Retrieve Any Data You Want with rtrvr.ai

and those tabs are still selected and I'm just going to press extract again and so I wanted these specific things from those tabs and our AI agent can then extract that information from the page itself just using like the natural language and it's also much faster it should be with the fewer columns you want So yeah so it got the exact things I wanted and immediately could even be something more complicated hi my name is Demitri panichi and I'm a content creator agency owner and AI Enthusiast you're listening to the AI agents podcast brought to you by jot form and featuring our very own CEO and founder idin tank this is the show where artificial intelligence meets Innovation productivity and the tools shaping the future of work enjoy the show hello there and welcome back to another episode of the AI agents podcast in today's

episode we have the founder of retriever AI arjan and I'm going to leave it to you to say that your last name um I should have asked prior how to pronounce it it is a pleasure to to chat with you about your app um it this this is one of those applications where you see it and you immediately get the application especially with the uh world we live in now uh and obviously the entirety of this podcast is around AI agents um I have been talking a lot about AI agents uh to automate Tas tasks on a practical uh level that you can kind of see the tasks being performed and I think retriever is going to be one of those products that gets uh a lot of traction in this year so if you want to introduce yourself and how the product got to

where it is today your story so to speak yeah I'm Argin uh chinali that's how you pronounce it but uh it's sounds like a little bit Italian if you could squint but uh yeah yeah so I uh worked at Google for five years on like vertical Federated learning so like uh had been very deep into the AI ml space worked on like setting up a new model ways of training models to train on data silos but when Chad GPT came out I immediately thought about like this is going to change everything like uh and I was thinking through uh applications for this and went through a couple of hackathons and iterated on this idea that like there's a huge scraping problem where people want to compile data sets from the web but uh they're manually either copy and pasting it or like having like uh

custom scripts to do it and it seemed like if the you can get a agent to like navigate the web and retrieve that data for you that would be uh extremely useful so I did a couple of hackathons iterating on like a toy problem of like for every municipality in California get its audit file so it's like you navigate uh starting from like a local government websites or like you know like all the municipalities and you could iterate from there so um from this toy problem I kept working on it and uh with my co-founder uh bavani we both were doing these hackathons both iterating and trying to figure out how to do this and we settled on the idea of a Chrome extension that can do actions on the browser your local browser and um uh retrieve this data to Google Sheets for you

and uh we looked at other players in the space like hey yeah there there's a solid Market here there's existing companies uh like Bine or like magical or other uh uh thunderbit or like brows. a that are like trying to solve this problem as well but they're not using AI web agents that can autonomously do actions on a page to do this so uh yeah so with that we uh kept working on the Chrome extension getting some traction with early customers and we released uh Christmas Eve so not the best time to release probably but yeah so uh since then uh we've been loading up new features and uh yeah maybe I can just show it in action and uh I I'm just actually curious before we get into the the um like kind of um you actually performing with the product um what necessarily

drove you to want to make something that was kind of in this space because obviously you brought up the other tools but what did you see specifically that they were lacking like Bine or any of these other products for example like what did you see that specifically that they were lacking that you wanted to to improve upon yeah I think it was more coming from the other direction of like playing around with llms to like do actions on a web page and like and what ways what like Target use cases could we uh go after with this like After figuring out like initial implementation and we saw bardine like Target for like uh for sales people or for uh like so it's like okay yeah we can do this better like scrape this page with just like a sentence you don't need to do any

clicks or like uh so then we like just went after that use case of like all right let's like make demo after demo on this and like really harness it so that it's completely intuitive as much as we can so like yeah these other tools you have to like click and show like de demonstrations of like hey this is like the table or these are the fields I want but with our retriever tool you could just say I want the name education experience and click and the schema is automatically generated and if you want like uh the more contact info like the email like it will do that action and like get that data as well yeah yeah interesting so I I just want to touch on this real quick because I think it's uh something important that you you mentioned is this also not

