Open Source Milvus Vector Database – Frank Liu, Zilliz

This is Textron TV. Hello everybody. Welcome to Tech strong.

I'm your host Rebecca. And today. I'm really excited.

As I am here with zilis the creator of the popular open source, Vector database millbit has recently raised 60 million and funding in this economy now with download surpassing the 1 million Mark and production users increasing by 300% from last year. I'd like to welcome Frank. Hey Frank, how are you?

I'm doing great. Thanks for having me Rebecca. No problem.

Hey, could you tell us a little bit about yourself? Sure. Yes, so, you know, I've been in and around I would say the AI ml space for probably the better part of seven eight coming up on nine years and you know, I first start off at Yahoo.

So a much larger company and you know, we were doing a lot of computer vision and trying to put a lot of machine learning models into production right after that. I actually did my own startup. And also what we were doing there is a very different where you're doing these using machine learning to do indoor positioning on these very low power iot sensors.

And now I'm here at here at zuliz and we're doing a vector database right working on structured data. That's phenomenal. Now, it's really exciting you guys once again have to say congratulations on the 60 million in funding.

What can you tell us about that Arnie? Are there any big plans in the future? Yeah, I know.

Absolutely. So yeah, I mean, I guess first off, you know, I'd like to thank our investors for believing in our mission and believing what we're doing. But I think you know this latest round of funding I think is really going to help us scale out our cloud services as well as our teams here in South Valley, right?

So we are a company that is based in the San Francisco Bay Area. We're actively looking to bring folks on technology not just to bring focus on but I think to educate a lot of the industry about what a vector database is, right? What is unstructured their processing?

Why is it something that Pretty much I would say close to every organization should take a look at and you know, I think building out the platform building out everything from engineering to go to market. That's really what we're going to be using based trying to funding for. Well great.

Well, but that said I have to you know, just dive in and say what is unstructured data. Yeah, great question great question, unstructured data, you know, it's got a fancy name but really unstructured data just means anything that can't really be stored in a table-based format some extra traditional relational database anything that is doesn't really have a predefined data model. Right?

So for example, strings numbers tables or even Json documents. We look at database like among the DVDs are meant to store structured data. If you look at unstructured data a great example that I like to use is a human generated data.

So images video audio text natural language. These are all examples of unstructured data, but I think there are a lot of less mundane or less boring examples too. Right?

So for example geospatial data, right user profiles, what are some other ones that I can think of, you know iot streams of iot data. These are all examples of unstructured data and there's so much more of that out there then structure data today. No, probably if you look at in 1970s or 1980s most of the data that was being generated with structured.

But today, you know, I would say coming up on 80 and soon to be 85 90% of all day today is unstructured data. So we need a way to be able to store index and search across all day. They find that fascinating you say that 90% of the data that's out.

There is unstructured and it's not until now that we're starting where I'm starting to hear of it. And I'm just like okay that's fascinating now just so viewers and you know listeners out there understand so structured data, that would be kind of like just a basic explanation would be, you know, your date calendar date, you know. Similar to that, correct?

Exactly, you know a date anything anything that you can really store in a table or Json documents, you know these wide column stores. You can also consider them structured or semi-structured databases. But yeah, you know and I think you know following about thinking of following up on that is, you know a big reason as to why You know unstructured data is probably not, you know, probably not a very common term is that there's really, you know, I would say up until today up until you know, the era of of AI there really hasn't been a great way to allow computers to understand all of these different pieces of unstructured data, right and you know what we have today at our disposal or these really really powerful machine learning models and you know, they they generate these embeddings are these vectors right these floating point bacteries that are really really good representations of the input.

So the analogy that I like to use this and so if you have We have all these machine learning models and these are really the languages that computer speak. So, you know you and I we're speaking English right now folks. There's a lot of languages out there Jim and French Italian, you know, and these ml models, you know, the the embeddings from these ml models are are the words right are essentially the way that computers Express or embed if you want to say a particular piece of Auto structure data, and that's why you know, I think that's what we're kind of coming up on an era where you're gonna see more and more folks start to use vectors or start to use these embeddings in their in their businesses.

No, that makes sense after you know, I spent some time cyber stalking, you know the company and you know and just reading the blog and it was it was kind of like a Eureka moment. It was like wow, you know with the you know as much, you know with all of the videos that are floating out there and pictures and none of that stuff is classified well, and you know, when we searched through Google that's kind of that. I mean they have a pretty good job.

We are pretty good way of structuring the data to find, you know, the image based on what you're looking for and it just had me curious or you know, it hit my hip my Without my brain where I was just like wow, this is this is really great. And so I wanted to ask you because you mentioned the vector database the vector database. So what are some well-known use cases for the vector database?

Yeah, so you mentioned video you mentioned video being one of them for for us in particular our Vector database our core Vector database. Our core technology is open source called Novus and you know, we've looking at the the folks who have used movies for different applications. I think we see a wide, you know just tons and tons of different applications.

