Unlocking AI Vector Databases with James Luan, Zilliz CPO | EP 130

0:00I think it's already taking a lot of my jobs as an engineer. I think yesterday I was trying to fix a bug. Like, you were like, I would say it's like, take me a couple of days to fix a bug. With the AI, it runs four hours. Well, I just have meetings like they, they are actually running at the same time. So fixing to themself gave me a very strong advice. So I think it's already partially take off my job. Hi, my name is Dmitry Benicci, and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents podcast. Brought to you by Jotform, and featuring our very own CEO and founder, Ida Kintank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show. Hello, and welcome back to another episode of the AI Agents podcast. We're here today with James Llan, the co-founder and VP of engineering at Zillis. So it seems like you have a master's in engineering really excited to chat with you. Obviously, AI is kind of a thing that

1:03naturally has come out of the world of engineering, and I'm kind of curious, what was your first aha moment when you experienced working with AI? Could it be in 2023, 2022? When did you first kind of have an experience of the LLM and went, oh wow, this is going to change how we work? I think I'm not a typical AI person. I just like the first 10 years of my engineer and during more like building infrast, really focused on how high performance code can be working in a large scale environment. The first, how a moment for me is even before LLM Auto comes out. Yeah, I think back to five, six years ago, we are trying to build a time series set of this, where we try to use some machine learning algorithm to predict the time series status, especially stock price. We actually use like the last 10 years stock price and training our like in-house models.

2:06And it's a small one, not huge, it takes us like several GPU cars and a couple of days to do the training. And it works pretty good. Like we're actually running the like more than 100 days and the prediction actually is more than 70%. So that's the first time where I realize how well machine learning works. And then the second time of course is like when machine, like preacher model comes out when we see bird, when you see like image-night little kind of selves, then I realize using database itself won't be good enough for understand all your data. We definitely need to be complying all those like remote-als together, all those like preacher model together with some infras. That's why we try to build the back of your base. And of course like when chaty-b comes out, see wow, that's something we're like looking into although like preacher model is good enough but it's still kind of like limited on certain tasks.

3:06But the first time when I use chaty-b-t, I think November, 2022, like I am actually the very early age adopter for chaty-b-t and the first time I use it, I know, wow, that's it. Like we need to build something like really make sure that bacteria based can be working together with chaty-b-t. And after one month, I could read a paper I saw RAC and that's where I realized, okay, some people is already doing that. It's really cool. So for the last couple of years, it would be several RAC systems although like it's not super cool already. Like everyone is building the same thing but I think by that time like RAC really shocked me like how good it is, like we just put all the our document into RAD and it works perfect. It works like it's already like at least a one year like a worker like has a coworker has a one year experience on our product. So pretty. Absolutely, no, I totally agree. And I'm curious like just to learn more about this

4:07because I think vector databases are really cool, RAC really cool. Could you kind of explain to the audience a little bit more about what those are exactly and how they work? Because we've had plenty of people be on this, we've discussed this topic on the show. I don't think we've had anybody who's worked at like a vector database company explained those topics to us before. So I think hearing it from the horses mouth so to speak would be awesome. Interesting. Yeah. I mean, like a learner models, they're really powerful. They are trained with a lot of different material so they have a basic knowledge. But the challenge over here is first of all, the knowledge actually stopped on certain amount of time, also in time because each pre-tree takes like more than like Joe Mouse. So when you start the printer, you actually have the knowledge is already like stable. There's no way you can add a new knowledge into it. That's the first problem for learner models. And the second problem is of course it hallucination.

5:09I will say it's a bug, but it's more like a feature for all the learner models because they have some kind of like ability to do reasoning. And when you do reasoning, you're really sometimes, sometimes you're good, but other times you're wrong. Like the way to fix that is you have to give them enough, I would say prompt or contacts or just enough information to make sure that just like human right, well, you try to answer our question. What you can do is just search on Google once you get enough information on Google and you can compile those information into one solid answer, probably you've some kind of reasoning doing model round of search, get a better result. Similar things happen some library model as well. So when we call red, the full name is actually retrieval argument generation. So what we do is we first retrieve all the information, it's not gonna be only vector base. You can also use a graph database or you can just search on general web page,

