Build AI Agents using Integrail (Halloween special)

so yeah brief int introduction on my side um Anon is the CEO of integre which is a platform building agents for various purposes and this is going to be more practical talk compared to all of the other previous talks which we usually had um people talking about papers or whatnot this is going to be much more practical and actually Anon is going to share some free credits uh for you when you register on the platform and then additionally when you enter my name you'll get additional discounts so you can start playing um I really hope to see some cool app that I want to start using myself like we were discussing Anon and you were talking about having an agent that basically like crawls archive and sources some papers that would be amazing or just like analyz his email um so yeah with that Anon you can take it over and I see some repetition mirroring on your screen so I'm not sure if you're sharing the actual slides yep okay there they are all right thank you Alexa can you guys hear me well perfect yeah okay super so first of all thank you for inviting me thank you for having me over and actually it's really exciting and uh nice to talk to AI enthusiasts uh that are like myself because the inspiration behind building this platform was actually to enable people who have ideas uh for different ages to be able to create them really quickly and um to start playing with them uh as fast as possible and it's interesting that Alexa is mentioning uh the papers so before I go into more detailed introductions and everything since he mentioned it already uh I just want to mention that the main um the main inspiration behind integral was actually Minecraft paper uh I'm sure a lot of you have have read it that was the paper where the guy has built an agent that explores Minecraft world uh via well it was a multi-agent with agents working together with each other with agents that were able to generate code Etc and that paper kind of inspired me quite a bit towards trying some of the similar ideas in the real world and that's sort of one of our not really end goals but sort of midterm goals uh to build an autonomous agent that is able to work in the real world using the concept similar to what were introduced there such as skill library and uh Vector search for skill agents Etc but I will touch upon that uh a bit later once we get into the D there was Nvidia paper right like I think Jim fan shared it I think he was one of his ear yeah it got quite a lot of reception I would say nice one more uh Logistics remark um if people can turn on the cameras they'll be amazing to see the to see your faces when because I really want to make this one more interactive where people ask questions to Anon and just like yeah what whatever pops to mind I don't want to be asking questions only on my side I want to let you guys also interact with Anon and and get answers to whatever questions you have and then one more thing is if you have a question if you want to ask it verbally as opposed to just textually in the in the DM just raise your hand and I'll be curating who gets to ask a question so we can we avoid people talking over each other all right thank you Alexa okay so to started off guys uh we are giving away uh credits for external apis the platform itself is currently uh free for the usage we're not charging for any license that's all down the road we just want to see as many people building different agents uh as possible because uh also what we see in the market is that we are mostly limited by our imagination so we get inspired constantly by our potential customers by researchers with whom they're talking so people who have good ideas for agents they're able to implement and build them and I'm sure uh with a group like this we'll be able to build something really cool and amazing so in terms of this credits um we will be doing live demos today quite a bit uh I will show the actual link to register uh in the app uh I would maybe ask you not to go and register just yet uh just in case because I will be showing Demos in the production environment uh we didn't really test it with the uh you know hundreds of people working simultaneously if that happens so if you could hold off registry Until the End uh but yeah so we are offering free credits uh in general uh for everybody who registers but with uh promo codee Alexa gorich you'll be able to get additional credits on top of it uh when this ones expire because we are very interested in like I mentioned people who are excited about AI who want to build different AI agents and we want to work with you as closely as possible and apart from that apart from this immediate sort of nice giveaway uh we also are thinking about having a program for AI Builders so again for people who want to build different agents uh we are willing to provide all kinds of uh incentives because once again what we have found is that it's a really huge wave with you know gen and with large language models especially and with what trying to see uh starting to see in terms of applications towards building different agents but the challenge is that people uh are not really getting it that much so I believe everybody who is into this agentic space we need to work together to educate as much as possible we want to launch also University and absolutely free University to educate people on how to build agents how LM Works how they can work together to provide advanc functionality etc so we just think that there is real value in the community in exchanging ideas and we want to encourage that as much as possible yeah and this link in the bottom that's our also Discord group grou uh where we are discussing support issues and just brainstorming agents Etc BR pasted is also in the chat you can

