Name: AI search in action: how Adobe, Cohere, and AWS turn complex data into business impact
Uploaded: 2025-10-29T17:01:05.023Z
Duration: 42 min 26 s
Description: AI search in action: how Adobe, Cohere, and AWS turn complex data into business impact

Transcript for "AI search in action: how Adobe, Cohere, and AWS turn complex data into business impact": Okay. Let's kick off. So my name is Heather. I am, run our webinar program here at Cohere. And before we get into the content we have prepared today, I just wanna share a couple reminders so you can have the best experience possible. So first of all, sometimes, when you join, the browser will mute the speaker. So if you can't hear us, there are some buttons at the bottom of the screen. There's a loud speaker button, and a help button. And if you click through those, you should be able to turn on your sound and hear the speakers. And then over on the right hand side, you'll see there are a couple chat couple tabs in the upper right, and there is a chat tab and there's a q and a tab. And so we welcome you to put any comments into the chat tab. We hope that there's some really lively conversation there. And then we ask that if you have any questions for our speakers that you put those in the q and a tab on the far right. That helps us stay organized, and we'll have some time at the end of the webinar for live q and a. And, also, please put in your questions throughout the webinar. It prevents us from having a flood of them at the end, and that way we can get through more of them, in an organized fashion. And then in between the chat and q and a tabs in the upper right, there's a docs tab, and there are some additional resources that, are related to the content that we'll discuss today. And finally, everyone who is in attendance today will receive an email with the recording. So, please look out for that after the webinar, and you can watch again, share with your colleagues, all that good stuff. And with that, I will turn it over to Nick Morales to get us started. Thank you, Heather, and thanks everyone for joining us. We're really excited about today's panel, and demo with Adobe and AWS, you know, focusing on real world applications of Cohere AI search capabilities. As Heather mentioned, we wanna hear from you. Please drop questions in the QA section, and we'll make sure those get answered. Let's jump right in and start with some quick introductions. My name is Nick Morales. I'm head of customer experience and partnerships. I get to work every day with our customers on their AI transformations journey. And I'm joined by some of the best experts in the industry when it comes to search and enterprise technologies. So let's start. Niels, can you introduce yourself? Cool. Yeah. Happy to be here. Niels working as an IT of AI and search at Cohere. And, yeah, I've been on a space for search embeddings for quite a while. So started it in 2018 when there was maybe, like, three people in the world interested in embeddings and vector spaces. And it has been, like, a fun ride. A lot of cool models, innovations came out, and, like, yeah, very happy that Adobe is adopting our models. And, yeah, there is so much potential going forward, looking forward to it. Thanks, Nils. We look forward to hearing from you, later in the webinar. Vikram. Hey. Hi, everyone. Vikram Matoff, principal solutions architect in the AWS strategic division, based out of San Francisco area. And, I've had the opportunity to work with, like, bleeding and cutting edge technologies like Adobe over, like, the last several years, based out of Silicon Valley. Thanks, Vikram. We really appreciate the partnership with AWS. Abhije. Hey, everyone. Great to be here. I'm a senior engineering manager at Adobe. My team and I are driving agentic gen AI and AI capabilities in Acrobat. It really helps our customer kind of drive the maximum value they can from Acrobat. Thanks, Abhijin. And we'll start with you. So, you know, across enterprises, there's so much valuable knowledge that's just blocked away. It's buried in PDFs, emails, images, documents, unstructured structured, you know, across legacy systems. And, you know, now we're seeing that AI can finally help surface that information and turn it into action. But with the rise of AI agents, enterprise search is no longer about just finding the documents. It's actually about enabling intelligent decisions and workflows. So, Abhijit, can you kick us off by sharing how Adobe, one of the one of the most innovative enterprises in the world, is using coherence models to power AI search in the real world? Hey. Great question, Mike. We do know that a lot of information is captured, but it becomes very difficult to extract knowledge from them and kind of take actions there. So Adobe has been on this journey to enable our customers to do that. We started with the AI assistant solution that work with single documents of multiple formats. And now we have come out with a a solution that's p f spaces that helps you collate a set of documents into a collection and kind of work on all these actions fairly decompressing the cycle between, comp comprehension creation and collaboration with others