But I think I'll go over a couple of pretty big ones and also some ones that I like as well. I would say the first big one. I like to talk about is video analysis video analysis and especially as that relates to targeted advertising.

So if you look at you know, you're mentioning video You're mentioning YouTube videos, you know, there's something like 720,000 probably more hours of YouTube videos uploaded per day. That's a lot right that is that is a ton and you know again, I I don't really know what algorithm YouTube uses in the background. My guess is that it's based off of titles or it's based off the description of the video or perhaps even some comments in these videos, right?

So, you know being able to analyze actually what's going on in each of these videos. We can take a very long one cut in 10 second clips and embed all these turn them into vectors and be able to search across all of these vectors. That's something that's really powerful and for you know, recommendation systems or recommended systems is another big one and so if I have products or five ugc that I want to recommend to a user or to let's say customer that's one thing that you can do in a factory database, you know, another big one is actually personalized search or semantic search so you know by another great example that I like to use of that is, you know, electrical engineering and computer science are two very very related fields, right but There's really no overlap there in terms of the text itself, right, you know traditional text search engine might take computer science and group it with let's say social science, which you know, these are actually two very different fields, right?

So, yeah, so with you know leveraging the power of modern amount of deep learning, you know, Vector databases are able to really group to group these things together semantically and enable you to do much better semantic Tech search or personalized search. These are you know, three of some of the big ones You know one of the ones that that I really like it's something that we actually did for Cleveland Museum of Art. And I really like this one in particular because I think it really shows folks the mission that we're trying to achieve which is, you know, Cleveland Museum of Art.

You know, they don't have like a massive. They're not like Google or Apple where they have like a massive these masks engineering team a ton of Engineers and a ton of resources, right? But still they were able to create this really really awesome app called movies called art lens AI.

I encourage you to go check it out. It's really cool where you can actually take a picture of let's say a particular Landmark or take a or you know do a particular pose and you can find Art that's very similar to that particular. Let's say pose or picture that you took and yeah and they, you know the team over there hats off to them.

You know, they they did this and I want to say that we're able to build this and I want to say like a week or two weeks, you know with Bill this as again, they're like, he's infrastructure. You know, it's always one of the examples. I like to talk about because we really try to We really try to you know democratize and make available to everybody a lot of this unstructured data processing.

And I think that really, you know, that particular application really shows really tries. I'm really highlights that really well. That's interesting.

So with milvis they were able to create this app that allows you to just take a picture. Let's say of any structure or impose in that painting and then anything else that's similar to it would would pop up. Yeah, so something similar in there in Cleveland Museum of Arts database it would pop up.

Yeah, so you would be able to do it that's like a reverse image search and you know, I think it it also illustrates. I think the wide variety of modalities that we can Target as well so, you know embeddings. You know vectors themselves being using them as a representation for you input data that allows you to represent.

Let's say You know just tons of different types of unstructured data, right when you're one of the things that we did with compass is they actually had these floor plans and you know, they turn them into embeddings as well and they're able to do this floor plan search and search for houses and build this that way. So yeah, really cool and I'm looking to see what the future holds that is pretty neat. So I mean as a as a consumer myself if I'm on I'll go out and say Wayfair.

com and looking for a particular piece of furniture. Yeah, I see it. You know I see the furniture at another furniture store.

I can just take a picture of it upload it into let's say Wayfair was using you know millvis, then it would up with everything that's similar in their database to it. Yeah, yeah and semantically similar to it as well. And so you'd be able to search, you know through.

The images and not just based off of the product text or titles or descriptions. Interesting. Okay that is neat now to help our viewers better understand how and why are vector databases really important?

Why are they so important right now? Yeah. Wow, that's a that is a loaded question.

Um, so I think you know, I think that's sort of touched upon a little bit earlier, but I'll go a little bit deeper into it as well. I think you see so much more unstructured data out there, you know 80% of all day to day is unstructured. And that number that personage is only rising and we also have these great tools.

We also have these amazing machine learning malls based on based on these detail networks that allow us to represent a lot of our input data and as I want to say that this particular type of you know vectors search itself using vectors as a tool in search. I think that's been around for quite a while. So, you know back when I was you know, right for graduation when I was at Yahoo.

We were actually doing this very very large scale similar research as well. But we didn't have anything like like Nova so we didn't have anything like a vector database and that I think you know development for that particular project. I think if milk was around during that particular during that period of time this is back in like 2015-2016.

I think it would have made everything so much easier. I think people are as we moved to 2022 and coming up in 2023. I think folks are realizing why you know why Vector databases, you know, why vectors search and why using embeddings is so powerful and I think you know Having a database that's going to abstract a way that entire layer for you managing managing and searching your unstructured data.

I think that's something that's really a powerful and something that hopefully most folks would be will be adopting very soon. No, okay, that makes a lot of sense now explain what some of the companies like on your website you guys have some really great clients that are using not this like Walmart Ikea eBay now, could you talk how they're using Vector database to improve the customer experience overall? Sure.