copy paste it there okay so uh a couple of words about uh people behind integraal uh myself uh Anton antic I'm co-founder in the CEO uh I was in software pretty much all my life I ran VMware I was with Microsoft uh then eventually I joined a company called theme software you might have heard about it rner timf who is my current founder he was the founder of wiim software that was an amazing journey we grew from zero to 1 billion on sales in 10 years it was the fastest growing European uh company uh rir sold it to inside Venture partners for five billion several years ago I made an exit a bit earlier we were doing different startups but then with this AI wave we sort of set together did several brainstorming sessions and decided to start uh building inter and also in terms of my personal background I'm a physics major so even though mostly in my career I was doing business roles but I do enjoy research I do enjoy coding myself and I'm actually focused on the product at integraal uh quite heavily and also I don't know if there are any husk lovers among you but I just recently finished writing a book on husk that should soon uh see the light of day so if some of you are interested I'll be happy to discuss are you using Husk in the background for anything to implement on the inter platform yeah parts of the back end based on husk yeah nice that's hardcore I rarely hear people using H in production I think Facebook had for spam filtering I heard they are using hll but it's like not really wiely adopted yeah I wonder why a bit of a steep learning curve that's what I'm trying to change also with that book um all right so once again uh I think with a group like this uh I'm sure you've been following research and uh you've been doing research yourselves so there's no need to explain things in depth but the way we Define agents and the Agents that we want to build on our platform are roughly shown on this diagram so basically the best way to think about agents is uh like about human beings uh they have sensors just as we as human beings have eyes and ears uh to perceive the environment around them uh we have hands and feet and mouth to do the change in the outside world and the Agents also need to have something to be able to do a change in the outside world such as ability to do API calls or to press some buttons in the browser Etc and then of course agents have longterm memory they have shortterm memory they have a brain which is usually a combination of different llms working together or in the simplest case just one of them and then this idea that I really love from Minecraft paper uh we are actually building a library of skills for agents uh so that they can use them eventually even dynamically uh to find the specific skill for a specific task but uh that's kind of a bit of a longer term Vision but these are the agents we want to work with we want to build another key Point um just uh from the business point of view uh and all also in terms of differentiating with um potentially some of the other companies or similar project that are trying to do something uh we firmly believe that the best way to build up either personal or corporate AI is to have a team of Agents uh Team of a agents that are specialized for specific tasks uh which they can perform uh very well but another key point is that they have to work in the shared context So currently the most popular as all of us know it's a vector search Vector databases uh we also started experimenting with graph rag because again there are some papers that show that combining Vector r with graph rag provides even better results but the point is if you have a team of agents that are focused on your tasks on your workflows that work in your corporate or your private uh knowledge context that's the way to win uh because generic agents are not good they don't know you they don't know your corporate data they know your personal preferences um and the agents who are specialized and work in a shared context is the right way to uh solve the problems and also another maybe a little bit naive but a big motivating factor for myself is that I really don't want us as Humanity to end up with you know like three chat boards from Microsoft of Google and whoever that will control everything I strongly believe that if we have ai even if it's not like ji or whatever but if you have useful ey that knows me my preferences that they fully control they know how it works internally and theci knows me very well what I like what I don't like what I'm interested in this agent can really serve me in my best interest not like generic corporate agent that is build to make money on me I want to enable people small companies whoever to be able to have this really helpful AI that is working in their interests and I think this is a way to a better future than uh what might happen in some I think it's worth mentioning that you guys are using goal of the open source models on your platform you support LMA you support mol right correct yes and um those are the ones with which we love to experiment the most because they also allow for much more flexibility in terms of you know moderation and all moderation in terms of data privacy Etc so it's very important yeah um right so what we are trying to build what we have built already because those um autonomous agents are a bit over step in the future so the current platform has a no code visual editor which we will dive in uh in the next slide basically we have an integrated Vector memory basically we're just trying to make all of the terms and all of the parts of building a good Agent easy to use for people without the need to code