6:09general search engines to get enough information and combine those information with your prompt, fit it into library models and they can give you better answers. Yeah, I think that's a way we can give a lot more knowledge and also accurate knowledge. Yeah, I really appreciate that and yeah, retrieval augmented generation is, it's something that a lot of things are using now. I think it was like one of the immediate quick gains for agents over, I mean AI agents were practically just built on Bragg for a while and that's still kind of as to an extent. How does that kind of compare with, or sorry, let me stop. So let's go into a little bit of what you're doing over there at Zilliz, right? I'm curious kind of what the background was, how you and the people there got together and founded Zilliz, what was the moment where we got to do something with this, we got to start a company and then we can get into more of the nitty-gritty

7:11of what you guys do. Yeah, which we started long before larger model comes out is more like close to where pretty training models are really dominant, like bird, like rest net, yeah. So we actually saw the opportunity that people have so many ultra-details that there is no way that you can understand that the traditional big data or relational databases. All of our founders actually from database industry, like we built several different databases, so we know how to build a stable and scalable system. And for the first year, we really focused on building a TPU, really accelerated databases, like using TPU to process massive model data. I think it's a smart idea, but the same is by the time we didn't get a very good product market fit, like just because TPU is a little bit too expensive and also the memory is limited. So we saw that that might be not good enough

8:11for traditional workload. But when we take a look into, I'm sure that as things are pretty much different because when we're processing those vectors, we saw a high computation density where TPU can really shine those and since we have TPU background, since we have database background, and when we have a user come to us, say, I want to build a reverse image search, we just think why not should we just build a database for all those images, all those reviews, all these, and make sure that people can easily processing their data as you use our database. That's where the worst first record database comes out. And we named it Mewis, we open source it, so become popular, and that's where the story begins. Yeah, that makes sense. I'm curious, how many founders do you have? We have three. Yeah, our CEO is former Oracle employees.

9:16I'm from Oracle as well, and we also have another co-founder works on product site. Yeah. Got it, okay. I already view like first time founders. I actually joined a little bit later after the company begins, so they added to like, yeah, joined founders before. No, I met like, are they all like, is this their first company that they've like founded? I mean, I would say you has a small set of, like, before we found this company, actually worked for several other startups. This is the time we actually found this company together. Yeah. Very cool. Nice. Just kind of curious about that. And what do you think if you had to describe it in as simple terms as possible? For the founders listening, who might be using standard databases, not using vector databases, how does Zillis Cloud solve the complexity of scaling vector search applications compared to other solutions?

10:19I think the first important reason is that we actually designed for vectors. I'll like many of other databases, like PC vector, like, elastic. Some of them are like, focused on traditional methods, others, probably just traditional search. From day one, we actually designed system for vector search. Yeah. So, computation is very important for us. I'll like traditional databases, like, think about this. Like, future with one single float, couldn't like take you too much computations. The blockers would definitely be on, like, IOs, network. But for us, from day one, we know computation is super important to us. Yeah. That's why, first of all, we leverage a lot of the different new hardwares, with example, TPUs, and also, like, CPU SIMD instructions to help to accelerate the database. I think that's not quite common for all the other databases. The other reason special for us is actually,

11:20we are still new. I will say, like, I'll like the other, like, model when there's like every six months, you get some new vendors. Like, database industry is more like, a lot of those projects that has more than 10 or 20 years of history. So, we are still pretty new, like, because our products are still like five, six years. Yeah. And we actually, papers like after when Kubernetes come out, so the whole new speculative business is actually running inside Kubernetes. With the help for Kubernetes and the public cloud, like, it becomes super easy to just for people to scale. And also, we just store all the data on cloud using, like, storage, like S3. So people don't really worry about how to purchase all the data. Yeah. Interesting. And what do you think is, like, one of the most common, like, industries that you work with customer was? Well, I think we cover a lot of different industries.

12:23We have several really shiny use cases. I would say, for example, RAP is one of the semantic search, image reverse search. We also help people to do multi-modality. Like, for example, drug discovery, like, we do search, even like cross-modality, and even use tags to search, we do, so all these. Yeah. The industry is, like, one of the reasons industry would be working very closely is more like legal and healthcare. Oh, you also have a lot of users from different use cases. Like, one of the largest delivery company users to do recommendation systems, some of them using us to do, like, multi-modality search. So yeah, I won't say there would be a specific industry. Yeah. OK. What are some examples of companies that you've worked with and, like, what does their relationship look like? You've talked about a couple of different ways you can help people out. How do you, like, how does it happen?

