#127 Clemens Mewald: Redefining the Boundaries of Artificial Intelligence & GPT-4

0:00One of the key aspects of our platform that goes beyond just language and especially large language models is that it's already multimodal. If you get a document, sometimes you want to extract things like a logo or a signature or a stamp or like a figure. So we also run convolutional neural networks, just basic object detection models to extract those images. And our goal is, and our ambition, of course, with the iHub is that we will cover many more modalities as well. So we're really shifting towards content in general. Any text document, any images, audio, video is all per game for us in terms of like being able to capture in the future.

0:38Hi, I'm Craig Smith, and this is Eye on AI. This week, I speak with Clemens Maywald, an expert in AI and enterprise software, about solving unstructured data challenges using AI. His company, Instabase, enables users to engage in interactive conversations with various types of content, such as documents and spreadsheets. And he introduces us to the AI Hub, a marketplace where developers and users can contribute models and applications. I hope you enjoy the conversation as much as I did. Now here's Clemens. Thanks for having me. I guess the very short version of my background is I'm originally from Austria.

1:29I went to school there. and got engineering degrees there then went to MIT Sloan to get an MBA and that's what led me to Silicon Valley so right after Sloan I joined the Google Brain team and I guess my claim to fame there was that I was the first product manager on TensorFlow so I joined in 2015 just when TensorFlow came out and when it was open sourced and then over the next four years I helped build Google's modern AI infrastructure build TensorFlow build TensorFlow extended it uh basically how Google is doing AI today uh then after this I joined a company called Databricks uh to help build the data science and making businesses there um of course you know Databricks started with spark and I helped um build a lot of the product portfolio around AI platforms and then uh late last year I switched to a company called Instabase uh which is like yet another category more focused on unstructured data, actually, which I guess is the big shift.

2:26Because, of course, both, well, Google mostly, but of course, at Database, most of the data, of course, that people operate on is structured. Yeah. Yeah. Tell me about Instabase. And you've launched something called the AI Hub. Is that right? Maybe you can talk about that as well. uh but first about instabase it wasn't clear to me you're you're you're sort of grouped among the uh ocr crowd uh is is that what you do or or yeah give me an introduction to and maybe a little history of instabase yeah makes sense yeah ocr is a very dusty word by the way um so uh but yeah i I guess also like a very brief history also how I ended up here, I guess.

3:15I actually first met the CEO Anand in 2019 through a mutual connection, actually someone at Index Ventures. And she knew that I was interested in enterprise software and just applications for enterprise in general. And she said, you should look at this company. It's very interesting. And then I had this mind meld with Anand, the CEO back then, because my vision for how data-driven applications should work in the enterprise was perfectly aligned with his. but I had just started at Databricks as it was just bad timing but fast forward three and a half years like here I am and I guess the story of Instabase is interesting because it was conceived of as a platform for enterprises to run data-driven applications and that's a very broad statement right and it's a very broad product to try to sell and product market fit was found initially in what you would call document processing like intelligent document processing because it it turns out actually one of the biggest problems in these big enterprises is actually unstructured data.

4:16There's a lot of attention with the snowflakes and the databases of the world on structured data. So anything that you would fit into a spreadsheet or a table, but it turns out that anecdotally speaking and there's also like some research that about 80 % of the data is unstructured. So that's documents, audio, video, like anything that's lying around that's not in a database. And that turns out to be a very big pinpoint for big enterprises. So Instabase started getting into using AI to process unstructured data, most importantly, documents, to solve very high value business processes, mostly for financial services, insurance and big retail companies.

4:54just to give you a couple of examples uh you know in uh large banks that do mortgage origination or um approved mortgages when you apply for a mortgage you send in 100 pdfs that are your bank statements and your like poker statements and your w-2s and some poor soul needs to open each one of them look at what is the document what information do we need to get out of it some of them are not actually digital pdfs some of them are like pictures taken with a phone So you need to extract all of that information out of there and then process it. So this is really where Instabase found product market fit.

5:29And that category of problems is where we've seen early success and have grown significantly over the last couple of years. Mostly of course, with the application of modern AI techniques. So Instabase was, as far as I can tell, actually one of the first companies that successfully applied transformer-based models, real business problems, like bird-type architectures. And now with Generative AI, of course, the whole world is changing. It turns out that you actually no longer need to annotate data. And I guess we'll get more into this later. But the AI Hub product that we're launching or that we did launch this week is the next iteration, if you will of our product and how you actually approach to solving these problems and that's one of the key aspects of instabase as a company actually a customer of ours just told me the other week using instabase is like having a subscription to innovation because we keep innovating and we keep you know like adopting like the latest in the eye and the i hub is really like applying generative models, like specifically GPT, to the document problem.