to share that, those actionable items. And in this journey, Cohere has been an amazing partner with, for us. Their emitting models is an amazing technology that we have really kind of leveraged to deliver these values to our customers. These models have really enabled us to not just establish trust in our AI solutions by generating, citations, attributions that users can validate that the answers they are getting. The conversations they are having with the AI assistant is grounded in truth. And in addition to that, as the technology has moved forward, we have been able to leverage the technology to support multiple languages across, a large set of, population ranging from, French, German, Italian, Brazilian, Portuguese, Japanese. So a huge range of languages. So multilingual was a technology that we clearly adopted. And then going forward, we do know, in addition to text, a lot of information is captured in figures, images, charts. And that is where we really leverage your multimodal capabilities to, generate those, to unlock those, knowledge that's, kind of trapped in those formats. And that really helped us move forward, our solution. So it's it's it's an amazing solution. And and with that, I think the model was another area that really helped us accelerate the the velocity with which we were delivering this these capability. It's been an amazing, amazing, collaboration with, these, collections. We realized at certain points, the collection are reaching a a level where the context windows for LLMs were running short, and that is where we started to leverage the reranker models that are provided by Cohere. And and it really helped us make sure that the quality of our answers is amazing for our, for our users. So, yeah, that's how we have been leveraging Cohere, and, it's it's great to see Cohere keeping up with with the the trends and in some cases, really driving the frontiers of of this. Thanks, Savage. You and your use case, you know, require, you know, requiring the the performance, the multi linguitily linguistically, the multi modality. You know, can you talk a little bit more about, you know, what about Cohere solution stood out as you were evaluating options to support your use case? You know, how are you getting the most value? So, again, when we are evaluating different, we're able to see that, this was coming forward. There were different dimensions we were evaluating the solution. Those were based on quality, the performance, the cost associated with it. And in addition to to that, one factor that we really looked at was how well does this, how how well are you able to drive the technology forward. So we didn't want to get locked into a state of time, where you validate the current state of technology in terms of quality, performance, and cost. But we also were thinking about, hey. As the technology is evolving, are we able to get all those capabilities and and really, shown? Yep. Thanks thanks, Abhishek. And so, you know, would be great to, Vikram, transition to you, and and, Vijay, we'll see the demo later today. So really excited about that. Vikram, can you elaborate more on the AWS cloud and how the role it played in facilitating this this partnership and this use case? Absolutely, Nick. So at AWS, our role in this partnership is really about, enabling choice, flexibility, and scalability for customers who want to build intelligent applications like Adobe Acrobat AI, using Cohere model in this case. So Cohere language and embedding models are available to customers on Amazon, SageMaker and Amazon Bedrock. On Bedrock, customers can access Kuya's model through a fully managed serverless API, meaning there's no infrastructure to manage and developers can integrate AI directly into the workflows, with just a few API calls. Bedrock also has features like, basically, augmented generation, drag via knowledge bases, our guardrails for responsible AI, prompt management, model evaluation, to name a few. And, very recently, we also released something called Amazon Bedrock agent core. This is our modular set of capabilities to build, deploy, and operate production grade agents securely and scalability using any framework and model. So all of this makes it easier for companies all the way from start ups to enterprises to build trusted context aware AI workflows, like what Adobe has built. And, then on the other hand, we have Amazon SageMaker. This gives teams the flexibility to fine tune, deploy, and monitor coherent models in a fully customizable environment. Ideal when enterprises want to deeper control over performance latency or model adapt adaptation to their proprietary data. So, you know, going back to your question around partnership, what's powerful in this partnership is that Adobe gets to leverage both the innovation velocity of Cohere's model and the security, scalability, and compliance foundation of AWS. So truly a better forget the story, models from Cohere, use cases from Adobe, and a secure, scalable cloud platform from AWS that brings it all together. And, the design pattern, if you can call it that, can benefit any other enterprise. Absolutely. And then just piggybacking off of that a little bit, we