13:27I think one of the most interesting use cases we've been working with people is robotics. Like, we actually working with some embodied AI company, like they actually collecting a lot of reviews from their robots. And we helping them to get the fun tuning data set or, like, training data set. By, for example, I'm the data scientist. I'm just trying to tap in. I want to find a case where, on my left side, there is a dot. And on my right side, there is actually a stop sign. And my robot gets through and they didn't stop. So we actually converting those tax information into embedding and tax. And they use that to actually searching the bacteria base. With that, like, prompt, it helping people to find a lot of the reviews. And probably their robots just failed in that test cases. So they're just trying to get more data

14:28to, like, keep fine tuning or doing some posturing on their robots. Sometimes it's just all Thomas trolling cars. Yeah, that's probably one of the reasons that use cases will be working out. So they come to you and they ask for it. But how does, like, implementation work, you know, like, how does the, you know, how does the relationship work? Like, I'm just kind of curious how the whole business operates. So they come, they come to you and you're like, we have this problem and you guys custom fit a solution. Oh, we have, I will say we have two modes. They usually we have it's our managed service. So I think the majority of those users just using the call service. So we don't even need to talk too much if they're like very standard use cases like RAG. So what they just do is actually pico model vendor and users as a record base, just use it as APIs. So it's pretty simple.

15:29And yeah, the other way we do, for example, for the embodied AI coupling is like more like a new use cases and people has a lot of issue on how to find, how to make more accurate. I mean, the search more accurate and how to reduce the cost. For that, that kind of use case. So usually we work close together. Like, sometimes people need to like tuning their models on our side. We also need to tune in a lot of parameters. And yeah, with that use cases is, is you really like we, we spend more time with the little data centers and trying to find some bad cases that are fixed at our end? Yeah. Very cool. Well, that sounds awesome. What's your, I guess, long-term goal as a company? Like what is the, it seems like you're kind of positioned as a leader for enterprise grade AI. And as, you know, AI matures, you can go a lot of different routes. What's like the long-term goal with what you guys are doing?

16:29I think we're just trying to reduce the cost and make sure that everybody can fully utilize their ensure that us. Yeah, because the day we have to start the company, like Wacker Search is now something really new. We know that Google has a strong interest on doing Wacker Search, Facebook by that time also open source, the library called FES, which means they've already put those Wacker Search into production for a very long time. I think things from 2014 to 2015. But back to six years, I don't think any kind of individual developers or even startups, they have the access to use a vector and to understand their data. So it's especially for after data is too expensive. But it becomes more and more important especially on today's AI word. Yeah. So the first day we found this company,

17:30we decided to open source the product because we want more people to use it. No matter it's uncalled, no matter it's in our managed service or it's just on their laptop. And after two or three years, we had to hosted the MiOS for just for the same reason because we see with a lot of AI applications, it's still a lot of overhead for people to manage their own vector searching for us. So why not we just helping them using our expertise and let them to just focus on building their own applications? Yeah. And for the last couple of years, what we're actually trying to do is just reduce the cost. I think we just reduce the search cost by almost 10 times. And with more cost reduction, we actually see data goes even more than 10 times and we've seen a lot of different new use cases. Just because for some of the reason, originally like record search is too expensive

18:31for some of the not super valuable data, but once you reduce the cost, you actually unblocks a lot of different use cases. So I think we'll be like pushing really hard on reducing the cost and make sure that the vector search is actually super easy for everyone to just integrate with their own system because what we foresee is probably in the next couple of years, we see more and more even smaller side of like one person's side of what two people start up. And they could be something really cool and they definitely need a lot of different logo bricks to make sure that they can integrate their own system and we definitely want to be part of it. Yeah. Yeah, very cool. What do you think is the next cool thing that you're maybe working on to improve what you're doing already to the next level? Well, last, I would say this year,

19:32we actually built a cool product which we call as a cloud context. So I would speak for a cloud code. I use it from it actually released, but I soon found our code repo. It actually lost a lot of context because our code basis huge. It has like more than two million lines of code combined with Python, GoLand and C++. So even with very experienced engineers, you need a lot of time to get familiar with the code base. I start to use cloud code and find at the very beginning, it works, it works great, but when I try to test some more complex functionalities, it actually sucks because we don't have any kind of search and just using group is either slow and they also lose a lot of context. So that's why we try to build something like combine our vector search and make it more like an MCP server

