Everyone, welcome my, welcome to my talk on introduction to vector databases. So over the next 30, 40 minutes, I want to take you through a general introduction of what vector databases are and how they help you search over your data in a much more organic, organic manner. And then I want to introduce you to the basic concepts of vector databases, how we take those basic concepts and scale them up so that you can search over for hundreds of millions of data objects, if not billions of data objects. And then I'll end off with hopefully, what's an engaging demonstration of the power of vector databases? So before we jump into vector databases, a little bit about myself. I'm a developer advocate at Weaviate and I'm an engineer and data scientist by training. And the first time I really learned about vector databases, that was a bit of a light bulb moment for me where I thought, why doesn't everybody just use vector databases? They're just a much more natural way to search over and deal with your unstructured data. And so if the only thing that you get out of this talk is that nugget of insight around what vector databases can help you accomplish and how they can help improve the projects that you're working on, then I'll have considered my job done. Okay, so the essence or the main idea behind vector databases is that these allow you to go from a classical approach of keyword search, where in, let's say a SQL structured database, if you want to look for an item, you would have to do some sort of keyword search or exact matching search. Vector databases transition from that idea and allow you to do a similarity search or a semantic search over your data. This is a fairly abstract definition, so let me further explain this using an example. So imagine you've got a bunch of documents that you store in your database. The first one here as an example is how to build a rest API. Imagine you have thousands and thousands of these and you want to search over these documents. The traditional approach would be you go into your search bar, you type out the word python, and the traditional keyword search would go in and would essentially look for the word python anywhere in your document, right? So imagine you only have two documents here. The word python is nowhere to be found. So it says that it returns nothing and there's no articles that have this keyword. The more fundamental problem with this traditional keyword search is that it technically doesn't even understand whether you mean Python the animal, the snake, or python the programming language. And that gets to the heart of the problem here. Keyword search is just doing string matching to find out the document that you are looking for, it doesn't actually understand the context and the meaning behind your search. And so on the other hand, if we do a semantic search, we go in, we type out the exact same word, and now it recognizes that in the context of the documents that you've provided me, you're probably not looking for the animal, you're looking for a programming language for data scientists, which is the Python programming language. And the essence here is that it understands your query and it matches it with a similar document that's in your database. And in order for this to happen, it has to understand everything that's in your database, the documents as well as your query, and then it has to match it appropriately. And this functionality of understanding your data and then going in and matching it is exactly what a vector database does. And instead of just being able to search over a couple of documents or thousands of documents, it can search over hundreds of millions or billions of documents. And that's the scalability of the vector database that we'll talk about in a bit. So really, if we want to achieve this end goal of semantic search, similarity search over our documents, there's a couple of hurdles in our way. The first hurdle being that we need to take our unstructured data, we need to be able to process it, understand it and search through it. And that's a difficult task. The second tall order, the second hurdle in the way is that it's not enough to just be able to search over and understand a small amount of data. We have to be able to do this in a scalable and secure way so that we can search over hundreds of millions or billions of documents. And so those are the two hurdles that are in the way before we can achieve this idea of semantic search that I just presented. So we'll go through and see how vector databases address each one of these issues. So the first issue, how do we understand and search through our data? Well, we use machine learning models to understand the context of our unstructured data. And so every time I say unstructured data, I mean either natural language, text. So this can be documents, books, emails, or it can also be images, video, audio. All of your unstructured data, we pass them through a machine learning model. It understands the context of that data and then we use the output of the machine learning model to learn about these context in our vector database. So let's dig deeper into this concept of using ML. So let's say you've got your data over here on the left hand side, upper left. This can be emails, videos, audio, images. We're going to pass each one of these through our machine learning model. And our machine learning model is going to spit out a vector associated corresponding with each one of our unstructured data objects. And the way that I like to think about this is that our unstructured data is in human understandable format. We can read an email, we can view and understand an image, but the machine or the computer can't understand it. So these vectors that the computer generates are machine understandable. So