anything uh we have a benchmarking tool which some of our customers and users love quite a bit because it allows you to systemically check the performance of foundational models and your agents on your data and this I think is also very important point because usually when the new model comes out people test it on GMAT or or on the lawyer bar exam Etc and it's all nice and good but it doesn't give me any useful information whatsoever how this model will perform on my data on my task here you just create your own questionary you give some guidelines in terms of how to grade the answers you give and you get very systemic um uh assessment of both your models or your agent are doing uh but I will also show a little bit on that yeah it would be cool to see this later on the platform yep slide um and finally the fourth piece is we have a one oneclick deployment into Cloud so once you have built an agent even though we are fully no code as I mentioned I myself actually like to code sometimes and sometimes uh you Cann not Implement a full application with just the agents so you just deploy them and then you can use your team of Agents via API uh if you would like so there is a full API access with some sdks and the range of this SD case will be expanding obviously but then we also have a standard uh chatboard interface on the web through which you can interact with your agents uh and we are working on mobile Chrome plugin and Discord is actually coming up real soon and I promised Alexa to put it on his server as well as one of the first trials but uh obviously these agents work very well with Discord just because I guess one cool thing would be if we can brainstorm some something interesting that we can deploy later on our Discord that will be that will be a success let's see y um right and then so sort of under the hood we integrate with all kinds of gni models uh as Alexa mentioned it's llama it's open eye it's anthropic it's Google it's uh whatever the people want it's different business applications I'll talk about our approach to Integrations uh also a bit in the future and of course built-in rag uh via Vector memory via web search and um uh via the graph r that we adding and all of it allows us to implement all kinds of workflows personal or uh business all right let's then just go into the platform itself so um just you'll probably have to remove the me Google if you can move that widget somewhere on the so people can see it hopefully don't exit the Call by doing that okay perfect like this yeah so here on the main page uh

you see your agents if you just register obviously you won't see anything here but then we have a bunch of sample agents uh which you can uh customize so use them templates by clicking the customize button and it also gives you some uh explanation of what is inside of those agents if you want to go into the Integrity actually this documentation is also written by another agent what we have found is that they are very good at documenting and explaining stuff if you give them some code uh that we use behind the scenes and sort of the uh the reverse task of that is generating the agents from what people are asking us uh that is also very promising I won't be showing it today but we will uh Implement on the platform quite soon um so to build agents you go into agent studio and then into the design agent part so let me first build a simple agent uh from scratch uh slowly so that you get used to the interface and then we will look at some more advanced examples without necessarily repeating all the steps so to create a new agent you just press this create new agent button uh you get a kind of a template which includes agent inputs uh a single llm node and agent outputs so that's basically the simplest agent you can build and if you click on any note there will be some additional settings uh system prompt user prompt temperature chat history length so basically if you keep chat history at zero or empty it just behaves as a transactional agent uh in some cases it's very convenient especially when you want to transform some data uh and I'll show a couple of use cases of that as well um but if you are really into it and you know what you're doing you can click on this simple more button and we will actually show all the other parameters that you can use with the LOL such as you know top etc but we're not going to go there that much uh now so here I would just um change the model into something bit more exciting let's just take GPT 40 uh then we can go to agent settings write something like my first Co agent then save it um and then we can click on the agent inputs and put anything into user prompt and press this run button and it will start running so it shows the status it was running for some time then it shows that it finished then you can go into this into every node basically and look at the output that is being generated or you can go in the output and read it here but of course again this is just a designer area you will be able to um work with the agents from chat as well I will show it in a bit but let's make this agent a little bit more interesting so let's Anon maybe just uh if you can demonstrate how people would actually be using it later like the deployment I think that will be useful before we go into more complex agents agreed so let's go into chat with agents here it is uh on the left the agent that we just built it says hi hello how can I assist you today uh actually as mentioned uh this agent uh we didn't put anything into chat history So currently it's a transactional agent and it's also a nice demonstration in itself because sometimes people get confused they think there is some magic behind you know this chats and uh whatever