Webinar
Delivering the Most Performant Vector Search with Zilliz Cloud
Join the Webinar
Loading...
Jan 2024 Zilliz Cloud Launch
Join our exclusive webinar to explore the latest enhancements in Zilliz Cloud, the most performant managed vector database built atop open-source Milvus. Tailored for developers and technical professionals, this session will delve into the forefront of vector database innovation.
Learn how we're evolving Zilliz Cloud to reinforce its status as a top-performing managed vector database, ready for enterprise production and integrated into the AI ecosystem.
This webinar is a must-attend for developers and AI professionals looking to deepen their understanding of vector database trends and advancements. Join our VP of Engineering in this insightful session to learn about Zilliz Cloud's role in transforming unstructured data and AI applications through unparalleled performance and scalability.
Reserve your spot today and be part of the future of vector database technology!
Thank you so much for joining us for today's session, um,for Zilliz Cloud January, 2024 product launch,delivering the most performing vectorsearch with Zilliz Cloud. I certain need to cover a few housekeeping itemsand then we'll get right into the session. So, firstly, the webinar is being recorded, so if you haveto drop off at any point, you will get the accessto the on-demand version within a few days. If you have any questions, I think someone already didto mention that they cannot hear,but feel free to just paste them into the qand a two at the bottom of the screen. So before we get into the features, uh, I really wantto just, uh, talk about what ulu cloud is.
So firstly, it's really based on themost widely adopted open source vector database,bu we still invest like lots of engineering resourcesto keep making BU better. Then the stable BU version will continueto deploy to new sales cloud. So that becomes the core of business cloud. Um, but on top of that, we actually tryto really just re-engineer and optimize, um, vu. So you are gonna get a better performanceand the lower TCO, um, versus to vis.
Um, the other thing we really try to do is really go, uh,above and beyond the Mel visand just try to build a few essential enterprise feature,which including both more flexible deploymentand also also security feature to, um,create this production ready platform. Another really important thing is, um, we wantto build the unstructured data platformand just help you to connect to transform data. And so we build a few features including pipelines,which is public preview right now,and more connectors and integration. We're gonna talk about that later as well. So basically we wanna build this kind of platformthat is well integrated into the data, uh, dataand AI ecosystem.
Now let's get it right into this. Um, very exciting, um, launch highlights. Um, so that's the pillar I mentioned earlier. Uh, let's start with the optimized bu. So the really the key highlightof this launch is a Cardinal.
Um, Cardinal is a highly efficient vector search engine. So it is designed for our fast approximate nearneighbor search, uh, what we call A NNS. It can handle a wide range of tasksand support a various data formats cutting focus on spendand, uh, sorry, kind focus on speedand efficiency, allowing more user requestswith resource limits. It's optimized for both X 86 and arm hardwareand Cardinal offers a 10 times performance boostof a open source bu this make vector search quickand more efficient enhanced AI applicationand user experience. Another key highlight of this launch isafter a few months be beta, the most anticipated Novus two,drug three is generally available on sales cloud.
So I listed a few highlights here,and the number one is co-sign metrics. And so it's a advanced similarity metrics,especially popular among, uh, NALP, um,co-sign eliminate the need for prior vector nominization. Um, so if, um, developer wantto conduct co-sign calculation right now, it's easier. Then the second real important feature is ranch search. Then personally, I really like this, uh, this feature, uh,James will talk a bit more later about this one.
So this feature actually provide a more precise dataretrieval method for some specific use cases. Uh, it broadened the scope of vector queryingbeyond top K search. Um, it's potentially beneficial for recommendation engine,ensure more relevant suggestion for users. Another really key feature is absurd observe. Simply find data management in updatingand delete data sets is particularly valuable in dynamicenvironments where data consistency and ity are crucial.
Then let's move to the production ready pillar. Um, a really key, um,the engineering team actually spend like quite a few monthsto really build the more granular role-based access control. Um, so with our robust access control, you,we actually provide your like a four clusterand a project role. So that's like for operational permission,like you can manage your projects and billings. And then the even more important thing is we also providethe data layer permission.