20:32and integrate together with the cloud code. And it works pretty good like fix around like 30% of the problem like original cloud code cannot fix. Yeah. So I think next year we'll probably, one thing we want to build is probably another layer of a agent search on top of vector database and helping people to either search right now, people has to build an agent by themselves. But by right now, I think it's more like for every searching for us, you can just build your own agent, not only for code, but actually for general, like multi-modality data and expose it at the MCP server. So people could just either they integrate with the MCP instead of like implement their own. Yeah, talk a little bit more about MCP. I'm curious to hear kind of the improvements that it's led to recently. I think it's been such a whole new frontier of improvement.

21:38Well, I think at the very beginning, I think we kind of like the, one of the first couple of vector bases supported MCP because it works very smoothly with called, and you can easily use it as a two. I think it's become the mask for all the agent workloads. But right now, I'm kind of like a, I mean, a little bit worried about the future for MCP because I see when more people put those into production, they actually hit into different troubles. Especially sometimes agents still got runs more like not just like what do you traditionally do? It's just calling all the APIs and it has like 100% accurate. By right now, we see like with more tools adding into the MCP, sometimes learning models did the wrong selections

22:39and other times you just ignore the tools you have. The more tools you have, the more bad case you're hitting too. So hopefully like with some more powerful data models, this problem can be solved. But I think some of the developers already hitting the troubles when they try to build some like real-world traffic. Okay, what are some of the troubles there coming into? One of the troubles is sometimes just a model cannot help you to pick the right tools to use, especially when you try to have like more than a hundred different tools and expose all those to MCPs. They just pick the wrong tools or sometimes they just prefer not to use tools. So I think MCP is worse better when you have only like 20, 30 different tools. And if you have more tools, you definitely need to think about should have like only one layer or you probably just need more agent.

23:40And first of all, you probably need to separate into multiple different categories and use different agent again to pick different tools. But I mean, that is complicated. And so far, we don't have a real way to fix it and make sure it's a hundred percent correct. And sometimes that becomes headache when you view the real-world productions. Yeah, no, that's fair. I'm curious, you know, MCP to me has a lot of like real-world applications and you know, just general like asking a question of your email inbox or checking out a basic database that you'd use in any project management tool or anything like that. What are some of like the more advanced capabilities that a lot of people aren't maybe aware of that we're unlocked for stuff that you guys do? Because it's probably not as, I think it's still a world where a lot of people don't quite understand what it is and myself included. Yeah, I will say myself as an expert on MCP, but the way we use the MCP is more like another layer

24:43of access to databases. For example, if you try to use a shared database, what you usually do is either you use a SQL connector or you can use a like RESTful APIs or maybe they have their own SDKs so you can use SDK to access a database. So our MCP server works more like another like access layer to your databases and you can use natural language to query all your data. So instead of writing Python code, you can just tell the MCP server, I want to find the most similar result with this image, with a future, for example, the category is a dog or the color is yellow, then we actually converting those natural language into a real expression and running in our database. That's what we use as an MCP provider.

25:47Yeah, I mean, is it almost like, I feel is it new question? Is there almost a, I'm familiar with APIs decently, right? I do a lot of automation work. Is it almost like a unlock for a universal ability to interact with multiple tools? They're like one endpoint almost in a way. Is that a decent way to describe it? Yeah, I think the MCP itself is more like a standard or I would call it as a bus. So when you try to build your own applications, you might want to like have more multiple different data source and the traditional way is what we do. It's just a right to to or write a connector one by one and yeah, write a bunch of emails and pick the tools we want to use in that use cases. But yeah, with the MCP, I think it's first all,

26:49I think it's marching there for every winders, like when I want to expose my service, I right now I have a standard way to interact with all the agents instead of a computer, my own SDK or build my own protocol. And also for all the people who want to build applications, it's just like several lines of code, you just need to specify the target and the agent help you to do like everything else. So they actually save you a lot of time to do all the integrations. Yeah, that makes a lot of sense. And I think I've been using it a little bit. It's definitely beneficial for people like me, but kind of want to talk a little bit more about your opinion on a couple of different things in AI right now. I've once talked about like autonomous agents or once talking about the capabilities of these new reasoning models and a lot of them require like long-term memory to function effectively.