6:42And what's actually quite interesting and what we found is document is a very narrow, we actually are getting away from using the word document. Content more broadly. Because we actually support spreadsheets, HTML, like emails, PDFs, like handwritten notes, and of course code and like any other like artifact that contains information. and actually allow you to utilize that to extract value. And part of the big investment that we're making is to actually make this more accessible because it used to be that solving an unstructured data problem is a multi-month task that's very intensive. You have to collect data, annotate data, train models, and actually deploy them.

7:28And we find that that's no longer necessary with Genre API. I want to stop there. Yeah. Well, yeah. So Instabase, its first product offering was when and maybe describe how the platform works. Is it a web app or is it a piece of software that sits on the client's computer? and to use it, presumably you've got somebody in an organization who is wrangling data and maybe give me a use case where they upload a ton of data to the platform and then query it on particular things. Yeah, definitely. There were a couple of really good questions in there. But so the first one is, so Instabase as a company was founded in 2015.

8:30And the product as it is used today is a little more recent because initially Instabase tried to solve the document problem in the early days, just like everyone else with heuristics and like rules, right? So like you ran an OCR engine and then like applied scripts basically on the text. and then i believe uh about like actually like late 2019 like 2020 uh we started using transformer-based models like bert um to solve the problem so the product as it exists today or as customers are using it today is roughly like one and a half two years old and uh how it's being consumed is actually a quite interesting question because i i mentioned earlier like the mind meld if you will between me and the founder was how we envisioned that data intensive applications would actually be used and deployed in enterprises and one of the biggest problems in that space is actually portability of applications like you if you write an application for like let's say aws you can't run it elsewhere because there's no operating system if you will for these applications so instabase was actually conceived of as an operating system that provides like operating system type guarantees such as portability abstractions for storage compute and identity so that you can build these applications in one place and deploy them in another the reason why i say this is uh you can probably guess uh from the types of customers we serve so it's like top tier like banks financial services insurances they're very sensitive when it comes to security in data so a lot of them want to run these production workloads on-prem in some cases like literally in their data centers um so instabase runs both as a sas offering so we have a sas offering we just sign up and use the product as a cloud-hosted product so our customers can maintain it themselves in their azure gcp or aws instances and also on premise and the interesting aspect of the product is because it provides operating system level portability guarantees is you can develop in sas and deploy on-prem or like any other combination so it's actually like quite an interesting value proposition especially for highly sensitive data where like some of our customers like really only want to run the production workloads on-prem and then the last question that you asked was what are the types of use cases or like you know how can you imagine is this working one is i alluded to briefly earlier which is really mortgage processing right so you can imagine if you apply for a mortgage you send a whole bunch of pdfs the output eventually is all of the extracted data in a structured format in a database and someone actually like making a decision to give you a mortgage or not right and along that way there's a lot of things that happen in terms of recognizing that information from documents extracting it refining it validating it uh our customers also you like detect basic kinds of fraud it turns out when people like going to defraud banks when they apply for mortgage they do simple things like changing the income in their bank statement but they forget to update the sum in the bank statement so you can actually check that just by like checking the math um so that's one example another example just to show you the diversity if you will we just recently um closed i can't i I can't disclose the name, but it's a big ride sharing company.

12:00And if you imagine like every time they sign up new drivers to their platform, they need to take a picture of the driver's license, of their car insurance, like registration and like vehicle inspection and send that to the company before they can start driving for the ride sharing company. And that's another example where you need to take these pictures. And it's an interesting example because like that's actually an example where the quality of the picture is usually pretty poor. It's really just taking a picture of the driver's license, like in the dark, you're like sitting at home. But you want to, as much as possible, automate that process, right?

12:36Because if you have hundreds of people signing up every day, you don't want humans having to look at those pictures and extracting all of that information. So it's really from a foundational capability, pretty broad. And we started with documents, but we also have customers that do mailroom automation, right? So like you can say like in a big enterprise, you get a whole bunch of mail in that goes through scanners and then hits our system and we classify the type of document and then send it in like different directions with extracted information based on this. And one of the key aspects of our platform that goes beyond just language and like especially large language models is that it's already multimodal.