hear a lot about data security and and privacy and the strict requirements around that. Can you talk a little bit more about that in regards to the deployment strategy? Okay. Great question, Nick. So it's one of the first things customers ask us when adopting generated AI on AWS. So specific to security, security is our top priority. Our infrastructure and services are architected to provide customers customers with the most secure cloud, computing environment available today. AWS has meet the unique security requirements of even the most, sensitive workloads, including governments, financial services, and health care. As for AI workflows in AWS specifically, our customer data is never used to train the base models, and it always stays within the customer's AWS account boundaries. But so versus like bedrock. Any data sent to a model is encrypted in transit and at risk, and no data is ever stored or you reused by the model provider. For customers who want even title control, Cohere models can be deployed in private VPC environments or with SageMaker endpoints so that data never leaves the organization's secure parameter again. So this approach allows enterprises like Adobe to innovate confidently, combining Kuya's cutting edge AI models with AWS's proven security, compliance, and governance controls. Yeah. Awesome. And, you know, for for the developers who are on this call, I mean, thinking about how do how do they leverage this great embedding and re ranking and search, capabilities. You know, I these platforms sound like a great option for them to get started. Maybe you can call out some of the, you know, recent tools and services that AWS has released for their AgenTek AI, development. Sure thing. There are many to call out, but I'll specifically call out four, that I'm particularly very excited about. So, two of these are, like, agent tech AI development tools that we released recently and can be used for almost any use case and by any company of any size. So first up is Kiro. It's a agentic AI ID that enables something called spec driven development. So this takes it beyond white coding, the unstructured process of prompting an AI into networks. And it creates a formal documented plan before writing any code. So this makes Hero particularly suited for larger, more complex projects where code consistency and maintenance are critical. Second thing in the dev dev tools I'll call out is Amazon QCLI. It's an agentic AI assistant. In the CLI itself, our developers love CLIs even in present day. And this has many capabilities, but one thing I'll call out is multi tool orchestration. So through its extensible model context protocol, the QCLI can act as a orchestrator for multiple specialized agents. So as an example, a complex task like modernizing a mainframe application can be delegated to multiple agents specializing in architecture, performance, and testing. And as for building agents and agent tech, AI work for itself. Obviously, we released something called Strand's agents. It's a open source Python SDK, that helps you build production ready multi agent AI systems and few lines of code. And then Amazon Bedrock AgentCore. So this is something I turned on earlier. But think of AgentCore as a fully managed platform for building, deploying, and operating AI agents at scale, by providing a unified set of services for runtime execution, memory, identity, code integrations, and observability. So agent code handles all the underlying infrastructure security and reliability so developers can focus on building the agents' capabilities and integrating them with their existing tools and data. Thanks, Vikram. Very exciting. And so, you know, we're very lucky to have Neils Reimers with us today. He's the mastermind of AI search and who has had an impact that not just coherent but but beyond. In fact, during his research career, he created the transformers that are the foundation for today's semantic search applications. So, Nils, thank you for for joining us. And you're leading the team here at Cohere developing these enterprise search solutions. Can you explain, for the audience, you know, what makes these solutions so unique? Yeah. Happy to be here. So first, what we see in enterprise use case search is extremely critical. So if you take out of the box, like, very, very smart, but they don't know about your enterprise context. So it's kind of like a fresh graduate, super smart, got a PhD, but without any industry experience, any experience on your company. But if you equip that person with the right search tools, might be able to discover your enterprise knowledge, they can do amazing things. That's the same with LMs, with AI agents. They need the right tools to find the right information, and then they can take the right information and transform very good output. So search is like a very key fundamental aspect of at Cohere. What's different in terms of what unique about CohereSearch, like, we truly care about the problem. So what you see was a lot of solutions out in the market, open source solutions. They aim to overfit on certain benchmarks, on public benchmarks to to claim the leadership position, but then they