it's much more easier for a machine to understand these vectors for our data. So now that we have our data translated into machine language, now what we're going to do is project this and plot this out and create vector representations. So each object that we had a vector for here is now represented using a green dot on the right hand side. And so we call this a vector space or an embedding space. So we take all of our vectors and we embed them into this vector space that I've demonstrated using a 3d graph. There's two important things about this vector space. The first one is that it preserves similarity between objects. So, for example, if you look at the vector for these word cat, it is going to be close by to the vector for the image of the cat. And the machine learning model understands that the word cat and the image of the cat are similar in nature in these unstructured data. So it's going to keep the vectors closer together. So similarly, if you have the word chicken and the image of a chicken, those are going to be closer together in vector space. On the other hand, if you take a look at the word cat and the word banana, the vector for the word banana, those two concepts are further away, and so the corresponding vectors are also further away. And so in this way, you can take objects, or unstructured data, objects where you can show them to a human, and you can ask them which two of these things are similar and which two of these things are different. Now, you can do the same thing by conducting a distance metric around your vector objects in vector space, and a machine can do this much more easily. The second thing that's important to keep in mind is that even though I'm showing you these green dots in a 3d space where I have three axes, in reality these vector spaces are often very high dimensional. They can be 300 dimensional, 600 dimensional, 2000 dimensional, depending on how much data you want to preserve, how accurate you want the vector representation to be, you're going to choose a higher dimensional vector space. It's kind of like asking a person to read a book and summarize it into one page. And that summarization is going to lose a lot of data. If you ask them to summarize it into one chapter, they'll be able to preserve a lot more information in that one chapter. So the higher the dimensionality of the vectors, the more descriptive and the more information that they are going to preserve from the original unstructured data format. Okay, so now that we have these vector representations, that's not enough. What we need to be able to do is we need to be able to have millions and millions, if not billions, of these vector representations, along with their unstructured data object counterparts. And then we need to be able to search over and query these objects thousands and thousands of times per second. And this is this idea of being able to scale your vectors and search over them that a vector database really excels at. So in order to be able to search over our data, we need to first of all store billions of these data objects, both the vectors as well as the non vector unstructured counterparts. In order to have this happen, we need to have our machine learning models work at scale in a production level environment. And so the idea is that every time you have a data object, an unstructured data object come in, it has to go through the machine learning model. You have to perform inference, you have to take the vector that comes out, pair it with the object that went in producing that vector, and then you have to store both of these things into your vector database. And then of course the vector database, the keywordsto there being database has to support crud operations. Any of these green objects that you have on the left hand side here could need to be deleted, updated. You could need to add new data objects depending on the velocity of your data that you're dealing with. So a vector database has to have crud operation support. So when we're talking about scaling vector search and scaling, the ability to embed vector objects in our vector databases, that's exactly what a vector database excels at. So I want to go through the definition of a vector database, and then the problems that you need to tackle as you need to scale, storing vector objects on the billion object scale as well as searching over them on a billion object scale. So when we talk about a vector database, really these fundamental, the first fundamental is that you need to store the unstructured data objects and all of these corresponding vector embeddings. Anything that you want to be able to perform vector search over, you need to have the original unstructured object as well as its vector embedding. And then it's not enough to just do this at a small scale, you have to be able to scale this up efficiently to millions and millions of data objects. And so if we look at some of the challenges that a vector database needs to solve, the first thing that it needs to do is to be able to store both your data objects, your unstructured data objects, as well as the vector representations. You have to be able to perform similarity search. These idea of a semantic search over your objects as well as a structured filtering. So for example, in this picture on the right hand side, let's say I wanted to tell my vector database to query and search through all of the animal objects that I've got in my database and tell me which one of these objects I can make a recipe out of. So in these case it would sub filter, it would do this structured filtering where it would exclude all of these company logos, all of these fruits out, and then it would allow only the animal objects to pass through. And then you would go through and then say okay, which ones