13:21if you can imagine you know if you get a document sometimes you want to extract things like a logo or a signature or a stamp or like a figure so we also run convolutional neural networks just basic like object detection models to extract those images and our goal is and our ambition of course with the iHub is that we will cover many more modalities as well so we're really shifting towards content in general. So any text document, any images, audio, video is all per game for us in terms of like being able to capture in the future. And the AI hub is, is that, from what I understand, it's like a marketplace of, of AI models or, or products.

14:11Are those use instabase models or products or or is it a platform that anybody can upload a model or product to or do you curate models and products for specific use cases and put them on the platform yeah yeah that's a really important question so i'm going to tell you the the long-term vision and then i'm going to tell you what exists today um the long-term vision is really and like this This is actually how Instabase was conceived, that the AI hub is a community of developers and users where people contribute models, contribute what we call apps, and then other users can consume them. And it becomes this flywheel effect of people developing AI-driven applications, people consuming AI-driven applications, and it really becoming an epicenter, if you will, of applied AI.

15:06and uh so the goal really is that third-party developers can publish their own applications and they can be used by like the the audience of ai hub and there's like a revenue share model behind it um what it is today is uh three different components that we built so right now it's it's seeded by us if you will um so that the the headline app if you will that everyone is paying attention to right now is called converse and that is a interactive q a like chat type app that lets you have a conversation with content so that's the key distinction between that and something like chat gpt right so again chat gpt you're conversing with the model itself right in our product you upload content documents email spreadsheets and then you can have a conversation with that content which is quite interesting the second component is what we call build which lets you build reusable workflows or apps that automate this.

16:04So let's say if you wanted to extract information from a passport or a driver's license or something for identity verification, you don't want to do that in a Q &A type form. You want to actually develop an app where you can just send 10 ,000 PDFs and it gives you the structured output. So with build, you can actually build these reusable apps that automate these workflows. And then the third component is that app marketplace, where we have seeded the marketplace with 10 pre-built applications that we've built. And we will keep adding more and more over time. But we're also now heavily working on the ability for you to be able to publish your apps.

16:42Because you can already build your app. So if you go to the AI Hub today, you can build an app to, let's say, summarize a document, right? Or like pull information out of like an insurance contract. or even translate a document. And you can publish that app on the AI Hub. And in the future, we will have a review process where we can take a look at it, make sure it's secure, make sure it meets our requirements, and then is actually discoverable and usable by anyone who comes to the AI Hub. Are you familiar with Singularity Nut, Ben Goertzel's organization? he's he's kind of a wild figure out there on the fringe but um he's been talking for years about a blockchain-based marketplace for ai models or products uh that that you know anybody could access and then using smart contracts or, you know, use them and pay for them and that sort of thing.

17:51Yeah, it's a really compelling idea. What we found actually, and there's a key distinction by the way, I kind of answered a slightly different question than you asked because like you were asking specifically about models and products. So I think when you used the word products, that's what we're referring to as an app that really is targeted towards solving a real problem models on the other hand so we we also plan of course because internally we have a model repository of actually proprietary models that we trained on top of like bird type architectures that are layout aware models so they will also become part of the ai hub but what we found is that model marketplaces specifically are not really that useful specifically to businesses right in like real world use cases because models are just small building blocks of something that really provides value to you like no business can go to let's say hugging face and find you know like the the latest model and just like create an endpoint and start using it because any meaningful business process or like business solution usually has multiple models that it uses so has some business logic in between has like integration points right So actual apps that solve real problems are much more than just models.

19:08And that's really what we're focusing on and what differentiates our AI hub from what you would refer to as a model repository, which is really more focused towards data scientists. Or actually, I would say hobbyists and academics, because it turns out these days, I just published a blog post yesterday on this. most of these models that are being open source actually have restrictive licenses so that it can really just be used for hobbyists and academics but it's not really relevant for businesses. What is some of the competition then in AI app marketplaces or repositories or however you want to refer to this?

19:58So there's a couple of different components to this. If you look at the apps themselves, right, like that solve meaningful business processes anchored around unstructured data, there's a couple of IDP type companies out there that will sell you these types of solutions, right? like let's say there's an app that says identity verification uh that takes passports and like driver's licenses make sure that the names match make sure that the driver's license uh the the birthdays dates match and gives it a structured information the the big difference i think between our ihop and some of the competition out there is number one a lot of the the competition actually requires you to like sign up like for a very heavyweight like enterprise software deployment and it takes six months to like deploy it in in in your cloud architecture and then run it um our ai hub is now a really like sass in its fullest form which is the app runs in the i hub you can actually right now go to ai.inspace.com upload a whole bunch of documents and get the output and it has a pay go type pricing model where you just pay for what you use So the accessibility and how easy it is to actually use it is one key differentiation between a lot of the enterprise IDP offerings, if you will.