often, like, break down in production settings. So for example, there's an open source embedding model from, like, a big GPU vendor claiming very good scores on public benchmarks. But if you test and also claim a context length of 30,000 tokens but if you test it on text that's longer than 30,000 tokens, you get totally random output, zero scores because it was never really tested and designed for that. And inquiry here is, like, very different. We're not overfitting on benchmarks. We don't care about, like, overfitted benchmarks. We care about, like, real customer applications. Applications like Adobe, applications from other customers really see what do they care about, like PDF has images in it, has complex visuals in it. How can we make, like, very good search solutions on it? Well, we will invest equally amount of time in evaluating the models. So we have, like, over 400 internal evaluations on very different use cases, be it, like, weekly ads from my Walmart, Target in Hungarian so that you can search for, like, hey. Where do I get the best deal on bananas? To, like, the retail to the hotel industry, like, with the customer. From the hotel industry to find hotels in New York City, outside Manhattan without a bar. And this all goes into the model, and we see, okay. How can we equally improve the performance for all the customers for all the use cases and get better with every new model model iteration? And this is great for you. Like, in many cases, what I see when I talk to customers, they have, like, very specific use cases. Talk today to a customer from the lightning industry, so they are producing light installations in house. And I said, okay. No solution worked so far. We tested a lot, like, how can you search for, like, different light settings, and I tested Cohere at four. It worked amazingly well out of the box, and one was super impressed, and I directly went into production. And this shows kind of, like, how strong the model is, like, really in industry settings in unknown use cases, making it, like, an extremely good model, and you can be really confident it's by far the best test and model on the market and giving you, like, a very, very good search experience from your data. Absolutely. And and no better validation than what we're hearing today with an enterprise like Adobe and a partner like AWS to be able to perform, scale on top of all the capabilities that are required for this type of use case. And so, Neil, as we think about the future of enterprise search, you know, what what what's getting you excited as as we as we look ahead? Yeah. So, I mean, embedding sound, we rank a skip for steps. But it's far from being solved. There are, like, a lot of challenges. We know as an employee, if anyone who worked in in any large company knows, okay, there's a lot of knowledge. It's very hard, to get access to the knowledge, to find the right information. Sadly, reality is in many cases. You ask someone if they know someone, who knows someone, who knows someone, who knows the answers, kind of like you ask your manager, and your manager knows his manager, and that manager knows some colleague. And so I'm very excited about, like, seeing more intelligent search system, also combining AI agent with search system, longer running search solutions. So, traditionally, search was, like, very quick, like instant 100 millisecond. And now with agents, stuff like cloud code, you see, okay, sometimes it can take a minute, ten minutes. And if it really finds your answer, you're, like, very, very happy on it, and it's also changing, like, user perception. Like, before it was, like, hey. I need an instant answer. If I go on Amazon product search, I search for jeans. Hey. I want, like, hundred millisecond answer. But if I spend, like, a lot of like, in enterprise setting, you say, oh, there may be a big project we do or work we do in, like, parallel. And if the system spends, like, ten minutes an hour on it and then says, oh, yeah. Here's, like, two identical work streams. If you cut one, you save 100,000,000 for the company. You're very happy that it just took, like, ten minutes to say, okay. If I would go to, like, a consulting company, it would have been taking months and cost me tens of millions to find these inefficiencies. But with AI and search, I think we can enable it, and I will be very interested in how to build these very high value search applications. Mhmm. Absolutely. And no better way than to move move forward than to to see it in action. So with that, thank you for the insight, Nils. Abhijit, let's go back to you and, see the demo of Adobe AI search solution powered by Cohere. Sure. Definitely. So till now, you were able to use, Acrobat AI assistant for single docs and multi docs, but every time, you had to work in isolation, just you and your documents. Going forward, we have come out with PDF spaces. This is, this is a solution that's, almost kind of like a digital desk for for your, your tasks or a set of documents that you work with. Just so this is how I hope you are able to see my screen. Yes. Yeah. So this is how PDF spaces comes up. It's it's it's an AI powered workspace for your documents, not just your