but it's all just about information being put into the context of an llm so for instance here since there is no chat history we can say who is uh Hing way for instance uh it will give us some answer because obviously they know who hu is but then if we ask something like um which book of his should I read first or what was my previous message yeah it still have no idea because we don't have any chat history so it's asking I would need to know which author you referring to so let's remedy it real quick let's go back to design agents just to show one more thing this is cool but obviously most people will probably want to just have a URL and build an agent here and then use it in their own application or something like how do you can you show how we can easily have a URL endpoint that we can ping and like have the same functionality but like almost as an API do can you demonstrate that now yes uh we can uh let's save this agent so to deploy the agent in production again you can use it via chart interface here but if you want to build a custom UI uh there is this rocket button uh that says deploy to Cloud currently it's all there is you just have to press it once it says the agent will be deployed into the my account you say confirm the agent is deployed then we go into my account settings and we see all the agents that are deployed uh in this account so we have quite a bunch basic chat Bo let's find the one we just yeah here it is my first cool agent and then if you want to know how to access it there's a full uh there's a full link with instructions how you can access the API but then as I mentioned we actually have a couple of sdks uh in Python and in node uh but they will be more uh so that's really cool and if we had images or multiple modalities would you automatically expose all that through the API endpoint you have full control how you do it so for instance let's take a look at an agent that has image recognition one of this ones is what this lead enrichment agent so what it does it basically scans a business card or or anything a picture with any information about a person uh so here uh we can put uh an image let's just go I think I had yeah I had like a small picture with my email and then we say something like go so this is the picture that I uploaded then there is a model that recognizes the picture then we do LinkedIn search we read the LinkedIn we summarize something about uh myself and also about the company that I'm working for so what it this agent does it finds my name it finds the company I work for it finds my LinkedIn profile and then it does two summaries first one is about the company I work for so integral it did another Google search to find the integral read it and then it provides a summary on myself uh based on my LinkedIn profile so obviously I love this one yeah if you put it into an API then you can further integrate with your CRM system or with your table or whatever and you have a very fast uh lead enrichment tool that that's been built in like five minutes basically so this wasn't catched it looked very fast no it wasn't cached yeah okay that was very fast yeah it's the way it works um so I can show you how it looks like inside it's uh obviously more advanced than this one but since we switched to it already let's take a look so uh this one is quite a bit more complex but it actually demonstrates several interesting approaches uh that are useful when you build an agent so first one is a node and when you're building an agents add node is here on the top left you can just press it and then you have all kinds of different nodes that are available for you uh to use in your agents there are a bunch of jni nodes uh search Rel notes and then there are notes about Vector memory we will highlight some of them so this one is image to text it takes the image it gives a description then we have another llm also transactional that's a search query generator uh which analyzes whatever is in the image and generates a good Google search that searches just the linkedin. com site we could have searched the whole web but obviously we want some good information so we only search in linkedin. com then whatever we found here goes into a universal box into which we put a lot of thought and effort that can read pretty much any website as you know scraping websites is always an issue our readability box is able to read all kinds of websites that will not give you any information if you do a simple fetch um then we convert it to markdown uh because again if you do HTML there is too much uh too much noise too much junk uh for the llm and then we have another llm that does the summary on the person and then we have the same for uh for the company info but the way it works there is another quick LM that is all it does it extracts a domain name and here it's an interesting philosophical question actually because uh I myself being a programmer first I would think okay I should write a function that takes you know the main name out email name and then we do something but since LMS are so much versatile you can give the input data into any format it can be integral. EI it can be integral space do seven spaces sl. and it will still understand and give you the right domain name so this is the Beauty and the flexibility of LMS that you don't really need every time you need to transform some data into other data it's actually much more versatile to use LM yeah it's uh probably slower in some cases but I mean in case like this it doesn't really matter so all it does it just extracts the domain name from whatever was given to it we can see it here here's some instructions uh then it says uh to give structured output because it can actually give different outputs and we can select uh different branches based on what's Happening Here but the output it gave this time was name the main inputs attachment integral. so a structured Json output that goes into another readability box that reads my website and then there's quick question here do these templates these more complex agents already exist on the platform so when people come can they find inspiration and analyze these or yes exactly lead enrichment agent is actually one of the feature agents that we have right here so nice I actually have an idea because