So for example, you want to addor view certain datas, um, for the data layer we have threebuilding one customer row. Um, so basically just really help you and enhance securityand the teamwork while ensuring the user have the rightaccess to the right job. Then another really exciting announcement we wantto talk about here is we are available on GCP marketplace. And the even better news is that if you are sign upthrough GCP marketplace,you are gonna receive an extra a hundred dollars credit ontop of the standard a hundred dollars creditavailable to all new users. Okay, so the last pillar is about streamlineour unstructured data.
Um, we're really excitedof like a tremendous positive feedback. We heard from the confluentand Air byte collect connectors we launched last year. Um, right now we're introducing the data Databricks, uh,and Slack connectorsfor easy data migration and transformation. It supports both streamingand batch data import into those cloud and help teamand individual with tasks those updating modelswere uploading data. Okay, now I'm really pleased to introduce James, our VPof engineering for the deep dive of this new features.
James, welcome. Um, firstand foremost, uh, I really want to talk about vu. Uh, so for vu, uh, 2. 3, uh,it's been quite a few months since we, um, released itto the VU community, to the zills Cloud. Can you talk about why we don't have the feature priorityover there between VU and Zeus?Yeah.
Um, first of all,we are like very conservative about, uh,publish any new features to zes Cloud. Uh, the reason is, uh, we want to keep our, uh,little called, uh, very stableand also compatible with all the, uh, previous featuresand many of the newest, uh, features, uh,release, uh, in embed. So we want the users to open source usersto try all those features and provide us more feedbacks. So those features are probably in very early stage. So we, we might do some adjustments on both, uh,the interface and also the, uh, restrictions.
Uh, but on the other side, we want, we hope that the,this cloud can, uh, users can use more stable features and,and interfaces. So, uh, because we believe that compatible is, uh, goingto be very crucial, uh, for all the users in production. So, uh, that is our key goal to, uh, to keep around threeto six months, uh, delay compared to our open source newest. Yeah. Uh, this means that, uh, for all the call users,you might need a little bit patient, uh,with the new features.
Uh, in fact, we're just going to go in, uh,to release meals 2. 4 soon this month, uh,including some more exciting features like, uh,more embedding support, like, uh, scale, our inverter index,like Sparsing embeddings,and those features we are going to be in, uh,those called maybe mid middle this year. So also, I, I think, uh, uh, oneof the reasons we can do more restriction to sus isbecause, um, this not only, uh, ensures sustainability,but also guys, the users who wantto use this product in the, in the, in the right way. Uh, for example, uh, we limited numberof collections created on the cloud, uh,because we find that a lot of users who want to build their,uh, rack applications, uh, what the user try to do is, um,building hundreds of sof collections each collection for one tenant. Yeah.
But, uh, this put significantload on the, on the cluster. So in fact, if you want to build multi-tenant applications,the, uh, recommended, recommended way to do that is usinga feature called Protection Key,where you can just support millions tenants, uh,in one collection. And when you, when you, when you do search, you can like,uh, filter out the, uh, the, uh,tenants you want real quick,and that gives much better performance as well asthat's the burden of the system. Great. That's really great tips.
Uh, if you have any questions regarding howto use those cloud, please do not hesitateto, uh, reach out to us. Our support team. We're definitely very happy to help you. Um, then let's talk a little bit more about the mailbox 2. 3.
As I mentioned, one featureof funnel super interesting is Range Search. Not every vendor actually has this kind of feature. Can you just tell us a little bit more about Range Searchand what are some of the use cases best for?And then maybe show us a little bit demo abouthow people can actually leverage this like a power. Sure. So just said that, uh, you offered featureinto is research.
Uh, mine not, I, I, I would prefer for afterbecause, uh, uh, I mean a lot of users also like inspectfor, uh, after. Yeah. But for Ring search, uh, I mean, uh, mostof the Vector database, right?Uh, for now we, what we do is called a, uh,a n approach in near, uh, nearest neighbor search. And, uh, usually, usually, uh,what we do is perform a top key queries. So when you give a query vector embeddingand the database will return you, uh, top key,most similar nearest neighbors.
Uh, however, in many of the use cases, uh, we don't know,like, uh, what is, what exactly number of key we are lookingfor, but we just wantto retrieve all the data in certain distance read. Um, this need is, uh,particular common in anomaly detection, uh, in someof the image, uh, research. So, um, moremoreover, like the top key mode, um, the volumeof data you want to, uh,retrieve from certain region might be very large. Um, if I said I want to retrieve all the vectors which have,um, distance between 0. 5 to one,uh, cosign metrics, right?I might got a very large number of entities.