can we use in a recipe? And then it would let the chicken pass through that filter. And then you could do some sort of vector search over the post filtered objects. And then of course let's say you have new data that's coming in, you have to be able to take that unstructured data object, turn it into a vector store. Both the unstructured data object as well as the vector, you have to support create options, you have to support delete options as well as update options. And these, the main idea is that in order to make this scalable and efficient, you have to implement a nearest neighbor's algorithm, which we'll talk about in a second. So when we're vectorizing data, when we're vectorizing images or natural language, we use machine learning for that. And usually those are neural network or deep learning models. And you can use any open source model or any one of the companies that we've partnered with, for example cohere, OpenAI, Google's new large language model, to take in your data, vectorize it, and then the vector database stores that data. So once you've done that, this is similar to what your data would look like. And then what we need to do is go ahead and query our data. So let me show you an example of what querying a vector database looks like. And it's quite different from something like a structured query language query that you would see. So all of the green dots on the right hand side are your data points that are embedded in vector space. So now let's say a user comes along and they have a question. We're going to take that question, that query, and we're going to pass it through the same machine learning model and create a vector for that question, which shows up as that red dot on the right hand side. And the human understandable representation for this query is nothing other than a word. So in this case, the word being kitten. So if a user comes along and queries your vector database with the word kitten, that word goes through the same machine learning model, it gets projected into vector space and it creates that red dot. And the act of vector searching over your vector space is essentially the act of going through and saying, which green dots are the closest to my pink dot over here? And then the closest ones. Let's say the user wanted the top three closest vectors that are in my database. It's going to go through and pick the closest three green dots to my pink dot and return those to the user in the corresponding images or text for those dots. In reality, we don't actually go through and do a brute force search where we compare the distance of the pink dot to every other green dot. That would take hours and hours depending on how much data you have. What we do instead of doing a full brute force nearest neighbor search is what's known as a ann, an approximate nearest neighbor search. And the idea behind this is that instead of searching through all of the green objects that you have in your database, you kind of set up a hierarchy of layers over here where you enter your search at the very top and you look through a very small subset of all of your objects that you have in your vector database. And then as you go down this hierarchy of search, you search over more and more data points. And so in order to get to the final nearest neighbors, it takes a significantly smaller amount of time. You have to do significantly less comparisons and compute. The advantage here is that you can search over hundreds of millions of objects without actually having to do hundreds of millions of distance computations. The disadvantage over here is the a in this algorithm's name, it's approximate, so you're not guaranteed to get the closest nearest neighbors. For that, you would have to do a brute force search. But given the amount of time that we save doing approximate nearest neighbors, this is the only way to reliably do vector search at scale. You have to have some sort of approximate algorithm that can search over all of your data objects. Okay? So with those fundamental concepts of being able to embed your unstructured data objects into vectors and then put that into your vector database and then scalably search over them, supporting crud operations that defines the foundations of what a vector database is. Weaviate that I'm going to introduce now is an open source vector database and you can try it, you can go to our website and play around with it. And the main idea behind Weaviate is that it understands your data because it vectorizes your data. It understands your data in this vector space and it allows you to search over your data in a way that it understands it. So I want to show you a typical vector search pipeline with Weaviate just to give you a center idea of how all of these moving pieces come together. So everything kind of revolves around the machine learning model, the machine learning model that can be open source so it can be Resnet 50 that's widely available. It can be a model that a friend uploaded to hugging face and made available to the wider community. It can be a model that, for example, OpenAI has created and is an API for. So all of these models allow you to take your data over here and the data can be in whatever format the model understands and generate vectors for that data. And then you take those vectors and you pop them into VVA, and VVA stores both the unstructured data object as well as the vector representation for that object. So now let's say a user comes by and they have a query they want to search over your data using a concept. We're going to take that query that has to be in the same modality as the data modality that the machine learning model supports. Weve going to take that query, pass that through the machine learning model and generate a vector for that query