Robert Bell: 3rd party. Robert Bell: Hi, everyone welcome. We're gonna be starting the webinar in a couple of minutes, just waiting for Robert Bell: few more people to join Robert Bell: thanks again to everyone who's joining us. Now. Robert Bell: My name is Robert Bell. I'm Robert Bell: product enablement specialist in Netapp. Robert Bell: and I'm joined here in this session by Lupa Weisberg, who's our product? Solutions architect who specializes in Robert Bell: Fsx. Amazon, Fsx for Netapp on tap. And Gen. AI. Robert Bell: So thanks a lot for joining the session. It'sgoing to be a session which will be mainly about showing you Robert Bell: a solution, an actual live demonstration. Robert Bell: But before we show the live demonstration. I'll be showing a little bit of a few slides. About the topic. Robert Bell: so we just I think we'll start inless than a minute. Robert Bell: So we have quite a lot of people who've joined. Robert Bell: Okay, I think we can get started. So this,is a session which is focusing on bringing data into Gen. AI applications and the products of solutions we're going to be talking about are Amazon Fsx for netapp ontap and Blue Xp workload factory for aws. So obviously, this is relevant for anyone who is Robert Bell: doing Gen. AI on aws or planning to do Gen. AI on aws. Robert Bell: And this will be mostly a demonstration. Robert Bell: but I'll give a bit of background before this session Robert Bell: is being recorded. So if you have colleagues who you're interested would like to hear this. Then we will share the recording, and you can share this with your colleagues. If you have questions you're welcome to ask us in the Q. And a Robert Bell: window will have people answering that during the session Robert Bell: and if you need have any specific information that you want to, you want us to get back to you about Robert Bell: additional information. Then you can leave your details in the chat, and we will make sure that someone will get back to you with the information. Robert Bell: So I think we'll get started. Robert Bell: So this session really is all around generative AI, and Robert Bell: what we see now it's increasing at such a pace that every single industry is now Robert Bell: finding out ways to use generative AI Robert Bell: for their business purposes. And there's a lot of different Robert Bell: ideas and different applications in different fields. Robert Bell: But what's common is that everyone is trying. So Robert Bell: whether it's in production or just something that is being developed. Robert Bell: What we expect is thatin the near future everyone will be seeing more and more of these Robert Bell: Gen. AI based applications being applied in business people will be relying this more for business applications Robert Bell: and for finding new ways to become more efficient and to expand into additional markets or into additional products. Robert Bell: So Robert Bell: just to sort of understand why we took why Gen. AI ismaking such an impact. If we look back atwhatreally came before Gen. AI. It was Robert Bell: around just machine learning, which is something that has been around for many years. But it didn't really Robert Bell: an impact. Such an impact as Gen. AI has had. And I think that just to understand that Robert Bell: when you talk about machine learning, it was about taking a lot of data. Robert Bell: but structuring it and labeling it in a certain way and training the data for a specific purpose. So you've got data which is trained into a specific machine learning model and then deployed into a specific application. Robert Bell: So this was something that required a lot of, I think, a lot of effort, a lot of expertise, a lot of money, a lot of time. Robert Bell: And it was something that was relatively something that ifobviously, if you invested enough, you could do this, but it would have to be done repeatedly for each type of application. Robert Bell: What's changed in Gen. AI is that you don't really have to Robert Bell: prepare the data. You can actually just take raw data in different formats, unstructured, unlabeled Robert Bell: and train a foundation model on this data. Robert Bell: Once you've got a foundation model that has learned all that information, you can apply it in different ways Robert Bell: for different applications. Robert Bell: So once you soin other words, you don't have to repeat the whole process of Robert Bell: training, and you don't have to really worry about labeling the data. Robert Bell: Now, this has kind of evolved to the fact that companies have developed these Robert Bell: public foundation models which are pre-trained traditionals like Claude Jurassic, Llama. And Robert Bell: these models, you can actually just use them instead of developing your own model. You can use this and just adapt it for a specificapplication. So companies are now using these public models Robert Bell: to make virtual assistants forgenerating code. Robert Bell: But also, if you look on the business side, then it's for creating content, for designing products, for improving processes. Robert Bell: And all of these improving customer experience. All of these are things that can be applied just based on these public models. Robert Bell: But the key here is in order to not to just make a generic application that anyone could make. Robert Bell: The key here is really to make it unique. Robert Bell: And when you make it unique, then it can be something which is profitable. Robert Bell: So the goal here is really to add specific domain Robert Bell: domain specific data into the application to make it unique