PDFs, your Word documents, PowerPoint, Excel, and various other documents. You can even bring in your text. It helps you get things done, based on your collections. And and just to kind of get you started, there are certain curated PDF spaces that that are available to everyone who logs into Acrobat. These are, freely available to all of you. You can just go ahead and try it. These are PDF spaces. These are set of, documents that really help you achieve a goal that you work you are working towards. So I'll just, share how how it works. We start with a a a use case where we have collected a set of documents that talk about, being obsessed with how you deliver value to your customers. You collect all those files, you upload them, and and you can add them as and when you get more information. You can keep on adding them and just kind of curating your your workspace with a set of documents that really help you, help drive forward. And this is where the power of AI workspace comes in. You can you can come in here, get all the insights, high level insights, of your collection. You can see how do how you can unlock growth through customer option obsession. There are dimensions related to customer value. So these are kind of the the insights that you get immediately even before you you you are having to look at the documents. And I we did talk about how attributions really establish trust, in in the in the output that you are getting, and this is where you get you can really validate the the the text that's been generated, really is driven from the documents. The source material, you can you can go over and and validate that, hey. These things actually exist in in the document, and and you really are able to trust the AI. And so this is where you get started, and then you can chat with your with your documents. There are certain questions we made for you. You can barely go in and and have that conversation to enable you to see, is your understanding correct? Or if you really want to kind of get get a action done without having to go over all the documents, this is where you get started. And, again, here again, you can see we have activations which really help you help you see where where these insights and answers are coming in from, which you can validate. So this is just one way to to get started. Beyond that, you can create your notes. This is where you can add your notes. So as as you move forward, you you have certain insights. You add them. And these go back into the workspace and add value to your future conversations with with within the the workspace. And there are other ways where you can customize your agent. Then we were talking to a assistant. You can also talk to an analyst. Say, if you have a a workspace for financial documents, you can bring in the analyst. If you are a student, you can really bring in the instructor, agent, which which will really help you understand the entire entire, set of documents and and really enable you. So these are some of the things that are already available, for you. So this is one example. There are few more examples where, say, for example, you are researching about happiness. You can go in there, do your research, create your your reports, have some findings that will really kind of use in your day to day life. And and these are some use cases that are really enabled through this these workspaces. Here, we allow you to have more than 100 documents at once. So, if you are working day in, day out with a set of documents, these workspaces really help you, kind of get to your, actions quicker. And so comprehension is one part of it. The other part is you can really go forward with creation. You can create multiple things, new documents, new insights, and and that's how you move forward. So one of the things that we have added here is now you can collaborate. Till now, you were working alone with your AI assistant. Now there are multiple parties in in this workspace. There is you. There's the AI assistants, and you also have other collaborators. You can go ahead, click on share, bring in other collaborators so that you can work together to to get things done, whether it's for finance use cases, legal, or any other knowledge workspace. It can it could be shared between parents and their children to to kind of learn and and and many other, very interesting use cases, people are are are coming up with with how they collaborate in the these workspaces. So, I hope you all will go ahead and try it, and and really, really amazing product. Thanks, Abhijit. Yeah. Great to see it in action. And as we mentioned, the different capabilities that we've heard from the speakers today are are are are are are obvious in terms of the capability, the performance, the scalability. Vikram or Neils, anything you'd like to add before we move on to, QA? Yeah. Very quickly. So, obviously, it's like a great product. We've seen this. We have collaborated with you for a while, obviously, the courier team. So what kind of personas is this scaling to? Like, I from my understanding, it's various different enterprises and many personas. Any can you give us give the audience some insights into that? So I think in real