27:51How do you see vector databases evolving to support these like complex agent workflows where a simple retrieval might not be enough anymore? I think first of all, that's a trend. For the last, especially for the last year, we've seen like so many people who are trying to build an agent search. Yeah, I think the reason for that is like sometimes the question is not like why hop? It needs some kind of reasoning. Like for example, if I'm asking for if I said I actually travel to Germany where Germany really have a like famous site saying place and then people, the question might be where you could possibly be. Then I think the place should be the answer, but they actually need two hops. And if you do one search, sometimes you cannot get the answer. You need to like AI need to help us to split all the questions into multiple subquestions,

28:54query for a get the facts and combine all the facts and there might be still some gap between the information you have. Then you probably need another like several round of the reasoning and search. So that becomes a really common patterns. We've seen a lot of those getting to production already. Yeah, so, but I think for like agent memories or long-term memories, I think you surely use record is might not be enough. It's definitely going to be one of the most important part, but still sometimes you need to graph databases at the time. You also need to do a lot of post processing or batch processing on your data. So you cannot just serve your row data into your system because just like human, like when we learn, it's not just like we read the book one time. You also need to do a lot of practice. You need to do some some rise on the information you learned. You need to compress. Sometimes we have to read in notes for like further checking,

29:58but similar sense happens for agents like it isn't. You cannot just build your agent memory on the row data generated for your chat history or for your email, but instead you need to use modern models and some other infrastructure helping you to process it in the back end. And that's something we'll be really focused on and the work with several very large, I won't tell his name, but it's like very large AI application providers like helping them to build their own agent memory. Yeah, that's cool. I really think it's kind of a whole new world for me. And every time I have people like you and I learn so much and I'm not sure quite how a lot of these autonomous agents will kind of perform, like in the long haul. What is your kind of bullishness on agents being able to take over actual jobs? I think it's already taking a lot of my jobs

31:02as an engineer. Yeah, I think yesterday I was trying to fix a bug. Like, you know, I would say it's like take me a couple of days to fix a bug. With the AI, it runs four hours. Well, I just have meetings, like they are actually running at the same time. So fixing to themself gave me a very strong advice. So I think it's already part of partially take up my job. But I will say like it's a end for all the software engineers. We still see a lot of channels like, for example, I think the AI still need to be well designed, like still need a lot of design work and they also need a lot of review work, especially on their very complicated repo like us. Two million lines of code they think just around AI is probably going to fix several issues. But after a time we see like everything mess up, we already have several systems we need to re-factor simply

32:03because like it's actually web coding, we're putting the production very fast. But after like two months, nobody can just maintain because we cannot read the code. So I think definitely going to take a lot of work as a long way to go. Do you think there's like, I've kind of noticed this trend and it's probably the case for coding for you? There's like two aspects to AI, right? There's the output aspect and then there's the synthesization and the being able to ingest and understand information. I feel like we're kind of lacking on the secondary component on a lot of areas if I'm not wrong because the first instance I was exposed to this, we interviewed somebody on the show who was the CEO of a company called LEAI. They're a HR company, basically, right? They'd help you with an AI agent that's A-R-A, sorry, that's HR related, right? And they receive a bunch of different, everybody, sorry, receives a bunch of different,

33:05it's cold, resumes and cover letters, right? Now with AI being mass-produced. And what they're doing is providing an analysis tool, right, an analysis tool for your role in making sure that it's proper and maybe checking whether it was through what they're putting down these types of things, right? You have to combat the output and I'm hearing it from some coders recently. Well, it's amazing how much stuff it's outputting, reading it, synthesizing it, making sure it's working, right? At a better technical level than just like it functions. The refactoring, you just mentioned, kind of sparked something there. Is that something you're noticing? Like the inverse of output versus analysis, the analysis portion needs some work in the world of AI right now? Yeah, I think I would be more like to take AI as some very junior worker, actually a fast worker like for humans, we can definitely like work as fast as AI do.