21:23On the converse side, right, like where we're talking about Q &A with documents, there's also competition that's starting there. It's a little newer, right? Because like that's like the form factor, if you will, of like asking questions of your documents just came up with the popularity of ChatGPT. and there's a couple of small niche startups that have started in this space and of course the cloud vendors are all getting into this business as well I was just going to say on the cloud did you say the cloud providers are getting into this space? I mean there's AWS marketplace how would that compare to what you guys are doing?

22:04So yeah so AWS of course AWS especially is a good example because they always have 20 or 30 different services that are like slightly overlapping with what anyone is doing so the aws marketplace itself is is more of a um cloud infrastructure marketplace right so like and database is in the aws marketplace right instabase is in the aws marketplace so it's not at the level if you will of of meaningful business applications it's more on the infrastructure level if you will um aws i believe does have a more like like low code like app development type um type product similar to microsoft's like power apps right so i think like azure's like power app is probably like a good comparable there and all of these also have uh their own like marketplaces and like try to make the exchange of these applications easier um most of them have been developed as just low code app development platforms right it's like not with the idea that they're actually like data driven or ai driven and that actually makes a key distinction right because like what i found when i started to get interested in in the space if you try to allow people to build arbitrary apps uh you would focus really it would start looking like like microsoft visual studio right it's like um just like build a form like write some custom code and that's it but it turns out the data-driven applications are much harder to build, right?

23:31So by data-driven application, I mean any application that relies on processing a large amount of data in the background, running AI models, that's a different ballgame entirely in terms of actually developing and launching these apps. And that's what we're focused on. And I think a lot of the other marketplaces that have been out there for low-code business apps are slowly transitioning into that space because they realize that that's quite important. And then anything will be data driven. Yeah. And then for someone to use these apps, well, first of all, from the app developers point of view, I mean, I can see the usefulness of this.

24:15we've been talking my my friends and i uh that with the the sort of advent of generative ai there is just an explosion of sas offerings and apps and all sorts of things leveraging that, you know, transformer-based technology. And each one is out in the market shouting for attention. And there are different, you know, websites that, you know, Trustpilot or Product Hunt or these various sites that rank and review. But if you want to sell the app, you're kind of on your own. But, you know, a company that comes to mind that I know fairly well is called Accio. And they have a no-code web app that you can upload structured data to and then pick different kinds of AI models to run against it.

25:38and, for example, predict something or classify something. And they're out there selling it on their own. But if it were on something like Instabase on the AI hub, provided that you get the critical mass of users, people see it and they can use it directly right there. And then what, you would take a commission or something? How does that work? Yeah, so the plan is just like any other common app store, right? That there's like a revenue share model and like we would take part of the revenue, but the majority of the revenue would, of course, accrue to the developer of the app itself. But you bring up a really good point, which is the whole idea behind a centralized marketplace and uh and i dare to say like the word operating system behind it is the the point of like distribution and uh and you know like centralized trust right because if you think about it and like this is really the value proposition that we're trying to push which is we have been successfully selling into some of the most demanding global enterprises in the world it's like four of the five top banks in the u.s are our customers and as you can imagine selling into these enterprise is not easy, right?

27:04InfoSec and like approval processes and so on. So the whole value proposition of building on the AI hub would be that, yes, you could start an entirely new company, build your app from scratch, and then try to sell into one of these like distribution channels. And it will take you like a year, one and a half years before you could even contract with these types of companies. Or you could build it on the Instabase AI hub, and we already have the distribution and like there's already a commercial model uh set up there's like identity there's trust right it's like these companies trust us and like we spend a lot of time on the security and compliance side um so it's a much more compelling story um for app developers and that's where that flywheel effect would come in but do they have to build build the app on on ai hub or can they put an instance of their app on the ai hub um so yeah so right now the the the scope of what you can run on the ihub if you will is limited to what you can build on the ihub today um so you can just like take arbitrary code and just run it on our platform um but it's pretty flexible uh or it at least today the version today is like somewhat scoped to specifically content understanding but the software in the background is flexible enough that you could really build anything just to give an example um the platform itself and like the way that a lot of our customers use instabase today follows a generic pattern of um classifying unstructured data and then extracting information from it right so like most of the apps if you will look like it subscribe to my email server like read all of the emails and like classify if this is a complaint if someone applying for something read the information like send it to the different systems but on the same platform we built this converse app uh which is a an interactive q a like chat um solution so the ability to build anything on the platform exists and like we we've proven it if you will ourselves by building this this chat app um but yeah the it is a it does require like building on the ihop um and not just like running anything.