similar to that red dot that I just showed a couple of minutes ago, and we're going to pop that into weve as well, and we're going to do a nearest neighbor search around that query. And depending on the results, depending on the closest objects that we have in our vector database, we take that and we return these results to the user. And the cool thing about Weaviate is that it's modular in the sense that you can go ahead and plug and play whichever models you want to use, whether that's your own proprietary model, you can plug that in, you can plug and play any one of the partner companies models that are available. So if you want to use an OpenAI model, you can plug that in. We've integrated with Google's Palm model that was released just last week. You can vectorize your data using that model. You just have to specify which model you want to use. The other option that's also popular is you can bring your data pre vectorized to us and then we can store that into weve eight. The different data modalities that we support are limited only by what the machine learning model understands. So if you have a machine learning model that understands image, video, audio, all of these different data modalities, you can take all of those different data modalities, vectorize it using the model, and plug it into vv eight. Probably one of the things that I'm most interested about is what happens when you take the results, the list of results that are the closest in vector space. What can you do with them? And weaviate supports this modular output where instead of just displaying the results, you can pipe them through any other functionality so you can re rank them. You can do some question answering with them. You can even pass them to a large language model like Chat GPT, which is a really interesting application that a lot of people are interested in. So I'm going to dive into that a little bit over the next few slides. So the main idea behind this is we want to use vector search, this idea of searching over hundreds of millions of data, unstructured data objects, and returning the most relevant data objects. And instead of showing the results to you, we send them off to a large language model. And the whole idea behind this is that you prompt a large language model and you say, answer this question, given the information that is returned by a vector database. So the information returned by a vector database provides relevant context that the large language model can then use in formulating a response to your prompt. And so imagine you have a browser of Chat GPT open. You would say, answer my question. Whatever your question is, you would type it in, and then you would say, the vector database returns all of this relevant information that you need to know. And you say, here's everything relevant that you need to know to answer my original question. And so now the workflow for asking your Chat GPT model, a prompt is a bit different, where you can say you have a prompt that comes in, you pass it to your large language model, but instead of just letting it rely on its general knowledge of information, you also pass it relevant documents or relevant information. Usually how I've seen people solve this problem is they go through, they see whatever information is relevant, they copy paste that and they pass it along with the prompt into the same window. The problem with that is that there's only a limited amount of information that you can copy paste in. And the other problem with that is you have to do all of the relevant information identification and filtering yourself. So imagine you've got thousands of documents here and only some are relevant to the prompt that you're asking. You have to sit there and go through each one at a time and say, this is relevant. That's not relevant. Once you're done this, you'll have a subset of the relevant filtered documents. Then you can go ahead and pass the prompt over to your large language model along with the documents that you filtered over that you copy paste in. But if you think about what the point of the user is over here, they're just performing similarity search over your documents. They're reading the documents one by one. They have some concept that they're looking for, whether it's present in the documents, and then if it is, it goes through. If not, it gets rejected. It's not going to end up in the list of relevant documents. That's exactly what a vector database excels at. So if we want to scale these approach, we can replace that user that was sitting there and filtering our documents with a vector database. And the vector database, its main job is to house hundreds of millions of documents and search over them to provide relevant, the most relevant ones that are relevant to your prompt. And in the demonstration that I'll show, I'll actually show you a demonstration of how I'll have 100,000 objects that are stored in my vector database. I'll provide it some context to query over those documents. And then I'll filter the documents and send it to a large language model and then use generative search over the documents. And so in the example that I'll show you, I'm using Weaviate. The cool thing about Weaviate is that we've implemented an endpoint whereby you can pass in all the documents and then instead of returning the results, you can say, take these results that Weaviate returns these filtered relevant documents and send them to a large language model. And what you see at the end of that is the generated text as a result of the answer. So you see the customized response that the large language model is going to provide to your prompt as well as the relevant documents that you've passed in. So that's a lot of talking from my side. So what I want to do is I want to go over a short