Yeah. So it might exceed the limits of, uh, single RPC can handle. So in such Case Cent can be used together with, uh,iterator features. Uh, so you can just, uh, iterate, uh, the,all all the data one by one, make sure that, uh, it won't,uh, break the RPC limit as well as it,it won't break your memory. So through our testing, uh, with the iterator, you can, uh,scan 10 million, seven 60 dimensional datas with less, uh,with less than 10 minutes.
So you can also use this feature to export all the data, uh,all, all, all of the cluster. So that is going to be really useful for mostof the Vector database. Once you put in all your embeddings, uh,you get vendor locking, you can, you, there's no waythat you can pull off your vector embedding cell,and, uh, with researchand also hre, it definitely helps you. So what I'll do next is I will give a quick demoabout our, our, our research. So let's start.
So actually have, uh, build a, a quick demo. So in, in this demo, I have a bunch of, uh, datas, I have,uh, men py, uh, which is a Python file that, uh,I'll ex I'll execute each, each step. So the first thing I will do is let's quickly seewhat kind of data like I got. Okay. So, uh, actually I have, uh, uh, 10 K datasand, uh, with, uh, PK Field from zero to 10 K,uh, with, uh, some descriptions, uh, with also with, uh,width and height, which is, uh, inte numbers.
And also, uh, for sure, there's a full 32 lectures. So I, I also, uh, start, uh, meals cluster in the,in the, in the backend. So, okay,so the first step I will do is to runa preparation, uh, collection. So what we do in this demo is we connect to mul,we create a collection, then, uh, we create, uh, create,create a collection name demo, s and tt. We insert, uh, 10 K into this, into, into,uh, this cluster.
And, uh, then next step,what I will do is, uh, traditional search or, uh,or just, uh, a top key search. Oops. Oh, I need to change,change the name. Alright, so let's see what we got. So on the top, uh, I, I got the most similar resultwith a distance zero, uh, 0.
9 0. 9,which means it is almost the same, on the same, uh,into this, right?And, uh, if, if you go through it, you got, uh, some other,uh, IDs for, for this one, you got, uh, uh,cosign me metrics, uh, similarity to be 0. 8,which is also very similar. And, uh, uh, in, in this demo, I gotaround a hundred result. So if we take a look at what we did in this search demo,so we, what we did is we connect,we create a collection, we do a search.
In the search, we have a limitation of a hundred. We also specify a bunch of afus. Yeah. We use a co-sign metrics. Yeah.
And we also at the, at the end, we print outor we print out the recall, right?So that's the how, what, what do we call that? Oh, great. So the recall is 1. 0. So everything is great. So that is for traditional search.
So what will happen for ring search,let's run the other test. I have my correction in as,So we still get, uh, recall to be 1. 0. Uh, and if you, if you take a look, look at this, I'm tryingto search between, uh,range from 0. 75 to 1.
- Uh, no surprise, you still get most similar, uh, resultwith IV equals to zero, right?And the similarity is the same as what we do as top K. And if you screw down, then you'll see that you get a lotof, uh, uh, other IDsand, uh, for the, for the last result, right?The distance is around 0. 75. That is a range.
I specify if you take a, if you take out the code,So there is a site D uh, site difference. So when we do search instead of, uh, specify top gate,we also specify a range, uh,which is in the poems, right?We said radius equals to 0 7 5, which means, uh, uh,the least likely, uh, the,the least likely data should have a similaritymore than 0, 7, 5. So that is what we call it a research. Yeah. And research can be also used together with few drinks.
So under that case, what we can do is we can,we can do a filtering searchwhere we can have another condition, this equal, uh,smaller than 500. If you took a look at the last result, right?You'll find some Ws larger than500,000, uh, 500. Uh, for example, if you look at this one,you definitely see that the W is five, uh, 5, 7, 8, right?But what will happen if you run a Turing search with, uh,rings together,then you'll see the result is much lesser thanwithout the Turing, right?And if you take a deep look into all the result, you'll seethat, uh, all the, this is actually less than 500. So, and the most main part for the ring searchis you can work, uh, it,it can be used together with the literature. So let's say when I run the demoand they give me the first chain result, uh, I,I skipped eight of them, uh, the, uh, the, uh, top, uh,most likely one still, you got, uh, uh, 0 9, 9 similarity.