Yeah, so, you know start with eBay and they're actually using so they have you know, they have 50 million products on eBay probably more at this point and they you know, they have a lot of these users as well. And what what they want to be able to do is recommend products to users. They have this sort of product recommendation email, you know lists and they want to be able to recommend let's say based on what I have purchased in the past what I can you know what I might be interested in the future.

So that's really happy eBay's using it there when it comes to the Walmart, you know Walmart uses to actually search catalogs and they you know, they have these language models that they use to embed product titles and the descriptions and the vectors. All right, so, you know and then you know, when they when they do a search when I use it comes into Walmart, and they do is search the descriptions passed in and really embedded by Of language model though, they'll go and search across, you know, 600 million items or actually not really items vectors, you know inside of the vector database and that's how they do. You know, that's how they do large scale and scalable search that way and you know, I think you know, you see definitely I think based on these on these particular applications.

Do we do see product recommendation is being a pretty large one. There are there's as well, you know, as I mentioned your video analysis, especially short videos and personalized search as well. We see quite a few codes.

Okay, that's interesting. Now are there other non retailers that are using Vector databases and if you're allowed to talk about it. What and what other ways are unique ways are this is this being done?

Yeah, okay, so I would say in terms of so retail. I think it's definitely one of the components one of the particular use cases one of the many applications for Vector databases. We do see a lot of you know, as I mentioned for example, you know Compass as one of the folks that's using our stuff as well.

We have short video platforms and it's I you know, I think they are currently are all using the open source version of this, you know, it's really able to everybody for any folks. I think watching now I have encourage you go check it out. But yeah, you know, I think the the different, you know Opera, you know be gopedia there.

They're all using some way shape or form as well. And you know when it comes to you, you know when it comes to sort of some of the lesser-known applications, I would say. This is one that I didn't mention earlier is molecular search.

So yeah, you have these you have again. I'm not I'm not I'm not I'm not biotech. But you know you have these these when you do drug Discovery with AI when you do this sort of this ml driven drug Discovery you want to be able to let's say search for new compounds that might Target a particular symptom or particular disease and you know out of let's say 800 million or close to billion of these compounds.

And turning all of these into turning these molecular structures into a vector and searching across them. That way that ends up being actually a really really scalable and Powerful way to do things. So this is something that we did for again, you know, they're using open source movements.

This is something that we did together with a pharmaceutical company and another one of the really You know, it's pretty out there. It's a little bit different. I think from what most people most folks would imagine about your database can be used for I think definitely it's more associated with machine learning semantic text search these days.

But I think it just goes to show you that the the applications are really numerous. Right and it's beyond just search I think leveraging search to do more complex set of tasks. It's really what effective database is there for.

Got it. Okay. So this really this is really unique and you know to see where this goes especially going into, you know, next year.

What are you what are you for see how big the unstructured data market growing to be? Oh, wow, you know that's wow. That is a great question.

I would say, you know in terms of in terms of a total size of the market for unstructured data in the next couple of years, you know, I could see it probably getting getting up to something like 10 billion quite large. And really I think that's only going to grow in the future. You know, I think again if we look at structured data back in the 60s 70s 80s That actually started off with a really really strong need.

There are a lot of folks who needed to store data in these table-based databases and today there's so much unstructured data out there and you know, there's there's a need for that and that need is only going to grow larger and larger as we move into the future and I really do Envision a day, you know one day at a certain point where folks will say okay, you know, I'll have you know, we don't really plan to replace a lot of these traditional databases. That's not really what we're trying to do. But we're trying to Target a different type of data unstructured data.

I think you know as a move into the future organizations will have databases for both types of data structured and structured novice. Hopefully being with them. Yeah, so sorry no.

No I've seen that's exactly what I was thinking. I could see this coming becoming a necessity overall everybody, you know, everybody's focus is on the structured data and it's well if there's 90% or 80% of the data that is just out there that's not used so much is, you know, just that the possibilities are endless in this point on, you know, there's so much said and done there. No, thank you so much Frank.

Now, let's say if our audience wants to get out there and you know download try or request a demo or do they go to schedule some time with some somebody from your team? Yeah, and you know you can you can always feel free to reach out to me directly either through email. You know, I'm on Twitter as well also on our on our website.

There's our managed service is actually currently in in Early Access, but that is scheduled to come out. You know, that is scheduled to be GA to the everybody later this month or early next month. So Um, so yeah, you know that's you know on our website or you know, always feel free to reach out to me directly.

I'm happy to schedule demo that way as well. And yeah, you know, I'm really nervous of course is open source. It is freely available to everybody on GitHub mild us.

And yeah, I I look forward to chatting with all the different. infected database users out there All right. Wonderful.

Thank you so much for your time and look forward to hearing more about Milpas and zilis in the future. Of course. Thanks Rebecca.

Thanks for having the show. So take care.