29:19Right. A couple of things. How do you ensure the security of the source code if someone's building it on the hub?

29:31So there's a couple of aspects to this. One of them is because at least today, the public version is exclusively no code. Actually, it's like not even low code. It's no code. So you can't actually like write your own source code. we basically just configure an application and we run it so we have full control over what that code looks like um the product that we have and like where this will evolve will actually allow um developers to of course run arbitrary code and at that point there's like a like app level isolation right like if you think about multi-tenancy in these systems you have like process level isolation you have a container level isolation you have hardware level isolation um so there's there's uh the appropriate level of isolation to make sure that uh one app can't you like escape their boundaries and like access like anything else if you will and a lot of our customers by the way also would run this as a single tenant instance if you will where this whole like portability story comes in right so you could actually in the multi-tenant sas environment have someone develop an app but then if one of those big banks banks wants to run it they run in their own single tenant instance of the ai hub and like in that way um like they can ensure that there's like no intermingling of data with with other customers um and by the way the other aspect that i i want to call out because it's extremely important and we've spent a lot of time on this is uh the bigger concern actually for companies today is more the data security and privacy around the data that goes to llm providers like open ai and so we use open ai in the background and we actually have a very close relationship with them and we put all of the controls into place for our AI Hub product to get an exception from OpenAI for the data retention policy.

31:18So OpenAI by default usually retains 30 days of data to detect fraud, right? And like to, for QA and we have an exception to that data retention policy so that any data that comes through the Instapase AI Hub is not retained at all by OpenAI. They don't use it to train models. They don't look at it for QA. It's really just process it in the model, return it and it's gone. And that was extremely important, especially for enterprise customers to be able to trust the solution because, you know, there's a lot of noise out there as well in terms of, you know, like, are you giving up your data and like, what is your data being used for?

31:57And so on. If so, so if. if i have an ai uh app ai based app uh i can't put it on insta hub i have to i could go on insta hub and build something similar using your tools is that right that is correct and then once i've built it if it proves hugely popular or profitable. Am I tied to InstaHub? I mean, Instabase's AI hub in perpetuity, or can I then download and market it myself? Yeah, so the way that I would look at it is really, I think the operating system analogy is a good one. So I would really look at it as an operating system. system if you develop so like let's take the Windows analogy right so like if you developed a piece of software for Windows it will run on Windows and if you wanted to take that same binary and like run it on a Mac it wouldn't work because it's just incompatible right so the same would be true for the Instabase AI hub like if you build an app to run an Instabase it's specifically to that operating system and like it will run anywhere where Instabase runs but if you wanted to run And outside of the Insta-based environment, it wouldn't work because it just depends on some of the assumptions of the platform.

33:27I see. The converse model or product that you were talking about. One thing that fascinates me about OpenAI and Microsoft is that on the one hand, they're providing these APIs for companies like Instabase to use their models. On the other hand, they're offering solutions that compete with those people using those models. And when you describe Converse, you know, an ability to interrogate a bunch of data that you've uploaded. I mean, that's what OpenAI is offering directly, right? you you you upload your your data through the api and then um and then can can query or or analyze or summarize or whatever you want to do with it uh so how do you differentiate from what open ai is offering directly in in with regards to converse yeah um so yeah let me actually first make a categorical statement, which I think the fact that like OpenAI, Microsoft and others are both providing the services and competing with solutions built on top of the services, I think is like pretty common, right?

35:06That's really what cloud computing is, right? Like AWS provides a cloud infrastructure for you, but they also compete with almost anything that anyone builds on AWS, right? And it's the right thing to do. And like they run a business and they make money on compute and they make money with their own services. So I think that's not going to change. And I think that's just like a competition at its best, if you will. In terms of the differentiation, at least OpenAI is not actually doing this, right? So like if we double click on how this works, right? So like on OpenAI with ChatGPT specifically, you're basically using the GPT model behind it, right?