Robert Bell: and the most efficient. The most popular way of doing this nowadays is called rag. Robert Bell: It's retrieval, augmented generation. So this is a method of getting unique data into a public model or combining Robert Bell: a public foundation model with your own private data to create a generative AI application which is unique, and that bringsspecific value that only you can deliver to your customers. Robert Bell: So thissession is focused onspecifically on how to do rag, because rag isan idea or a concept or an architecture. But then, when it comes to applying it, there can be certain challenges. And what we're going to talk about is a really simple way to get from Robert Bell: 0 to rag in very, few very,easy steps. Robert Bell: We will also demonstrate the results of thisspecific implementation. Robert Bell: And it's obviously not the only way you can do Rag, but it's really something that Robert Bell: helps you bring data effortlessly into these Robert Bell: very,useful foundation models that are available on aws through Amazon bedrock. Robert Bell: So I'll go briefly over this andlike I said, we're going to demonstrate this. So you'll see in context how it would look or what it would kind of what,Robert Bell: kind of results it could bring. And this is the demonstration we're going to do is just a very basic thing, obviously depending on the kind of information that you're putting into your application, you get lots of different results. Robert Bell: So if we look at theidea of rag. Robert Bell: The idea of rag isreally it's based on just Robert Bell: generative. AI is saying, you know, build an application, connect it with foundation models, and if you're using aws, the best way to do that is just connecting your applications to Amazon bedrock, which has the public Robert Bell: foundation models that are available that you can access through your applications. But then, in order to do rag, you need a knowledge base. So you need a place that stores the unique information, the domain specific information that you want to feed into these foundation models. Robert Bell: So the way it works is well, 1st of all, you have to bring your data into the data source. So it has to be nearby Robert Bell: or part of the application nearby. The Aws environment, where you're building the application and connecting to Amazon bedrock. Robert Bell: Then, once the information is, the raw information is in Robert Bell: in a online in Aws where you want it. You still have to embed the data, which is a process that you do with one of the embedding models which is really just turning your data into vectors, which is Robert Bell: in a format that the Robert Bell: large language models or the foundation models can understand. Robert Bell: Once it's been embedded, then you can start using or users, your customers can start using the application you developed. So what they do is they ask a question. They enter a prompt or they do some sort of input query to your application through the ui, and then what happens is instead of just going to the large language model, which is kind of like what you do when you just use chat, gpt, and you ask a question, you go directly to the model, and you get an answer instead of that in rag. Robert Bell: What happens is the AI application looks into the vector database to find relevant data Robert Bell: which is relevant to the specific question. Robert Bell: So if I've asked a question about a specific Robert Bell: technology or a specific product, and that's something that is information that is in inside the data source, that information will be provided and given to Robert Bell: the large language model along with a question. So there'll be a question, you know, how do I solve this problem, but also with the information, the product information which is unique, or the research information that I've collected or whatever data I have there. So I'm asking a question, but also Robert Bell: enhancing this question with a lot of relevant information. Robert Bell: So the large language model which knows how to answer questions and structure, the answer will not only know how to answer the question, but it will also have the expert knowledge that you've fed into it in per question. Robert Bell: So if you have very,valuable data, then that responses that the customer will get Robert Bell: from your application will be expert responses, and they'll be unique to the application that you've developed. So this is really the secret, you know, whether if you're a Robert Bell: a healthcare company that has a medical research, or if you're an engineering company. You have a lot of insights about specific technologies or products or any kind of domain expertise you have. This can be fed into Robert Bell: the information. You're not changing the foundation model. The foundation model is stillbehaves in the same way, but by feeding the relevant Robert Bell: or augmenting the prompts with your specific data. The user is getting better answers. Robert Bell: So now let's just see before we move on to the demo. I just want to show just basically what you would have to do Robert Bell: to build an application like that. Robert Bell: Obviously, there are many ways of doing this, and there are many different tools for building it. What we want to show you here is really something that is, I think, relatively Robert Bell: easy to implement and doesn't require expertise. You don't have to be an AI expert. You don't have to be anaws expert. It's really something that is very,well Robert Bell: built. It's built in a way that is very,easy to apply. Robert