demo that shows you the power of vector databases and what they enable. Okay, so let me go into my Jupyter notebooks ide over here. Let's start off all the way at the top. So to get this demo working you're going to need the Python Weaviate client. You're also going to need to go to the link to download the data. And then the other thing that you're going to need is to go to weaviate cloud services over here and then log in and you're going to need to create a cluster. So these is a remote cluster, a remote instance of weaviate that is going to store your data and then perform vector search over it. I've already created this cluster over here and you can see that I've already uploaded about 100,000 unstructured data objects in there ready for us to query. So we're going to go in here and then we're going to go in, import our Python Weaviate client. The other thing we're going to need to provide is a WCS token. So the WCS token is important because once you create your cluster, if you enable authentication over here, you'll see that you have a token that you're going to need to provide to be able to remotely hook into that instance. So over here this token is provided from these console. And then if you're using any third party machine learning models to vectorize your data or query your data, you're also going to need to provide API keys for them. So here I've provided the relevant keys and then once the client is ready it'll print out. True. So a little bit about the 100,000 objects that I've uploaded, these are just Wikipedia articles that weve chunked up into paragraphs and weve got a bunch of metadata around them. So for example, we have the id for the Wikipedia article, the title, and then we have the actual text from the paragraph, the URL, the wiki id, so on and so forth. We're going to take all this data and we're going to batch it and upload it to our remote instance that we just created. Before we can do that, we have to create a database schema that lets weaviate know what data is coming in, how we vectorize it. So to do that we define a class and we give it a name. So these are going to be Wikipedia articles, a description over these and then we specify the vectorizer. So because our data is text, we want to vectorize text. So we use a specific text to vec machine learning model. In this particular case, this is a multilingual model and that's going to be very advantageous later down the road. So I'll come back to that in a second. We're also going to specify exactly the distance metric that we want to use when we're conducting vector search. So in this case it's going to be a dot product. So when it's comparing how far away your query vector is from all the other vectors, it's going to conduct a drop product between those two vectors to find out which two data points are closer together, which ones are farther apart, and then we go in and we name all of our properties over here. So we've got the text of the Wikipedia article, we've got the title, the URL, the metadata that we went through in our pandas data frame. So once we're ready with the schema, we're going to go ahead and create the schema and then we go ahead and batch upload data into EVA. So we're going to start off, and we're going to say start off with a batch size of 100. So upload 100 article paragraphs at a time and then we'll set this to dynamic so that this can increase over time if necessary. And then we're going to go in and loop through our data and upload 100 objects at a time. So this takes some time. So in prep for this demo, I've already uploaded all of that data and you can see that all 100,000 uploaded correctly. Once the data is in there, you want to do a quick check to make sure that you've got all of the unstructured data objects are registered and counted for. The way you can do that is just by doing this simple query and counting how many objects there are. There's 100,000. Another way to verify that is, as I showed, if you go into the console, you'll see how many objects you've got in here using this object count. So now that all of our data is in these, I want to go through and show you these different ways that you can search over your data. And I'm going to start off with the way that I started this talk with classic word search. So this is boring old word search where I'm going to go in and create a filter where I'm going to look for a specific word in the titles property of my objects in my vector database and I provide in the exact word that I want to match. And then I'm going to print out only the first one. So I slice out and I print out the first object that it finds. So for example, if I go ahead and if I do a word search for the word avocado, it matches with the word avocado that's found in this document, that's in the title, and it returns this text from that Wikipedia paragraph, I can go ahead and do the same word search for data science and it goes ahead and it matches the word data science over here with a string in between, and it returns this. So nothing exciting, nothing interesting. But notice what happens when I do a word search for fast animals. Supposedly I'm looking for all the different or one of the paragraphs that describes a fast animal. It goes ahead and gives me this error, and the code is not wrong, the implementation is wrong. The problem here is that if I take away this word search function and I show you the actual query, I go ahead and I do a value string of fast animal and it looks for this string. What I actually get back is an empty list. And the reason why is because nowhere in my 100,000 objects do I find this string fast space animal. And that's one of the downfalls of keyword search that I was showing you before. If the actual string does not exist, you get nothing back. The database