34:08And the best part for a agent is like, you can do concurrent agent. I would just say some of the smart coders who is trying to leverage five concurrent agent at the same time, they're just a problem on one screen and let it running, then it's switched to the next project. And it's every most very fast. And the best part is they can also like offer you three different output. And you can pick one of your best. But yeah, I think you're right. I don't think AI is pretty good at evaluate all the results. And I would more like think myself as a manager when I work with the AI or more like a pack lead. So I actually gave them a lot of guidance about what should be right, like what is our code convention, how do you evaluate, you need to run all the tests and show us the test results, things like that and make sure that it's not like breaks humans like knowledge.

35:11Yeah. Yeah, no, totally. What do you think is the your favorite thing about being in the sage of AI as like an employee? I know obviously you mentioned coding and just like the ease of time there or the ease of effort there. What is your favorite like thing about being in the sage of AI because it was it's such a completely different workplace than it was even five years ago? Yeah, I think it's the best thing. I will say it's like a save me in tons of time so I can concurrently do multiple different things like not not only about work, coding is everyone's thing. But you would like just think about this like last month I have a travel with my wife to Japan. What do we have to do is a more like a travel plan about 10 days and there is like without the AI it takes me probably at least two or three weekends like going through all the website finding all the hotels and book tickets. But right now with the AI is more like four hours

36:13like I just gave him prompt and let them to double check all the facts and sometimes it can even help him to book all the tickets. So at the meantime while they are like processing it I can still like review the code or maybe watch them away. So basically speaking when you have energy like just give you a lot of possibility to run parallel tasks. And that would be like super helpful I think for startups. So that's also why we're seeing like a lot of like smaller steps they're growing faster. I've been seeing like a startup with 20-30 people but they can get like even 15 million revenues and that's like almost impossible for like five years ago. Yeah parallel task pathing is an insane thing that exists. It's so cool. Like I love it so much.

37:13It's very like for me it's so exciting. I use a lot of different tools at once now. I'm like multi-tabbing, multi-tasking kind of like it's crazy where we're used to be. We're waiting for something else to do work that we told rather than doing the work ourselves. Yeah, I think one of the things I want to mention is like I'm not a person who likes to build slides but for the work reason I have to build a bunch of size every other mouse. So the other good part for EI is like right now I just tried to use a AI to build all the slides actually using the gamma as my machine. Yeah, gamma, yeah. I think it's pretty good like saves me tons of time and yeah that's probably my third part, yeah. We move. Everyone uses decks like that. I don't really like but slides of course I can just throw my old power plants away. Yeah. No, it's amazing.

38:14I hated making decks before and now I can just prompt it be like I want this. I have this service I'm offering. Here's the PDF, here's the info, do it. And it's like yeah, okay, cool. I got you. And then 30 seconds later you have a 15 slide deck and you go, I would have been better than what I would have made. All right, thank you. I'm convinced. Is there any other cool tools like that that you use? Personally, I asked that because everyone's given me a very interesting stuff that they use. Yeah, really I'm, I think except for gamma, the other one too, I'm using very heavily is Notebook LAM. So when I actually reading new concepts, I've twist the old loads into Notebook LAM, let them to search more and then generating some reviews helping you to better understand that it sometimes is all those. So that actually make my understanding much faster.

39:15So originally what you have to do is actually search a lot of information and I have to read through it. But right now like we is with like 3, 2, 5 minutes video, it helps me to better understand everything. So sometimes I also create videos for our product and share it on the YouTube. So really helps a lot of people to understand some basic concepts. Yeah, yeah, absolutely. Well, I think that's kind of it for me just to kind of close things out. Let me know where everyone can find you so they can check it out. Yeah, first of all, you can definitely find me on LinkedIn, like just search for James one. And yeah, I think the first one is me. And I'm also like very active on GitHub. So if you search for me with which is our open source product or if you search for actor based, that's actually the first point on GitHub. I am very active on every issues and the PRs.

40:16So if you're free to contact me and we can also have any like check on this quarter set channel. It's awesome. And then also make sure to go to zillis.com at zillis.com, zilis.com, zilis.com to check out everything vector database related. Thank you so much for watching this episode guys. If you like this, I'd like to leave that this video a like. And thank you so much to James for being on the show. Very interesting stuff. Stuff I definitely hadn't known before and was excited to learn. So I appreciate you making the time. Yeah, thanks so much. Thank you.