Bell: And then, on top of that, you can then customize that to really many different things. I think the main advantage here Robert Bell: a lot of companies want to start Robert Bell: moving with Gen. AI as fast as possible. Robert Bell: so it takes away the complexity of really just becoming an expert on building applications. It gives you a very quick way to have an application up and running and to start testing it with your data, even if you haven't launched it yet. Even just as an experiment, you can build this application Robert Bell: connected to your data, and you can really get immediate results on Robert Bell: what kind of ideas, what kind of unique value you could give. And then, based on that, you can build your applications. So what you would do to build this. Robert Bell: The 1st step is really bringing the files or the information into this, theRobert Bell: knowledge base. Robert Bell: But it starts up by creating data sources. And we're going to show this.solution, based on Amazon Fsx phoneta ontap, which is a storage service provided by Aws which supports Gen. AI. In a very good way. Robert Bell: Now, one of the additional advantages of this is that fs, 6, for ontap Robert Bell: is, it has built in replication tools. So if you have data on premises, or if you have data in a different region in aws, it has this great feature called snap mirror Robert Bell: flex cache which you can use to bring the data very simply into the place you need. Robert Bell: So once that's done, you then have to create the architecture, the infrastructure for the application. Robert Bell: So that is done by a tool called workload factory or Blue Xp workload factory. For aws, it's the longer name Robert Bell: what that does it really automates the whole infrastructure building. So what it would do is it will connect the this Chatbot application to the Amazon bedrock Llms, we'll show exactly how that's done. But really it gives you a way to just choose which kind of foundation models you want, and then it connects all of the infrastructure so that you can start Robert Bell: working quite quickly. Robert Bell: Then, once that Robert Bell: is established, the infrastructure is established. That's the point. When you take the raw data which could be files, file directories could be huge amounts of information. You want to embed that into the vector database so that the user can start asking users can start asking questions. Now, this is a this is where kind of it'sRobert Bell: it can become complicated. But this is solved by workload factory. Robert Bell: What workload factory does. It gives you the option to say how you want to embed the data. This is something we can show you a little bit about, but that really Robert Bell: is customizing the type of Robert Bell: embedding to the type of data that you have. And Luba will mention this later on. Robert Bell: And the second thing is using guardrails, which is also very important thing. Robert Bell: You might be using data which has some sort of private information or personal pii personally identifiable information. Robert Bell: This is something that you can filter or mask from the beginning. Robert Bell: So in order to avoid a situation where private information is actually moved to sent over the network Robert Bell: or outside your network, you can actually mask that information while during the embedding process, this is also a unique advantage of using this solution is that you can make sure that there is no chance of any private information Robert Bell: being sent over the network. Robert Bell: And then, once that's done. Robert Bell: there is a built in chat bot in workload factory, which is something. This is what we're gonna demonstrate here, which Robert Bell: lets you sort of chat with your data. What you're doing is you're asking a question, getting answers both from the Amazon bedrock foundation models and your own data. Robert Bell: So this is what you can do to test Robert Bell: the accuracy of the response. And then, if there's something you want to update or Robert Bell: or fine tune, you can still do that. Robert Bell: so you can reach a point where you see Robert Bell: that you could like see the value of combining your own data with an Amazon bedrock foundation model. Robert Bell: Once you're happy with that, then you connect your Robert Bell: data to the access directory. Now, this is important, because Robert Bell: the files that you're uploading, you know, not everyone has access to these files. Maybe there are different types of users with different permissions. Who don't all have the same access. So by connecting worker factory connects the access directory to the data so that Robert Bell: a user who doesn't have permission to access information won't be able to get responses based on that information. Another import. That's another important security feature that comes built in this solution. Robert Bell: Then, once that's done, you caneither use the Robert Bell: this chatbot, or you just connect an external chatbot. This is something that is also available. Robert Bell: We can provide the access to an external chatbot you can use or develop your own. The idea is that the Chatbot itself just has to Robert Bell: have the user interface and the ability to ask questions. Everything else. The whole rag infrastructure here is built by workload factory. Robert Bell: So this is sort of an overview of what you would have to do. And it's really something that if Robert Bell: you don't need to be an expert, and it really can literally be built Robert Bell: ina matter of minutes. Robert Bell: Now, once