doesn't understand what you're trying to ask for. It doesn't understand the data, so you don't get this back. So now what I want to show you is with vector search or semantic search, you can really remedy this problem really quickly with semantic search. What we want to do instead is go over and search for a concept. Instead of searching for a value string or a keyword search, we want to search for a concept and that concept is going to be vectorized and we look for the closest concepts. So we have a parameter here, concept, and we're going to go through. And one other thing that's important here is we're going to call the near text search. It's going to go through and it looks for the text objects that are the closest to my query object over here that I'm providing. And I also limit the return to three because it'll be nicer to look at, it won't overpopulate the page in front of me. So I'm going to only return the three most relevant results, the three vectors that are the closest to my query vector. And then I have a function that's going to format and style this nicely. Let's say I go through and I search for this concept, a programming language used for machine learning. And the first result, the closest vector to my query vector is python, and then it's c plus plus and then central processing unit. And if you look at what we're doing here, we're literally just typing and chatting with our vector database. I queried it using this concept programming language used for machine learning, and it realized even though the exact string was nowhere to be found in any of these texts, it found the closest concept to what I was asking and it returned that. And that's the power of vector databases. You can never do this with a structured or unstructured regular SQL NoSQL database. The only way that you can do this is if your vector database understands your data as well as the query what you're asking for. And it has a way to quantify how similar or dissimilar those two things are, which a vector database exactly does. That's the main power of a vector database. Again, if I go back to my original query of fast animals and I conduct a vector search now, now it gives me relevant answers, right? So it goes through and it vectorizes fast animals and it realizes that a gazelle, a cheetah and a bobcat are fast animals, because when I vectorized these objects, there was some mention or it understood that these unstructured definitions, text definitions, are affiliated with fast objects. It could be mentioned somewhere in the object itself, or so on and so forth. And this is the power of vector databases. And if you remember, I told you that this was a multilingual model that we were using to vectorize our data. And that means that you can query it in any language that you want. So here I've queried it with, this means great movies in Chinese and it comes back and it shows me the most relevant data objects that I've stored in here. So, got goodfellas, totally spies. You can also query in Hindi. So this is the same query, great movies in Hindi, you get Schindler's list, the Dark Knight. So this is quite powerful. The flexibility of your vector database is only limited by the modalities and data types that your machine learning model understands and can vectorize. And again, vacation spots. This is in Farsi and it understands that as well. For the next part here, what I want to show you is this idea of generative search, searching over your vector database and then instead of just returning the results to you, piping those to a large language model, and then providing those results as context to the large language model so that it can answer a query or a prompt using those documents as context. So the context that I'm going to provide here is going to be from this return query. So I do a semantic search over my vector database for famous basketball players and it returns to me three famous basketball players. You've got Will Chamberlain, Magic Johnson and Will Chamberlain. The reason why it repeats will Chamberlain here is because the same Wikipedia article can be chunked multiple times. So I've got different paragraphs from the exact same Wikipedia article. So now what I'm going to do is instead of show you the results of this query, I'm going to tell Weaviate to take these results, send them to OpenAI's large language model, answer a question reading these results, and then give me the generated text back. And this process is known as generative search or retrieval augmented generation. So here the interesting thing that I want to do is this is the prompt that I would give to chat CPT. So I want it to write me some interview questions that I can ask. So this is something funny that's going on here that I'll explain in a bit that I can ask title and also how title would answer them. Here's some information about them. And then I've got text in here. So to understand where these titles and text come from, let's look at the actual query here. I start off and I query my client and I say that I want to do a near text search where the concept is famous basketball players. And this will return something like this that I've shown here. And what I want to extract from this is the title as well as the text. So it's going to give me the title of the Wikipedia article which is shown here as an example. And these the text of the Wikipedia article which as an example I've highlighted over here. And instead of showing you the returned objects, I am going to pass them over to the generative model using the with generate command here. And so in this case what happens is the title gets replaced with whatever title was returned by the vector database. The text here gets replaced with whatever text was returned by the vector database. And that creates