Uh, and the top, the top 10, you got 0, 7, 5, right?And you can still continue. And, uh, the 11th is actually a slightly, uh,far further than, uh, the, the 10th. And, uh, you get another 10 result,and if you keep going,you got another 10 and you keep going. Okay? So you got all the results. So with these iterative features,you can get a very large data set withoutworrying about your client get OM or opposite limitations.
Yeah. Great. That sounds very exciting. Um, then let's talk a little bit, switch the gearsto talk a little bit about the Cardinal. Um, so for Cardinal specifically, I would loveto hear the story behind, um, we already got nowherefor Novus.
Um, what made the engineering team thinking, Hey, we needto build another new search engine just specific with Zeus. Yeah. The, uh, the initial ideabehind constructing our Cardinal is, uh,to achieve a unified Vector index, uh,because, uh, email is, uh, at statusthat we have a library called no, where we tryto integrate all the sort, uh,vector search libraries like Fast HW disk to alignall functionalities with this, uh, vector index. Uh, we did a lot of modification on those indexand, uh, it spend us a huge amount of timeand also, uh, to, to maintainall the vector index have the same functionality. So, uh, one day, um,after, uh, release that we find that we, uh,we are not aligning certain featuresthat like this didn't support, uh, the featuring features.
So, uh, we got, uh, we got really struggledand we, uh, began to build a, our own index for,to support various the, uh, vector datatypesand also, uh, various, uh, datatypes like 4 32,like BF 16, like FP 16 binary vectors. Uh, we want to build a vector index that, uh,support different index tab, like both graph based indexand also a VF like inverted file index. Uh, we want to build a indexwith different ation algorithms, sq pq, uh,with different storage, like in memory on this,even remote storage, like S3, we want to build, um,our vector index with, uh, GPU with GPU support. So, uh, we, we did a lot of refracting on only, um, um,the new index make sure that, um, every partof the index is, um, profitable. And that is the beginning, sorry, of, of, of this cardinal.
Great. Then can we talk about,from the customer's perspective, what are some of the, um,good use cases, um, for Cardinal?I understand it's not only about speed,it's also about efficiency. Uh, do you mind talking about that more?Maybe I can show the architecture we can talk about justlike, um, align with the architecture. Sure. So as I, as I just said, so Cardinal is, uh,actually very, very plugable if you,if you took out the architecture, right, uh, they haveplugable content, uh, uh, colonizers, you havedistance calculators, you can, you can have co-sign metrics.
Uh, you can have L two metrics, you can have different kindof, uh, customized metrics as well. So it also, uh, split all the, uh, major part of, uh,uh, a index. Uh, we have a plugable index builder,which means you can build all kinds of different index. We can, we can have index, we, we can have,we can also have different searchers. You can use different algorithm to search on the same graph.
Yeah. So, uh, since it's going to be very, very flexible,and, uh, the interesting part is you can do a lotof different combinations. For example, uh, Hand W Index, they don't have any quantitations, right?But the fast probably has the most, uh,powerful quantitations, uh, algorithm. So what, after we combine them altogether, uh, it allows usto do quantitation on graph index. Yeah.
On the other side, since we have all kindof implementations, um, we can dynamically choosebetween using a graph indexor using inver index based on a lot of query conditions. If the top key is very, very large, that maybe use, uh, uh,graph index is not very efficient. So we peak FF index. If the top K is very small,then definitely choose graph index. Yeah.
And, uh,you can also implement different comparison algorithms. Um, we can automatically peak index between parameters. There is actually a powerful tool we call it as auto index. So it helps you to tune all the parameters, um, without,uh, adding like a specific knowledge from your side. It, it just did all the not tuning work for you.
Yeah. And thanks to all those opt, uh, optimizations, uh,our call, uh, with this release, uh, our high performance,uh, tire has achieved the performance improvement of,of more than 10 times compared to the open source solution. What we have, um, for meals. And, and the most exciting part is also it reduce 50%of our memory usage. So, which means for each one CU performance, uh, instance,originally it can only store1,000,007 60 dimensional datas,but now it, it can support one, 1.
2 million lectures in yc. Yeah. And on the other side for, uh, our disc index, uh,with, with this cardinal, we also achieved the 50%performance in improvement compared to our first generationof, uh, uh, those cloud index. Yeah. And model over kind has, uh,ex extensive optimization for scatter filtering.
