Rethinking your data stack: The future of AI apps and vector databases

So, uh a quick question before I kick off - who's familiar with vector embeddings and vector databases? Is this like a new topic or do people know what this is? Yeah, I see a couple of it. Yeah. Or are there some people from who this is new? Some new? Great. So that's, that's one of the things that I wanna tackle because it's like a um um it's quite new technologies, quite new infrastructure and that's what I wanna talk about.

So the um, I'm one of the co-founders of uh of Weavate. So Weavate is an open source uh vector database and we founded the company in 2019. And the reason we started to build new type of infrastructure was because we saw that um, machine learning models that came back on the scene. So starting around 2015, we started to output vector embeddings. And one of the things that you could do with these vector embeddings was that you could do search with them, right?

So in the early days, we had search and recommendation how, when we had started open source and then everything that happened with the, um uh with the uh with the models and with the large language models and everything that came on the scene there was like that everything took off from that, right?

So that's what I want to talk about. So the first important thing is to understand what these vector embeddings do, right. So to get an understanding why we can better search with vector embeddings. And the thing is this, so what vector embeddings do is that they make a relational representation on different data objects. So the simple example that I always give for people to learn what they do is that if you store data or if you search with data. So for example, we store the Eiffel Tower in Paris, then we need to retrieve the data object for searching for Eiffel Tower or Paris. But what the vector embeddings do is that the vector embeddings um have a numerical representation of your data set. And now if you, for example, search for landmarks in France, you can find data objects too. So it's a completely new way of doing search.

But what's important to bear in mind was that better search was like around 2021 you know, 2020 2021 was the important use case for the uh for the vector databases. But one of the things that started to happen um was that we saw new use cases coming up because of the generative models, right?

So when we store information in the database. When we store information in this, in this example, we and we retrieve them based on the vector embeddings, we can actually feed them into the generative models. And this is an example of that. And that is an um um an application that we call um uh uh retrieval augmented generation.

So what we basically do is that we retrieve the generation of output out of the generative models with data that we put into the prompt. And that was the new use case that started to emerge for vector databases. And you've seen this in action because this thing became prominent because of the lines of CGPT.

So people started to ask themselves the question, hey, if we have our own data, but we want to have all the analysis of the generative models. How do we do that? How do we achieve that? And that is the uh the u the new use case that started to emerge for um uh uh for vector databases.

So one of the things that sometimes we refer to it as something we call the R stack, we are basically saying that we intertwining the database, the vector database with the generative models and the um uh uh and the factorization models. And now there's like a lot of places where you can get these vectors, right.

So for example, if you look at V8, you have integrations with all here at AWS with Bedrock SageMaker, you can use the open source models of the I Cohere whatever you want that all comes out of the box in the database. And today we see that the majority of these use cases are on this. So basically, you have a query, retrieve documents from the database and stores and, and it's used to generate results. So it's like the quote unquote, you can create agents with it or quote unquote your own GPT like uh applications based on your own data.

However, one thing that I'm super excited about that is also emerging out of this and it's something new that we are that I'm, yeah, that I'm just very excited about is that this is a one way street. So what we're doing here is that the moment that we get data out of the database and we put it into the um the generative model, we use the output and then we're done. But one thing that we can start to do is that we can start to create generative feedback loops.

And what is the feedback loop? The feedback loop is that we store data in the database. We have a generative model, retrieve the information from the database based on the the vector embeddings that are created and loop it back and store it back into the database.

So one use case that we have one example that we have is open source. It's on our website is around Airbnb data so what you can do there is that the Airbnb data is on purpose missing descriptions or the description fields are empty. So what we do is you query the database for all listings, all Airbnb listings without a description, you have the host name, you have the price per night, you have the information on the location. You use the generative model to generate a um uh a description for the listing and you store that back into the database with the vector embedding. And now you can search for things like, you know, a place in New York to walk by dog or those kind of things, something you could not do on the uh uh uh on the original data set, but you can do that with a generated data set.

So that is like a next step that I'm very, very bullish on what we see with vector database in combination with the generative models. This is focusing now a lot on language and on language models. But by no means I think that it will be, it will stay there. It can also be images, it can be multimodal, you can all do that based on your own data.

Um so one of the things that we're focusing on, we, when we started the um uh the database is that we learned that a lot of people like some hands that went up earlier that some people actually like, hey, we don't, it's new to us, right? So we want to build these AI native applications, but you know, we don't know the nitty gritty details of how these models work or how they operate, especially in combination with the database. Do we choose our open source models? Do we use services, those kind of things?

So one thing that we started to focus on as a company, besides that core infrastructure is helping people through the content that we're creating trainings that we do to the webinars and those kind of things to learn and to help you get off the ground, building your own AI native applications.

So one of the things that I'm going to do in a bit is that I'm gonna invite you to our booth because we have these demo stations where you can actually see that in action. So you can just have a one on one session with people to learn and to see how these um uh uh how these retrieve augmented generation applications, vector search in general, et cetera, how that works. And depending on your, you know, your level of knowledge and expertise, people can double click on it and do a deep dive. So you can go high level, you can down, go go deeper down into the details if you like, but there's something very other important because if you go to the booth, you can win a lot of prizes, which is also non unimportant to mention, right.

So the um going to booth is, um, we are at 1620. That is right over there. It's a, it's a big booth right in the corner and I would invite people to take the demos to go um to the material, to talk to the people and to learn what you can build, attracted databases and these generative models.

That is what I wanted to share with you. I have a couple of minutes left for some questions if you have any. And uh but last but not least, please see us at booth 1620. Thank you.