you mentioned this scraping tool that's very good and can do arbitrary scrapes so you saw that I launch correct Engineers uh yesterday or two days ago yeah let can we maybe build an agent on the flight that does the following I just give you an URL for the like applicants like a job post and you basically extract text no matter the document into a text readable output and we we we test it now live that I think they would be really cool and probably more interesting than going through pre-built agency if you think we can pull it off quickly uh so what what's on the URL Brian says Alex promoting correct Engineers absolutely not I just found the perfect example to test the platform so what's going to be in the URL that okay so let me quickly share the screen on my side yeah see can you see the screen yeah okay so it's not a plug for my platform you should check it out but like so basically all of the job have some link that leads to their job offering you see so it's going to be very diverse for dynamic like a their website uh for onslot it's actually I think just like a GitHub read me we are hiring and then LM Studio I think is also like no it's actually doc form so this is a bit more complex so as you can see everybody has their own ATS so this is again this is drawboard greenhouse iio so can we build something quickly where I just give you this URL and you give me the relevant text uh from the URL and we printed out somewhere no that is very easy of course uh you you're going to give me the actual link to their job posting right yeah y okay yeah let's build it um just need to share the screen again yeah sorry second and if people have ideas for what we can test on the Fly I think that's much more fun than going to pre-built agents uh um yeah I would still like to show a couple cool ones all right so we will do that as well let's go to the agent Studio let's just take this first Co agent so okay the input is going to be URL we just put this readability box since it can read any URLs uh we give it uh all right we just connect it to the attachment markdown is fine this we're going to kill content we're going to put into doesn't really matter let's put the system prompt um so then think user prompt might make more sense for gp4 in the system prompt we can just say you are like a URL reader or something yeah it's it can do both um all right uh your task is to summarize the job is it a job offer or application um it's a job it's like um how would you define IT people describe what the role entails and all of that so it's like job offer I guess with details about the offer given to you like salary or whatnot maybe if you can extract salary information and the position that would be cool as as explicit things we're looking for okay yeah let's for given to you please respond is the following format jobri summary salary in avilable and uh else I don't know skills required skills required and array of skills in JavaScript um yeah let's just try this one quick and do yellow okay let me just paste you some of the existing ones let me not take the hardest one first let me give you the easiest one was maybe the first one I opened so I'm just gonna face the URL in the chat here you can copy paste it okay Dynamic okay let me just open it there you go okay let's try running it here first and then we will see the nice format in the chat format MH okay run live demos let's see this is the ultimate test if you if you got it robust enough for an N hog idea we actually didn't agree up front this was like a random idea that PO to mind while desing the readability finished let's not see what it gives here because it's nice formatting here we'll just move here so there we go nice okay it messed up the seller let me see whether there was a celery information no they did not have it let's quickly test one more that has information with CER yeah just me a second let see which one has salary information gave a nice array of skills yeah yeah uh um ahuh let's see this one there it is available but it's maybe a bit trickier to scrape it because it's in the header let's see if it picks it up yeah let's check it okay so it should basically be you can see it on the top yeah next to the location let's see if you can pick it up ah yeah that's interesting okay where is start here this away hide this and then just Bas it here MH Moment of Truth drum roll ah it messed it okay it probably requires a bit more work couldn't it couldn't read the website so this is one of those rare cases when the readability breaks because okay because it's what dynamically render JavaScript or whatnot okay got it no this was still super cool it's fun to see obviously all systems fail and you have to put in a bit more effort but this was amazing this amazing tuning yeah you can obviously get around all of this issues and we constantly fight un readability to make sure it supports what people need so yeah since