35:453.504 as a knowledge engine and as a reasoning engine, which is you ask a question and the model will give you an answer to the question. Yes, you can upload data in a limited form, which is like you can in the prompt give it some data, right? So you can say, hey, here's like some text that I've written, please rewrite it. So you can provide it with some amount of data, but you can't upload a PDF, right? Or like you can't upload in a spreadsheet, or you can't upload like a picture that you've taken with your phone because chat gpt is a chat box and it just takes text and the way that the model works is it it has a context window and that context window is like limited right i think it's 4 000 tokens and they have like an 8 000 token for gpd4 and like a 32 000 token uh as well but that so 4 000 tokens like 8 000 tokens i think is like roughly four pages of text so it's not a lot right like you can paste some text but it's not a lot Our product Converse, and we've spent a lot of time actually making sure this works, takes any amount of data.

36:50So you can actually have an example, you upload like a 400 page financial document and ask a question about it. Now, the interesting thing is you could not take 400 pages of text and pass it to GPT. That's just not going to fit into it. Not even with an API call. Well, I thought that you could upgrade to larger amounts of tokens. Yeah, so the largest amount that they currently have is the GPT-4 32 ,000 token window. That will cover roughly 20 pages of text, but it doesn't cover 400 pages, right? And I think I know that they're working on a larger context model. And by the way, there's also other providers of these LLMs, like Anthropic is one of them, that are working on providing bigger context windows.

37:42But the context window is never going to solve for the corpus retrieval type use case, right? Like where you say, like, hey, I have like a thousand documents. These models will never like grow their context window to take like 1 ,000 documents as a context, right? So the way it actually works and what we've spent a lot of time implementing for Converse is when you upload all of these documents, we actually like pre-process them, we create embeddings with them, and then we store them into a vector database. Specifically, we use VV8 for that purpose. And then when you ask a question, we first do the vector search in the database.

38:18And there's different strategies depending on what type of question you ask, right? So if it's a retrieval question, we basically just find the chunk that's most relevant and then we send that chunk as context to the GPT model. But if it's a summarization question, it's like if you want to summarize a 400 page long document, you can't just send a subset of those 400 pages because you want to summarize the whole thing. But then you need to do a more advanced technique such as recursive summarization, where you first summarize pages, then you summarize chapters, then you summarize the entire document.

38:52um so all of that uh infrastructure if you will and the logic around it is necessary for you to actually be able to like point it to any content and then let it answer a question about that content and gpt itself can do that there's other products that are built on top of gpt that are attempting to do this um it's actually interesting we tried to run a benchmark i almost i i was thinking about doing a humoristic blog post about this because we we were trying to a benchmark of our performance on long documents against other products that support long documents and what we found is that no one supports these long documents we tried we tried to run it as a couple of other products that i'm not going to name uh but they just don't support it um so at least today as of like my best knowledge you know like uh the ai hub converse product is the only one that like really performs well on like very large documents and if you think about it common use cases that we actually see in like financial services or like in i mean like even tax preparers like anyone where you may actually have like retrieval type questions like an insurance company says like hey i have 10 000 contracts which of these 10 000 contracts are exposed to like a pandemic risk right um that's uh that's like a question that requires you to actually like look a very large corpus and i mean that may be millions of pages right in aggregate it's like a different an entirely different way of approaching it so long story short uh the differentiation to like the llms themselves is like very significant right because by the way in a way to summarize it and this may be a controversial statement but the llms themselves basically act as a knowledge engine and reasoning engine right because like if you are gpt it basically like it uses the knowledge that it has that it was trained on and it like reasons to give you an answer what we do is we only use it as a reasoning engine and we point it to your data as the the knowledge engine if you will right and that's by the way just to call out like that important point is that's also how we actually of hallucination in our product.

41:06Because now we actually ground the answer in your data. And the model will not answer questions that it doesn't find in your data. So we actually have fun examples. You can point it to a court summons document and ask if someone is guilty. And it will tell you that it can't give you that information based on a document because that's for the court to decide and it's not part of the corpus. Yeah. Who did you say you use for your vector database? It's a product called VV8. Okay, yeah, because I've had Ido Liberty from Pinecone on the podcast. I don't know. I assume you're familiar with Pinecone.

41:52And I met somebody recently who has a company that sounds similar. I'm not going to be able to pull the name out of my memory bank right now, but also leverages large language models for their reasoning.