that's established, then you can start at Robert Bell: using theChatbot, and at this point I'm going to pass Robert Bell: over to Luba, who will show us a demonstration both of the Chatbot how you would use the chat, but also some of the features that I mentioned here. Luba Vaisberg: Excellent. Thank you very much, Robert, and I just want to invite everybody who is interested or have a question. Use our Q. And a. Or a chat options. Luba Vaisberg: if you want to get an answer on, or something isn't clear or more information. Luba Vaisberg: right? So hopefully, you should be able to see my screen now. And just as Robert said, I'm going to demo Luba Vaisberg: and show you how it is. How simple it is really to create your own application. And with your own data. Luba Vaisberg: And I have 2 examples for you. I'll start with example that will cover alright. This will cover Luba Vaisberg: what is rug, right? And show you how Luba Vaisberg: you can utilize your own information for that. So to do that, I'm going to use a file that I've downloaded for the FDA and covers the treatment for Covid called the mov. Luba Vaisberg: So before I can use it. What a person need to do is create a knowledge base. Luba Vaisberg: and you'll see how simple it is. I'll call my knowledge base Ls. For a life sciences. Luba Vaisberg: I can provide the description. Luba Vaisberg: and the next part for me is to choose the embedding model I want to use in the chat model. Now within worklot factory, Gen. AI. Or blue Xp workload factory. Gen. AI, we actually allow you multiple places and options to define how you want to work and with what you want to work when you work with a Gen. AI, Robert mentioned, there's many options there. And Luba Vaisberg: you can choose the relevant options for your data. Luba Vaisberg: So with my example, I'm going to use the default options here for the embedding model. I'm going to go with the Titan model and for the chat model. I'm going to choose the cloud. The 3.5 Luba Vaisberg: part is just choosing the Fsx on which I want to build my vector database Luba Vaisberg: where I will save both the vectors. Luba Vaisberg: And let's call it this role in this case. Luba Vaisberg: And the Luba Vaisberg: I'm also going to select a snapshot policy. And this is for a the option to recover if there's any issues. Luba Vaisberg: Now, this process can take, as you see after a minute. And what's happening in the background is that the system itself? Right. There's nocoding required on your end will create a new volume, and we'll install on this volume a new vector, database Luba Vaisberg: that will later be used to store. The vectors, including metadata of the file. Metadata, can include things like when the file was covered. Luba Vaisberg: scanned, who has access to the file, and much more. Luba Vaisberg: and the when it will be done, I want to show you the following. So Luba Vaisberg: I'm going to start with a fresh knowledge base. And I will not currently add any of my data source, and I go directly to the knowledge base itself. Luba Vaisberg: So you'll be able to see that I have here my new knowledge base. And as you saw, it took moments to create. Luba Vaisberg: And right now it's in public mode. This means that when I'll ask a question. The question will go straight to the cloud, because Cloud is the Luba Vaisberg: the chat model that I've chosen, and the question will be answered, based on the information Claude has. Luba Vaisberg: So if I'll ask what is mov. Luba Vaisberg: and I want to remind you that Mov is the treatment for Covid from my document. But since I did not enter a document. Luba Vaisberg: let's see what Claude will say. Luba Vaisberg: Obviously Claude does not have access to my document. Luba Vaisberg: The answer I'm getting is that mov is a file format Luba Vaisberg: that used primarily for video. Now, it's absolutely correct. Luba Vaisberg: The problem is that it's not really relevant for Luba Vaisberg: what I'm looking for. Right? It does not have my data. Luba Vaisberg: So let's just add my own data to do that very simply. I'm going to click here on edit Luba Vaisberg: a data source, I will be able to choose the Fsx. Where my data resigns Luba Vaisberg: in this case. My data is hosted here in this specific volume. Luba Vaisberg: I'm going to add some information for my active directory. Luba Vaisberg: And Luba Vaisberg: we'll just give up. Luba Vaisberg: Yeah, whatever. Just a sample of the IP, Luba Vaisberg: the next case, and this option where I can choose, I can select whether I want to look at data in the entire volume. Luba Vaisberg: or I can actually specify a specific folder. Again, it allows you to. define and Luba Vaisberg: see exactly the data that you want to use for your system. Let's just check this particular, a directory. Luba Vaisberg: and the next part. And this is where Ithink a lot of the options come to Luba Vaisberg: and are available for you. Here is the chunking strategy now, like Robert mentioned earlier, chunking strategy is very important. When you look at your data and decide how to treat it. Each file have different structure. There's difference between files that maybe are literature right? Or files that contain technical information Luba Vaisberg: or files that are just a emails and much more Luba Vaisberg: choosing. The right chunky strategy is crucial to get a good result, you can think of it as taking a picture. For example, let's say you have a picture that has a forest and the sky there. Luba Vaisberg: and you want to ask a question, let's say Luba Vaisberg: how many clouds there are in the sky. Luba Vaisberg: You want to make sure that when you send the data Luba Vaisberg: to the relevant model, you only send the parts that contains the skies right? There's no point of sending parts that contains the trees, obviously, because there's likely no clouds there. Luba Vaisberg: But you also don't want to send too small of a chunk. Luba Vaisberg: Such a small chunk where you lose the structure of a cloud. Luba Vaisberg: That's why, say, choosing a relevant, chunking strategy is mandatory is very important. Luba Vaisberg: It can also be relevant when it comes to different Luba Vaisberg: languages right? Not. Every language builds in a Luba Vaisberg: what we maybe look at standard sentence model. Luba Vaisberg: And this why we allow you here the option to choose between 2 different chanting strategies. Luba Vaisberg: One is sentence based, which means the data will be split into parts based on number of sentences. Luba Vaisberg: or the other option is the chunking strategy by characters where the data is split by characters, and here we also have the additional options of overlapping your data. Luba Vaisberg: So I'll go because my file is a simple text file. I'll go for the chunky strategy of sentences. Luba Vaisberg: And what you see here as well, is the permission aware? Option, and I leave it as active. This means that when I'm asked a question Luba Vaisberg: a the AI agent so been for the Luba Vaisberg: the heart of the solution will 1st check whether I have access to the relevant files right? And it will only generate an answer from the file that I have access to. Luba Vaisberg: This means that, let's say, even if Robert and myself both will ask the same question on the same knowledge base. Luba Vaisberg: It's quite possible that we'll get different answers because we have different roles. We belong to different teams, which means, obviously, we have different Luba Vaisberg: permissions. Luba Vaisberg: And one of the things that I personally really like about this solution is when you're using Fsx for ontap. Luba Vaisberg: you already have permissions configured Luba Vaisberg: on the data itself on the Fsx. For on top itself, which means you don't need to worry about those again here. When it comes to this AI application. This part is already done for you. Luba Vaisberg: While I was speaking you can see that the data has been embedded. Luba Vaisberg: and if I'll go and ask exactly the same question, a. Luba Vaisberg: What is MOB. Luba Vaisberg: And send it? You can see, by the way, that my data set is now on organizational mode. Luba Vaisberg: which means, when generating an answer. The system will 1st look at the data that is connected to that. And you can see I have a different answerbased on my file where it says mov is an oral drug used to treatment. Luba Vaisberg: And I also have here a citation which means I don't just get information on my question, but also what file or files were used Luba Vaisberg: to generate this answer. Luba Vaisberg: The other thing I want to show you is Luba Vaisberg: a new feature that we've added about 2 versions ago. Luba Vaisberg: And this is your ability when you Luba Vaisberg: chat with your data, to actually remove Luba Vaisberg: and reduct any pis of your information Luba Vaisberg: to use that we use another service from a netapp called classification. Luba Vaisberg: And it's also available for you. Here in Blue Xp. Luba Vaisberg: You have your internal governance. Luba Vaisberg: You have the classification. Luba Vaisberg: and if you already have it and are using that, you can actually use both tools together. Luba Vaisberg: So let me show you an example. Luba Vaisberg: Just to save some time, I already recreated the knowledge bases. Luba Vaisberg: But both knowledge, base road data and knowledge, base guardrails, data have the same data. In this case they have a file that contains some credit card statement. Luba Vaisberg: So if I'll go here, and this is the where I did not. Luba Vaisberg: initiate a connectivity with the classification model, which means Luba Vaisberg: my all the piis will still be available. If I'll ask a question about credit card number and Iban in the document. Luba Vaisberg: I will get an answer containing both information. So you can see here I have both credit card number and Iban. Luba Vaisberg: which is obviously problematic, right? Because I might not. Luba Vaisberg: You know, this is a private data. We don't want it to be accessible tousers. Luba Vaisberg: So I want to show you the behavior of a system Luba Vaisberg: where I have the same file. But I've actually Luba Vaisberg: enabled the connectivity and the connection to the classification service. Luba Vaisberg: and if I'll ask exactly the same question. Is there a credit card number? And Iban in the document? Luba Vaisberg: You will see that the answer is yes. Luba Vaisberg: the document mentions both, but they have been removed. Luba Vaisberg: and all I needed to do to enable this Luba Vaisberg: is just go here to the data, guardrail and choose, enable. Luba Vaisberg: or obviously disable. If I do not need it Luba Vaisberg: so very simple thing to do. Luba Vaisberg: And I think that's it from my presentation and my demo. Robert Bell: Okay, so we'll hang on here in case there are questions. I'm also sharing a screen Robert Bell: that shows some resources that you can refer to. Robert Bell: The top. The top linkat the top is theway to access Robert Bell: workflow factory. But there are also some Robert Bell: links here that you are welcome to Snapshot and Robert Bell: follow these formore information about the solution. Robert Bell: As we