the the pre-prep side. Uh how well are you are you processing the data um through your storage through data pipelines and using models. So that's that's the part that we tend to focus on with our orchestration engine. >> Um the other part that I think uh turns from a demo to production is >> can this keep up when new documents are coming because this isn't a oneanddone situation. >> Yeah. that even the bank that I was talking about, they still have 5,000 documents coming in every day. >> And that changes the context. >> That changes the context. >> Context, the volume, velocity, >> it changes the the I'm going to get technical, but it changes the HNSW graph that you actually have to create. And every new vector you put in has to recreate the graph. This is a very uh computationally intensive process. And so being able to demo it on, you know, a really large CPU is awesome, but being able to then run it at scale with GPUs, with Nvidia, KBS, etc. all that packaged in is what you know uh makes the difference between a really nice demo and the big boy of a production. >> Yeah. And agents are going to change that. So I guess my question is with agents >> and I'll just throw another another category uh dependency is these AI factories and Dell you you guys have performance has been off the charts since we started covering a factories three years ago. >> You have the horsepower now. You have the compute and the combination of these large scale systems is a dream scenario for software. >> Yep. >> And computing all these things were once very difficult. You need HPC computers. >> Oh yeah. >> I mean Boeing would design a wing with them and take them months to figure out how do calculations. We've seen those demos. Now it's like in milliseconds. >> So that you have massively high performance and you got agents who could code who could >> right find their way around the data. So how does all that change the production landscape? What would you say is the biggest sea change or transition or threshold we have to cross or cross? I mean just um you know uh Rashank mentioned all the challenges that people face when moving from pilot to production and scale performance all that when you when we're talking about agents right a simple agent is one turn you get an answer you know you can wipe code it and and that's that's one way but then in real production scenarios an agent is multi-turn meaning you're having a conversation with it it's probably calling five different tools it's getting context and so on. >> So think about an error in one step and it compounding over multiple steps >> and then you're soon the trust is lost, right? It's no longer production ready. So I think I think this is where um every agent is at the end of the day it's searching for something, >> right? >> And that search has to have the right recall and precision >> so that the error doesn't compound in production as the as Rajank mentioned there's so much data sources. >> You know you're going to have billions of documents. Yeah, it's not as easy and there's a lot more going on. It's not It's the garbage in, garbage out. Clean data is always good. Um scale is a huge issue, >> right? >> Um you guys have an event on October 6th >> where it's going to go into deep dives with practitioners and experts. Yeah. >> You know, >> there's a lot of interest in okay, if there is a compounding negative, which just happens when there's a positive benefit, you got to trace it. Observability comes in. So, >> there's a lot going on that now is out there. There you mentioned graphs. Graphs are now a nice piece of it. You have low latency, full text search, unstructured data, >> latency, lineage, traceability. >> I wouldn't even take out SQL. I mean, I know it's old school, but it's it's irresistible for agents to be able to combine the answer from a document and actually fact check it in in a data warehouse and say, "Oh, I wish I could just send a SQL statement down to a query engine and I would want to combine those two together." Like that's >> I mean SQL is a great use case because like to do a SQL query you actually have to form the business logic first then type it in. Exactly. Now you just say I just want an outcome and it does the business logic. >> But it's going to happen all the time behind the scenes. So >> yeah, we're seeing like SQL query demand actually skyrocket because of this. >> Like you know humans actually tended to really consciously think about the best query to write. Agents don't do that. agents write a query, look at the answer, they write the query again, they get the another answer. They keep writing this until they're really happy about the answer, which means that they're finding 10 times more queries than before. >> What's great about your businesses and your partnership is that all the stuff, the mechanisms that are all good for data are going to stay good. There's more good stuff coming. All the mechanisms that we use to kind of interact with it, interrogate it, wrangle it are going to get automated, >> right? >> And that's going to go away. It doesn't it doesn't go away for the use. goes with the user but it stays in place and is managed by an intelligence layer. >> Yeah, I think it's turning programmatic meaning it's it's not people writing the right SQL query that you can then imitate. It's giving them the reins with the context layer and the ontology and saying here you go and >> okay so you agree with what I just said then that's a good thing. >> Yeah. We we have this agent builder product for example for customers to build their own agents and that's generates ESQL which is our SQL >> version which is a pipe language which is very well suited for agentic workflows right so you can pipe the output of one to another and so on >> so yeah we we are seeing this automated adoption every tool call could be a SQL query ESQL query and so on right so absolutely >> that old generation or modern legacy or whatever you want to call it is abstracted away from the user everybody wants that answer. >> Correct. >> That's kind of where this AI plat data platforms are going or has to go because >> yeah, >> that's what the users want. >> Well, the I guess we don't have a traditional users anymore. The >> agents are the users. So, yeah, I think that's the advice we're giving to organizations. >> Customer satisfaction score. Soon they're going to be ranking the product. >> Agent satisfaction [laughter] satisfaction. We're going to work on that. >> Guys, thanks so much. Uh shout out for the October 6th event. All the practitioners going to be here. data AI platforms are really really important. As the AI infrastructure continues to grow, this next innovation is going to spawn massive innovation around new applications, new agents, fleets of agents, humans and agents working together. They're going to be coding. It's going to be governed and secured, of course, effective. Gentlemen, thank you for coming on the cube for this for this event. Thanks for having us. Okay, I'm John Furrier for the Dell data AI platform with Elastic. Thanks for watching. >> [music]