we were talking about API deployments Etc let me just uh really quickly show you another one I headed here because basically what we want to do here is we want to make these agents available uh pretty much anywhere so we we're going to do like standard uh plugins into to all the popular systems but yeah sorry let me first show you the agent so a common use case uh responding to the emails we actually have several customers that uh we're working with to do that but before we go there let me just show you real quick the whole point about the vector memory so um as I mentioned what is it uh here in the slides yeah the best approach is when you have a shared uh context shared knowledge context and the Agents that work in this context so basically this platform was built around that every account has different memory sets that you can use let me just switch into integral account because it's there uh so there is a memory Set uh which has lots of different uh memory pieces that's for the demo purposes you can easily create a new one it's basically a dedicated collection in a database with vectorized data and then it's very easy uh updatable via this interface you can upload any word file or any PDF file and it will become part of your vector memory or there is also a node so you can update your vector memory from the agents themselves as well and this way you kind of get a self-learning agent because it can expand the memory that is available to them can you pick the embedding mod used in the background or you guys picked it for us we picked a couple but we are open to putting new ones it's very flexible basically um then what can we do yeah so there is this agent that's called email responder it's a very basic agent as well what it does it takes an email uh as input then it does Vector search in whatever uh let move it away so once again Vector search there are options here I can go I can choose different Vector memories I only have one in this account so I choose this it finds several pieces of memory I put them all in the context and I ask my responder which is a Cloe this time uh to respond to an email using context and this makes all the difference I mean this is the most basic stuff for you guys if you work in thei if you're a researcher but when you start talking to actual customers for them very basic r this or or Google search based makes all the difference and people get amazed and actually it does give good results so basically all it does get an email it does Vector search it puts it into context then provides the response but the beauty of it is that I deployed it already into production and we have this small test plugin uh into Gmail but we'll have similar for the uh Microsoft as well and I have actually sent and a test email a while ago already that we can test it with the email says hi I'm interested to test the integr platform we need to build the customer support board can you help so here is this plugin very fancy interface uh just one button I click it my eyes are bleeding from that color yeah I'm not the best designer will definitely make it better with the design so we get okay here now we just go back and then refresh and we see there is a draft appeared automatically I didn't even do anything now I go into the draft and you see a nicely professionally written email with the context it actually says we'd be delighted to help you build blah blah it talks something about Integra because it's using the context it even tells go and sign up for free trial here in the link I didn't tell anything to it I just took everything from the vector map so basically if you build your vector memory and then you have a team of agents that interact with all kinds of different systems you can build anything you can literally I don't know do whatever you want the sky is the limit this is amazing and I have a question so you had an email you replying to you opened up the widget and you and it basically what read the email and then using the corporate data it kind of give you a draft and then waiting for you to be approved yes exactly it creates the dra automatically so all I need to do is send or I can edit it a little bit and then send but basically yes it reads the so incoming email is read it gets into this agent we just looked at here via API this finds some relevant information on our corporate Vector database gives a response using this context and then sends back the draft email which we then create as draft in Google this is something I would love to try on my side like as soon as possible I get my only concern here would be how do I know what happens with obviously email data this is a test account probably is very sensitive and like how do you like give customers peace of mind that like nothing right basically the only info that's being sent is the actual current email when you press the respond button we actually have a more advanced functionality that we are building for one of American universities where we have a an anonymizer d anonymizer uh in the loop so basically the way it works it takes the email text it uh changes all the names into something like person one person two then it gets processed on our side generically then the anonymizer does the backwards restoration it says instead of person two John Wayne and this way you can be sure none of your data gets sent anywhere because these animiz ders are llama based and you can run them on your own you know infrastructure if you like Etc that's