mentioned before, you're welcome towrite to us in the chat. If you want us to provide you with more specific information. Robert Bell: And this is workload factory is really it's a free tool, something that you know, if you're using Amazon bedrock, and you're using Fsx for ontap, it's really so easy just to go in and try it out Robert Bell: because it's it doesn't cost anything. You can really try out this Gen. AI application Robert Bell: pretty simply. Just to get an impression of what kind of things you could do with it. Robert Bell: So we'll hang on here in case there are any additional questions. Luba Vaisberg: I will see you. Yeah, thanks, Gary. Luba Vaisberg: to mention that some things that we don't have in this demo today but are available if you're interested in is even external applications that you can use for your external users. BothRobert and I pointed in. Today's session is more about how you manage that and how you administrate this solution. But obviously, you know, if that's something you want to give access to your Luba Vaisberg: end users, right? Could be within the organization could be outside of the organization. We actually at net up here, created Luba Vaisberg: a code that will generate for you an application. What Luba Vaisberg: you can refer it to. External application Luba Vaisberg: will securely connect to the data here. Luba Vaisberg: and will allow you to give access to your end users to this knowledge base. So they'll be able to ask questions, get answers without having to obviously go through this administrator process. Luba Vaisberg: This is also available and just reach out if that's something that is needed. Luba Vaisberg: Okay, excellent question in the chat in the Q. And a, could you specify data that needs to remain private not only credit card number, but social security also a amount. Great question. Luba Vaisberg: So we have quite a long list of data that is private is based on different things, like a Jdpr hipaa. And more actually, we cover things from a driver license. Luba Vaisberg: what I saw here, like credit card numbers. Say I, ban Luba Vaisberg: medical information, and much more. Maybe I'll share a link here, for the list of all the Luba Vaisberg: the data that is considered private, so you can look at it and see. Luba Vaisberg: But I can tell you that the classification service within Luba Vaisberg: a Netapp is quite let's say, a comprehensive Luba Vaisberg: and cover a lot of options when it comes to private data, right? That private data is let's say, it's something that keeps growing on right. What it does it mean? Luba Vaisberg: Private data and obviously, structure of private data are different in different countries. Luba Vaisberg: things like social security number in the State looks different and very different in India, obviously. Luba Vaisberg: or in I don't know Luba Vaisberg: in France, or something like that. So we have quite a lot of options to do that. Luba Vaisberg: And I'll just say, going to Luba Vaisberg: paste the link in the Q. And A, yeah. Robert Bell: Mostly, it's mostly around personally, identifiable information. So any information that Robert Bell: that could reveal some personal information of Robert Bell: a customer of a user ismasked. So it's Robert Bell: essentially something ready tocomply with the regulations like Gdpr or Ccpa and other and hipaa other Robert Bell: regulations for protection of private information. Robert Bell: So this is something that is, I think. Robert Bell: very efficient. If even if you're not sure whether there is private information, if there's a lot of data. You just want to apply these guardrails just to make sure as an extra security measure. Robert Bell: But yeah, definitely, something thatgives you more flexibility on the type of information that you can use in any case with the solution. The data doesn't leave your virtual private cloud. Robert Bell: It all stays within. It's stored within the Vpc, so it's not actually the data is not stored on some external application. Robert Bell: But merely when the information is being sent for, the prompt. If you want to just make sure that there's no pro personal information being sent, that's something that Robert Bell: I think that the safest way to do that is in advance while the data is being embedded, so that there's no option for this information to leak out in any way. Robert Bell: So there's a lot of features that we didn't cover here, that obviously this is a very extensive solution that has it's based onRobert Bell: ontap at the core of the storage and ontap is really one of the most advanced storage systems available today. It's got so many features. So these are also included. But I think the main idea here is not so much about all the things you will be able to do. In addition, it's more about the fact that it's just so easy to get from 0 to working Robert Bell: Gen. AI application based on your own domain information which could be, it could even be on premises. And you're creating a mirrored copy in aws. So it really gives a much shorter time to market with, or even time to just do, a pilot or proof of concept to really start working with Gen. AI faster than your competitors, which is the goal. Robert Bell: So I see there's not any more questions right now, so I'd like to thank you all for your participation. We'll be sending a follow-up mail, so if you have any questions or anything, you want to get back to us and ask more, then please feel free to reach out. Robert Bell: So again, thank you for your time, and I hope this information was useful for you. Luba Vaisberg: You everybody.