the whole prompt. And that prompt gets sent to the large language model. And this single prompt parameter that I've provided here essentially makes it so that every single data object that the vector database returned gets passed one by one to the large language model. And then you get the equivalent number of responses back that you can show. So here I'm just going to show you one of the responses back. So if we scroll down here, you can see that the relevant context that was provided was Will Chamberlain and the text for this Wikipedia article title. And then the generated text that I got back from OpenAI was what was your biggest challenge as a basketball player? So it's essentially made a question answer session where I ask questions to will Chamberlain and he responds based on what the large language model understands of the context that I provided it. So this is the power of generative search that comes built in with weaviate. And this is just one query that I can do all of this with. The other interesting thing that I can do is I can say write me a heroic tale about, again, take the results from the vector database and then pipe them to the generative model and here's some context about them. This is the actual paragraph. And so it goes in and it generates me a story about will Chamberlain and all of the context that's provided here. It comes from the general knowledge of the general knowledge of the large language model as well as these context that the text and title that the vector database provided it with. Another interesting thing that you can do is instead of passing in one of the unstructured objects at a time and generating a prompt result using one object at a time, you can group all of the objects that the vector database returns and pass these all together to your large language model to answer a more complicated question. So for example here, my question to Chachi Bt is which of these basketball players mentioned in text is the most accomplished? And it has to choose at least one and explain why. And the titles and the text here are going to be replaced by 15 of the retrieved documents that we v eight returns to me. And so here I go through and I'm going to show you the articles that were provided as context. First of all, so these are the names of the articles that we provided, the large language model as context. And then this is the answer that the large language model provided to us. So all the way over here and it shows that Will Chamberlain was the greatest basketball player, the most accomplished basketball player. And it explains why, given the context of information of not just will Chamberlain but all of the other basketball players that we asked it to compare Will Chamberlain against. And then one last example here, kind of got excited when I was putting this together. I go ahead and I say, give me five famous basketball players. Weve goes and searches over my Wikipedia documents. It returns the titles for those basketball players as well as the context, the paragraph for those basketball players. And then I ask story of tell me a story where these people, all these basketball players fight each other and then I give it context. I give it information that the vector database returned to me and I pass that in over here. And so if we look at the results now, first let's look at the context that was provided to Chat GPT here. So here we gave it the information and the name for Will Chamberlain, Magic Johnson. Again, a repeat of Will Chamberlain, Scottie Pippen. We also gave it information for James Naismith, which is what the vector database returns. James Naismith is not a basketball player, he's the inventor of basketball. But we'll let that slide for a second. Let's have a look at the generated story that we got. So in this generated story we're essentially customizing the response of Chad CPT to these context that we provided. So now the story starts off and it's saying that will Chamberlain and Magic Johnson were both legends, so on and so forth. And it tells a very intricate story. It's got Scottie Pippen in there and it's kind of pitting them against each other. And then it shows that Chamberlain stood tall, which it thinks that once Chamberlain is the best. So of course he's going to win at the end of the day. But that's the power of vector databases and what you can do with the inputs, the outputs, how you can chain them with large language models to get these large language models to answer your prompts grounded in the context that the vector database provides. And so this is one of the most exciting things around vector databases right now, where vector databases essentially act like long term memory for these large language models. They can go and retrieve ten of the most relevant documents to a prompt over millions of documents, and then answer your prompt, given that information as context. Okay, that's the end of my demo, so I'm going to go back and go over here. So I hope everybody enjoyed this intro to vector databases. If you have any questions, feel free to connect with me on Twitter LinkedIn. You can join our slack community. The entire team is around, we're happy to help. We'd love to get you to try and use weve eight. If you have any cool ideas, let us know. And then if you do have any questions as you're using WeV, feel free to shoot us a message. If you come up with anything interesting, if you come up with cool implementations, cool projects that you're using Weaviate with, give us a shout out. We'd love to talk to you. You can also blog about it. Always love to see what the community uses this open source tool. For me thank you to Con 42. This was great. I really enjoyed this, and I hope you guys enjoyed this as much as I did. Thank you, everybody. Take care.