Yeah. So, um, MU is actually the first, uh,database introduced pre pre, uh, which has, uh,almost become a standard feature in all accurate databases. Yeah. Building on the basis of perfusion, uh, we implement,uh, a lot of optimizations in Cardinal, especiallyfor the connectivity for the graph index. Uh, and we also have a specified optimizer, uh,select a different filtering algorithm based on differentdata ologies, different, uh, query conditions.
And this leads to a strongfiltering performance improvement. Yeah. So if you want to know more details about, uh,performance, maybe you want to try it by yourself. We actually have, uh, open source, uh, library called, uh,vector B bench. So, um, we have bunch of test cases.
You can run this test test cases, amount of the, uh,mainstream vector dbs compare their performances. Yeah, it's very exciting to hearthat the journey continues. We're still, we continue to build better things on topof cutting to benefit all our customers. Um, then let's move to the role-based access control. I did a little bit of research myself, um,and I realized those cloud, um,role-based access control is the most advanced comparedto any other vector database vendor on the market.
We have the four row on the con, uh, like control planeand another four row on the data plane. So James, I just really wantto hear from you why the team think this is a very importantthing to build since we, I think we probably spend thepast three, four months to just like do a bit by bitto have such a granular, uh, RAYeah. I mean, I, asthe status securityand being enterprise ready is our first priority. So, uh, this applies not only to our cloud service,but also for the all the open source users, uh,for all the open source users. We also recommend them to enabling TLS, uh,user authentication and also our back.
Yeah. Uh, for our cloud. I think it, uh, we, we also spend a lotof time working on all the, uh, enterprise ready features,all the complaints, all the, uh, data security features. Although I know for many of the A IGC developers, uh,currently, we might prioritize to getting their,their logic to work. Yeah.
But, uh, I mean, data security is still going to bea primary concern. Yeah. This is especially true when your vector, uh,data can ally be reverse eng engineered to, uh, create,recreate some sensitive or, or original informations. Yeah. Further your, your risk of for data.
So, uh, REC is going to be one of the key featuresof our, uh, data security. So, uh, this call supported roles such asadmin regret and also read only. Yeah. Uh, it helps, uh, usto management all the internal users viaenterprise, uh, admin. Yeah.
And, uh, first of all, we, we plan to introduce ldap,uh, authorization and also, uh, customized role featuresfor our, um, enterprise tier, so, uh,for our next major release. So, uh, stay tuned for those who haveSEC security re, uh, requirements. And, uh, you should also look at our BIOC, uh, this were,uh, this release allows youto store all your data in your own VPC, uh, with, uh,Zillow's only deploying our control point in our A PC. So this approach, maximize your security of your data,ensure that, uh, all the data remains in your control. And, um, this is under the, uh, most strict, uh,privacy protections.
Yes. We're gonna announce the BIOC bring,bring your own cloud soon, so stay tuned. Uh, lastly, let's talk about theDatabricks Spark connectors. Um, James, do you want to just share your screen?Let me stop sharing first. Sure.
So, uh, let me quickly introduce the, uh,SPARK connector. Yeah. We, uh, actually provide seamlessly integrationbetween Spark and combining all the data pro processingand ML features of, uh, spark. We, uh, with all the, um, vectors data storageand search, uh, uh, compat, uh, compatibility of s the,the integration actually enables various interestingapplications, uh, for example, uh,in efficiently bulking their data into, uh, mules, uh, movedata between multiple mules or like some other storage Yeah. And analyze the data in mules with, uh, leveragingthe spark and lead.
So I can give a quick demoof what, uh, spark Connector can do. One moment. Yeah. So, uh, what if you are a developerand you try to, uh, search on all your structure data?So without this, uh, smart connector, uh, what you haveto do is, first of all, you, you haveto run Spark Job, a Spark job. Uh, what they did is read out yourafter data, doing pre-processing, um,put all the data into an embedding model, and thenafter embedding, you just, uh,store all the data into another storage.
Your's going to be S3 or your local storage. So, uh, from there, then what you haveto do is you have to, uh, write some code, uh, put all the,uh, datas and push them to a Vector database,and the user, then user can run a search on the Vector DBwith our new, uh, spark connector, uh,since become much easier. So, uh, spark directly can run some mul, uh,utility functions that will directly, uh,put your data into other open source mul, as wellas your results code, uh, with very simple code. So I'm, I'm just going to give you a very quick demo abouthow we can run a Spark connector. Yeah.