super cool people had some questions in the chat Let Me Maybe address them before we go I think there is a ton of cool ideas here um so one question was Rafael asked if if it's possible to use a custom llm um and integrate it into your platform you know in an easy way I guess what that's what Rafel was asking yeah we uh we don't have it yet uh built in a automatically uh but we can definitely do it manually but also what we have sort of our vision with Integrations which I mentioned uh let me just show you this slide once again because it's our near future road map basically uh so we realize we will never catch up with guys like zapier who have 1 million Integrations with 1 million systems that they built in 20 years or whatever so we created what we call a universal API node basically it allows you to connect to any external API so if you have a custom LM you can use this one or we will have an easier way uh to do it via just LM specific interface uh but then basically what it does there is another agent behind it to which you only give the documentation to your API and then it generates automatically the node that you can use and do stuff like in Microsoft or in custom systems and Legacy systems Etc so we don't want to create huge pre-built library of Integrations but then anyone can integrate with any API they want using this node uh it's currently in beta but uh it should be available next week already um so that's sort of on custom okay and then one sagik asked whether we can maybe if you have more cool demos we can go through that but like he wanted to do a live screenshot of the page and read from that I think we kind of did that already right it's just that we missed the screenshot fun let me show something because that's another real cool it's like uh you know sometimes you can get frustrated with developers and like with myself even and you know I was I was saying that uh in many cases the limit is our imaginations it's like I am in this day and night and sometimes I think well probably this the agents won't be able to do and I start coding or I start doing something but then I decide okay let me try and turns out it actually can do it so let me show you something other cool thing um where is it my account uh there is an nent that's called react def I personally hate doing the UI programming uh I'm sure some of you love it I'm mostly a backend guy so at some point I got frustrated with lots of stuff and we decided to build this one and that's just a quick prototype but definitely going to improve it way more but that's the way it works so I just I take a screenshot like we said for instance of this card generic developer I go here I have to paste it as a uh as a file first put it into image and then I say go and basically what it does it tries to build a react interface using bootstrap uh from this image but obviously you can put any framework you like uh it does contain some advanced stuff in the um in the system prompt uh to relate the specific libraries I'm not going to you know run this live but I have pre-run it previously so you have seen that it generates it but I'm just going to show you the result it generates because we were amazed the first time we saw it so hopefully it's not the Green vidget from the email no so here is what it generated based so sorry it's just it's the card so it has all the same elements the Stars here the I here the placeholder even for the image and the buttons with all the right icons so this is what we made the screenshot of if you recall so I guess this is your version of of on Tropics artifacts feature which is fairly powerful can execute code as well right 10 bus and visualize stuff yeah but the point is you can really make yourself much more productive with some fine tuning and with working with right context so it's again we are just only starting to experiment ourselves and that's why I'm saying we are really open and looking for people who are as excited as us to come up with all kinds of crazy ideas and then try to implement them the original this is what it created so if you put the right picture it's pretty much what it is and this is a life react component with you know pressable buttons and everything nice Francis had one more question he said he he'd love to have a set of URLs in a CSV and then the agent would report to him every morning a summary of Abol I guess that's fairly simple to do because you already demonstrated scraping just a matter of how do you read CSV how do you how would that look like on on a high level like how would I provide CSV like um yeah that that's actually something that uh is not there yet but it's coming up either by the end of this week or early next week because it was something we really worked on and this is connected to that agent that we were discussing with you Alexa that reads the AI papers Etc uh because let me show you let me show you why um I think it was in gril account as well so basically as we discussed llms are very good in transforming uh the data from one format to another and any ideas from people on the call what you would love to see like and try and build maybe you can type it down in the chat so these were cool so far here yeah uh about the structured data uh so yeah we can read any link what we La currently what we're working on and what we will Implement is so-called mapping node uh that will be able to iterate over lists uh that's sort of the key feature that will allow you to go over many documents at once so for instance this agent it reads an archive list uh we were discussing this agent with Alexa we kind of share this idea I really want to have an agent that reads all the new papers for me finds the ones that are relevant for me uh gives me a quick summary and then I can discuss them in detail with this agent and the way to do it is to have an agent first that goes through archive every day and puts all the papers into Vector database so this is the starting point that we will make live next week when this mapping note is ready but here is the link to Archive let me just show how it looks I think this is just very powerful because you already have some Twitter accounts that basically do this precise thing and have 300,000 plus like followers