world, use cases are not limited just to, enterprises, but within specific subgroups within enterprises can also really leverage this. And beyond that, say, for example, for customer service use cases, you will have a set of guides that you need to follow. This is where this is the the single point of contact you can have for for all your all your policy or your, conversation, how how things need to happen, your knowledge bases. You can bring all those together here. So any any set of knowledge workers, you even HR, legal, marketing, product managers, management can use this. Sales is an excellent use case, and students is an amazing use case where, you can really kind of bring in a set of, documents that are interrelated and really kind of comprehend the the the the content. Yeah. That's amazing. That's it's great to hear that, you know, this is solving for so many different personas and, getting to so many different profiles. That's awesome. Alright. Okay. So, with that, let's You wanna enter into q and a now, Nick? Yeah. Let let's do that. Perfect. Questions come in. Maybe just oh, perfect. Yeah. We'll stop sharing the screen and then yeah. I mean, Nick, I think if if you wanna take it, I think we have some really great questions in the q and a tab. I'm seeing I think we we go to the first one that came in, to to reward that. There was a question from Jeff, about using multiple agents versus a single agent and sort of the the technical expertise involved in setting that up if somebody wants to take that one? Yeah. I can take it first first, Evan. So I I did thanks for the question, Jeff. I did put the response in the chat there. So it really depends on the use case. So a multi agent system provides benefits across accuracy, scalability, efficiency. Like, it does allow you for multiple steps, task, by having specialized AI agents to work together like a team. Right? So we've seen and that's the that's how we're seeing these system designs, well, we want we want at this point in time. But in contrast, I've also seen, like, single agent solutions, where it's simpler to because it's simpler to manage and better suited for, manual, straightforward tasks. And, going back to the the like, how how much technical, competence is needed for developing agents or multi agent collaboration systems, I I would like to say there are all kind of solutions in the market at this time, like, from, like, low code, low code to, very, like like, large scale systems like what Abhijit was showing for that for Adobe, that can be built and operated at scale. So I'll I'll put in a reference to Bedrock, where literally through click buttons, you can deploy an agent, and you can also have, like, a, primary master agent doing collaboration, for a multi agent system. Great. And then, I think another one that maybe, you know, we haven't covered too too much already. Abhijit, maybe getting back to you. You've probably touched on this a little bit, but, you know, what what's next for Adobe in terms of search with your product? It sounds like this is something that is already delivering a lot of value. And so, you know, where where do you go from here? I think, we will be expanding. We are actively expanding into agentic use cases with search. And and when you really bring in agents, search becomes a very critical part of your system. You really need to provide the right context to the agent to for it to function appropriately. And and agentic use cases, one area where where we are we are actively we have some components already live in production and more is coming in. And then, the current limit of 100 documents we feel, even though it it just is a huge number of use cases, we really want to push it to its limits. So, increasing the size of us the search space that and the size of our repositories. So these these are areas we are we are we are actively working on. And then, as we do that, we want to unlock search with in we have worked towards multimodality and images. There are other formats with with which we really want to tackle. And, when we start to do that, increase the repository size, bring in lot of, chunks that we we eventually search towards. That's where we rank, kind of technologies come in, and and and that's where we are fine tuning to make sure we we create the right context for our agents so that they can, really achieve the task they are working towards, whether it's to create answers or to create content for you in in terms of, presentations or or text or reports or help you analyze, your content. There are example of those agents already in our system like analyst, and you can create your own, custom agents. That's great. That's really exciting. And so then we have another question here, kind of going back back to agents for whoever wants to answer this one. Someone asked about the different agent profiles, so an analyst or if there's any others that that we wanna talk about. Maybe Nils, I don't know if this is one you wanna take or feel free for any any of you to jump in on these. Yeah. We do you wanna go first on, like, the different profiles you showed in the demo, like, an instructor or analyst? Yeah. Sure. So these are custom