So, uh, what I'm trying to do is, uh, tryingto dump some data into, uh, MySQL, uh, clusterand migrate all the, uh, data into. So, uh, the first, uh, there's, that's a special I do,uh, spark claim, uh, try to, uh,write some data into a data frame, right?And second stage, uh, I read, uh, I read the data stream,uh, the data frame into a my,and the read them out at another data, uh, data frame. Okay. And the critical part is here. So with just, uh, maybe 20 lines of code, uh, you're goingto able to run the inbound of all the order datasand put into ware.
Yeah. So first of all, you, uh, uh, decide what kind of, uh,tokenize you're going to use,and, uh, then the,that will transform the original MySQL data frame. So next step is, uh, you're going to, uh,specify your embedded models. So now, uh, what we're using is a word, two lecture,and, uh, tokens is our in input. And, uh, our output is, uh,actually a 1 28 dimensional vectors.
So after this, uh, transform, right?You get another, uh, data frame with all your, I embed. Yeah. And, uh, the last, the last step is, uh, to load this,uh, uh, mill data frame. So, uh, in the, in, in this data frame, you have to specifysome of the, uh, key pounds, like, uh,what is the mills host, what is the mills part, what kindof connections you going to load, uh,what is actually the vector field name, vector dimensions,uh, primary key field of this, uh, um, mural, uh,of this mul collection. And then you do a save.
Then what happens is spark, we goingto pull all the data out of MySQLand do embedding, put all those embedding datas into mules. Yeah. Uh, with, with this functionalities is not going to,not going to be only work with, uh,spark, open source spark. Uh, you can, if you are Databricks user, also,you may have a lot of datas in their platform. And, uh, they, we, we also announced, uh,integration with Databricks.
So, uh, with just the te of code, you can pull allof your data from, uh, Databricks, do all the in embedsand, uh, analysis to find all the inside of your data. And what's more with, uh, the Databricks connectors,you can do a lot of analysis on your vector data. Uh, for example, if you can do, um, distributed clusteringwith, uh, uh, spark ML leads,or if you have found you are anomaly that vector data, yeah. Spark can help you to do a lot of, uh, uh,distributed computations while meals can help you to, uh,be a scalable, uh, storage. So both muland, uh, sparkle gonna be, uh, work together very closelyas well as our results called and Databricks step.
That's awesome. And, uh, we have, James,my understanding is like we have published,and this called as well as the, uh, notebook,notebook on the website. So anyone who's interested can check, we can send you the,uh, the, the link in the email as well. Also, if you wanna check, um, thenbefore we close, um, James is just like curious, I want to,you already talk about some of the exciting, uh,things on Cardinal and the security front in the future. Um, but how are you gonna envision, uh,those cloud in the next six to 12 months?What's, what are someof the exciting things that will happen?Okay, so, um, I think the, yeah,I think the first priority is going to be reduce the cost.
So we aim to significantly reduce the cost of, uh,vector data storage in the next 12 months targeting2, 2, 3, 4 decrease. Yeah. With the, uh, continuous growth of a IC applications,uh, I think there will be a explosive increase in the, uh,volume of vector data by this year. So, uh, cost is going to be very, very critical for all the,uh, users as well as all the EIDC application developers. Yeah.
And, uh, secondly, uh,we are working on those call service. So, uh, increased response to the enrollment, ease of, uh,all the applications. Yeah. Uh, they,they have all the similar requirements, right?multitenancy, um, there are a lot of, uh,hot datas, but, uh, there are also like a lot of, uh,worm tenants and coten. So how we can separate all the datas they want query, uh,they want, uh, elasticitybecause, uh, their applications going very fast.
So we have to, uh, release our new service tier. So, uh, this tier will be, uh, designedfor offering fully on demand billingfor both queries and storage. So accommodating all the need for, uh, rack, uh,application designers with millions of tenants. Yeah, we can, we, we also have automatic scaling,uh, capabilities. Make sure that, uh, it can, uh, adjust your, uh,resources needed for, for your workload.