I think K probably does this like he buil some custom software for him that scraps archive and just like extracts the abstrack or whatnot summarizes it and then tweets out right but then here also a funny thing happened uh as I was uh sharing so here's the link you have seen it's a page there is code so I thought okay I need to build a converter and I started coding it and I wasted like an hour or something but then I thought what the heck am I doing and I just wrote llm and I asked it to uh basically as we did last time just to summarize the list of articles in this simple Json format and the beauty of it is again for any site reading information if the website changes something any divs any classes or whatever your custom written code you have to throw it out on the window and analyze the source CES again llms don't care that they couldn't care less in which format it comes so let me just run it and we will see unless it's ADV verial but yeah what kind of um what kind of data it returns so basically it starts generating this nicely formatted Json it's not so nice here but we will switch to u to chat to see how nice it looks but basically any Json any CSV any structured format the llm nodes can give you and then what we will put next uh after this agent is this mapping node which we don't yet have but we will have it next week as I mentioned so that it goes over every link reads every paper converts it into our Vector memory and makes it available to whoever wants to interact with the archive based AI papers so here is here is the very nice Jon the regenerate so obviously whatever you put as input they can change divs they can change classes whatever this agent will still produce this Json which is basically an array of different objects so we will put a mapping node behind it and the mapping node will read the URL we'll download the paper we'll converted into Vector memory and this way we will build up a public Vector memory on AI and we will there would be a cool product to build that you can build and like have it as opposed to just having a platform like one D app that people can actually use yeah that would help um okay I see we're almost out of time being cognizant of people's time and and also like a attention span um maybe we can start wrapping it up slowly if you had anything else to to share um yeah so yeah maybe real quick I'll just show that benchmarking because it's actually quite nice uh but for that sorry it's a bit of a mess with accounts I have to log in again because I have some of them in the other account so yeah benchmarking uh again as I mentioned you just create you basically create a questionary uh just questions and answers and then in answers you can provide some guidelines of how to rate them and then once you have created this questionary you can pick any base models that we support or any agents that you have created and then you just run and they respond to all the questions that you have provided if you use an agent they will use Vector memory they will use Google search all the power they have base models only use their knowledge obviously and then we summarize this results in a nicely formatted benchmarks that show okay sorry I have to put this away again uh they show accuracy they show cost efficiency the size of the bubbles is how much time they took then there is a summary table with all the info and then you can actually see every answer every model or every agent provided and also the comments that our AI grader gave to them and the points it gave to them so that you can find you it any way you like and this way you can ensure quality so for your agents um right and then yeah so just wanted again because we sort of shared the vision of the future and we are happy to brainstorm any ideas with you guys but basically sort of The Next Step the medium future uh road map what we are trying to build we're trying to build a autonomous agents that will talk with the user understand their intent create a detailed plan then they will pick simple Agents from the library from the skill library that we have and this skill Library gets developed every time we add a new agent there uh so it creates a plan for instance I don't know uh let's write a report on US economy in 2027 uh it will ask me some questions and your of what has to be in that report it will come up with the plan what needs to be done like research the web uh ask for additional sources um I don't know interview some people it will pick the agents that we have in the library that are able to execute those tasks it will orchestrate their execution and it will come up with the results so it's kind of a meta agent that works with this next level uh skill agents and this way we are able to get closer to whatever people call AGI I don't really like the term but closer to creating really useful versatile agents that are able to help people so by building up the skill library and knowledge Library via Vector databases and graph data databases and obviously you can build them in a custom fashion for your specific area or your company or your product etc that's kind of yeah but with that yeah invite you guys to come in to get those credits to start experimenting to start brainstorming and uh we are always open to discuss any ideas you have Anon thanks a lot this was super cool I enjoyed it I love the fact that you were actually able to build my idea on the Fly uh that was very fun and you know what I have this idea to have like a once a month or every two weeks live session maybe we can do them like together pick an idea from people and just build an agent during and talk about stuff I think that would be cool and progressively more complex ideas yeah yeah 100% thank you guys