instruction based agents that we have in our, digital AI powered workspace, PDF spaces. Analyst is one of them. You can create your own, agent that is customized to your use case. Say, for example, you are an, NGO that is working for care for for a a set of people. You have a a set of documents policy documents that you have to really follow. You can create your agents based on on on on how you you are managing your your your compliance with those policy documents. So, that's how you can kind of work with agents. We this is just a a step in that direction where we are enabling you to create agents. And and and I think Vikram touched upon it. You can create your agents on AWS, but you can also create your agents within applications. That's that's where we are enabling you. Cool. And, yeah, in general, for agents, I just see it's like there's, like, different agent in so implementation. There's, for example, the very simple React agent. You do, like, a reasoning, some two call, and then do observation on the two call. It's doing, like, 10 steps. It's very good, very easy, very nice, but it's sometimes, like, too limited. So if your prompt requires, like, a very deep reasoning, very deep research on a topic, it's yeah. I'm not really able to do this in, like, 10 steps. So 10 steps, like, superficial. It's good for, like, everyday problems, but not like, hey. I wanna analyze, I don't know, a 100,000,000 investment in NVIDIA. Is this, like, a smart idea or not? And then you have, like, more complex agents, like, the research that does, like, some exploration, parallel, tool calls, some reflection, some checking with the policy, some summarization of what it discovered. It's a lot more complex, but it can also go, like, a lot deeper into into your information. But the challenge was, like, the research is, like, very slow. It's taking five minutes to thirty minutes. In many cases, if you have, like, very simple question, hey. How many people live in London? You don't wanna wait ten minutes to have, like, a deep research answer to it. And so it can be interesting to find the right agent or gently loop for the right problem. Well, what we will see was, like, GPT five, but we see it's kind of like a router mode that you say, okay. This is, like, a very quick model. This has, like, low reasoning, high reasoning. And can I decide, like, hey? This goes to, like, a very quick agent that can be, like, very quick answers. This does some light reasoning, so light search, we like to call. And this is kind of, like, the pro version. It goes very deep into it, but also takes its time to get an answer. And can you build a system that intelligently takes the right agent from the right problem, which is inherently very complex? But that's probably how the future will look like, and what we see in reality. Like, sometimes we get answers quickly. Sometimes it takes a lot of time to get an answer. Yeah. I I I totally agree with that. And and and like you said, it's a spectrum from very simple agents to agents that are way too complex, running multiple loops to kind of get things done. And over time, we'll see more and more of different types of agents in the, in different applications where, those applications will rarely route you to the right agent and get you you should be able to get, a a lot done. Great. And we have time for one more question. So I think the last question, we can ask could also go to you, Abhijit, is how many documents are embedded within PDF spaces? Have you seen any concerns as you scale up? How do you, you know, see that process of of scaling further and further? So right now, we support up to 100 documents in, PDF spaces. And and like I I shared earlier, we really want to push this these limits to much larger numbers, and that is where we will be leveraging technologies of embedding models to, really get the right context to the to the aid assistance and agents so that they can give you the right answers or or perform the right actions on your behalf. And and and the role of embedding agents helping you search for the right chunks from your your huge document deposit trees kind of is the is the technology we, we will be leveraging. So it's, still well augmented generation. I think it has moved in that direction, and we'll continue to do that. And and and that is where when you start to retrieve a lot of content, lot of, content that may not necessarily be required once the the the context comes in. And and that is where we start to use reranchers to address, those kind of problems. So, eventually, yes, we want to really push the limit. But as of now, it's 100 documents, which which we we we have seen, kind of addresses or a very large number of use cases. And and and what yeah. Sorry. I want to add, you can create multiple collections. You are not limited to one collection. So one collection under documents, but you can create n number of collections, in your account. You can work independently with different collections. Well, I think that that's a very good, high note to end on. So thank you so much to all of our speakers. Thank you to Nick, our moderator. And, you know, we really hope that this was a valuable session.