Yeah. And third, I I will say, tooptimize the search quality. So in all the rack applications,search quality is actually very, very critical. So those call is, uh, set to introduce features suchas sparse meetings, such as heavy research between sparseand dancing meetings, uh,and query rankings, so those can help usersto find more accurate results. Yeah.
Uh, last one I would say is, uh, actually integrationand, uh, either of use, uh,we have launched this called pipeline, so designedto assist users in performing a, a series of, uh, um,operations such as data chunking, embedding,search and ranking. Yeah. So to facilitate two data in that out,uh, we are going to support more features in those cloudpipeline, including like multi model search, uh,including more user defined functions. Yeah. So moving forward, uh, those cloud pipeline, uh,we'll be working together, uh, with a lotof open source projects as wellas the other source projects, um,achieving a seamlessly data, uh, integrationbetween those called and oth other data providers.
So, very exciting news. Um, please to stay tuned, we definitely keep you updatedwhenever there's such a new exciting feature get out. Um, that's all the question on my end. Uh, is there any questions, uh, from our audience?You can just like pass your question in the qand a, um, box so James can tryto help answer your question. Let's give a few minutesand to see if anyone want to raise hand.
Okay. There's a question. Uh, how is those comparingto other manager solution like Pine Con or V Deviate?Oh, that's a good one. Yeah, that's a good one. Yeah.
And, uh, that's also my favorite ones. Yeah. So, uh, I'm not sure, like, which, which partof comparison you're looking for. Uh, I, I can give you some, uh, uh, general ideas. First of all, like, uh, uh, we have open source project, uh,ware and it has been running for quite a while.
So, uh, for those cloud we build down fully,fully on the open source solution. So for sure there is no, uh, vendor locking. Right. And also, uh, since you are asking for performanceand cost, so the easiest way to do that is you can,you can directly, uh, running our Vector V bench. So it has, uh, performance com, um, comparison for, uh,both Pine Coin for Vivid Cauldronand some other like major retro dvs.
So from our test result, I, I'm, I'm, I'm, I'm not sayingthat, uh, uh, it is gonna be a fair comparisonbecause it's, uh, the data is from our, our benchmark. Yeah. We, uh, currently we are actually, uh,much faster than all the competitors. So, and also since we offers, uh, disc, uh, disc, uh, uh,solutions, the capacity Terra,so we are not put all the datas into mid memory,but on the other side we put, uh, most of data on the disc. So it also, uh, we save our, our,um, memory cost.
So if you are building some A ITC verificationsand you are not looking for very, very fast, uh,performance, I will say that, uh,capacity optimizing things can help youto save a lot of money. Awesome. Uh, I also put the link of, uh,vector DB bench in the chat. Um, so feel free to check by yourself. Let's see if there's any other questions over there.
We'll give two more minutes. Yeah. And, and I know one thing I wantto mention is also about enterprise Ready. So that will also be a big, big difference because, uh, uh,because, uh, when we talk about Vector two db, right,it is not, not every Vector databaseis actually a, a database. So, uh, when we, when we mention database, what, uh,we expect is going to be, it has a lot of management tools,it have visualizations, it helps, uh, it,it has data backups, have migrations,have further recovery, those kind of stuff.
So when you pick Vector DB as all those are likevery important, uh, uh, evaluationas you have to do, right?Uh, and also functionalities. What, what kind of features can we support?Uh, I think where,where every vector base may have filtering searchmay have meta filterings,but the performance gonna be very, very different. Yeah. So you have to ask a,a lot about all the functionalities as wellas some enterprise features thatthat's gonna be very important. Yep.
Um,and since we don't have more questions in the chat, um, withthat, I just really wanna say thank thanks James,and to, to be here, try to share with us the whole audience,like whatever the new features, uh, in Zillow cloud,and we look forward to our next launch. Thank you. Alright everyone. See you guys.
Meet the Speaker
Join the session for live Q&A with the speaker
James Luan
VP of Engineering at Zilliz
James Luan is the VP of Engineering at Zilliz. With a master's degree in computer engineering from Cornell University, he has extensive experience as a Database Engineer at Oracle, Hedvig, and Alibaba Cloud. James played a crucial role in developing HBase, Alibaba Cloud's open-source database, and Lindorm, a self-developed NoSQL database. He is also a respected member of the Technical Advisory Committee of LF AI & Data Foundation, contributing his expertise to shaping the future of AI and data technologies.