only you know I mean we talk about scraping it's and on your website like it's tagged as like retrieve whatever you want with the internet retrieve research automate the web I think something good about products like yours and I'm curious your thoughts on this are well positioned to kind of for lack of a better term reduce or remove any RPA products that uh don't really have the level of capabilities because if you are you familiar with like Microsoft power automate or anything like that yeah that's another Target use case uh because we can you can record you demonstrating like a action and with that recording grounding the model actually it's like it will uh 100% like usually replicate that action so we are also targeting uh automations but like uh we're kind of limited in that we are a Chrome extension so we can only

do actions in the browser but we can automate those uh actions yeah so RPA as well as like scraping as well as like just being all around productivity tool yeah interesting okay cool well if you want to take a a stab at it here and showcase some of your best use cases I think that would be the that would be the next thing that would be be good for the audience to see I can uh show the screen here yeah and if you could when you're going through it like feel free to embellish kind of the work that you've done on the back end that we don't necessarily get to see obviously we're going to get to see like it work in action but like I always find it intriguing how you know someone as the founder like yourself can actually really understand what the

backend's doing in order to make it happen because it could uh seek someone's interest as to why you guys are different than other products y yeah of course so uh as you mentioned we're like a Chrome extension listed on the Chrome web store so it's like we went through Google's uh security and privacy review yep and um we can do a demo use case say you just typed in this natural language find retriever AI LinkedIn page and follow and our agent takes actions on the page right like search for retriever it can do oh the demo gods are not with me but uh it went to like r t rvr but it figured out that it's actually following uh it's already following uh retriever so um but uh before it goes off the rail so like uh yeah so what are we doing here is

uh similar to operator or like a Google's Mariner we're taking actions on the page but we're only using the Dom itself so like we're interacting with the HTML of the page and uh these but these other tools they're taking screenshots and figuring out hey this is the element here I need to click on this and that's like fundamentally limiting right so you would have to like take Scrolls you have to scroll through all the different things and like take screenshots there and then submit it to the model all at once but with the HTML this whole page is already there so you don't uh it's much faster and and it also allows you to do uh actions uh across tab so if you have like multiple tabs that you want to like extract right like uh sure yeah and uh so our um Chrome extension

could do actions on those tabs in the background without uh like actually don't know where it's going to hopefully it creates a Google sheet here but like um yeah so okay yeah uh so um what I just did right here was like okay I selected these three Amazon tabs of like Google or sorry pixel headphones and like other headphones and I didn't even Define a schema I was like figure it out right like if you just don't Define anything it figured out right like uh hey there's these are products I'll get the image and the product name and like these other attributes of these three different tabs right so you don't have to do any clicking you don't even have to type you just click a button and um it will scrape all of that information to this Google Sheets here and uh yeah so

beyond scraping we also have like research like kind of like deep research uh but like deep research with your own tabs and you could uh do actions across tabs right like you can if you have a bunch of docs you could just ask them or like say fill out this form with this PDF so like our custom approach of uh doing like this background tabs and using only the HTML uh allows us to uh because these background tabs don't get rendered by uh the uh by the browser so that's why you see operator only do on the active current tab okay interesting so with ours right we can um we're not we can do it on so maybe it's having uh issues maybe it's because I had those pages open like a while ago maybe I uh did not pray to the demo gods enough

uh clearly and uh so I think it there's an issue if it's like a longlasting page uh there like the Dom is like kind of uh deprecate like it's kind of like cleared from memory so I could so if it's like a new page like yeah you just see like it kind of uh I think I had those tabs open from yesterday and they were just like not loaded in the browser Okay so with this hopefully uh it could uh it should work fine now but uh yeah yeah it was much faster awesome yeah I should have uh but anyway so it's like we can get the image URLs and you could see and like compare these products yourself and that was without any uh clicking and yeah yeah let me let me just call something to attention here because I find this pretty interesting

so for those of the audience that are unaware to my knowledge at least most products like this for example a Bine you would have to make a predefined uh Playbook you would have to like and they've they've done some good imple implementations where you can like uh use some of their preset uh playbooks or you can use what's called a magic box you can say do X I want you to do XYZ thing and it'll make a predetermined Automation and then you have to find all of the page aspects that you then want to put into for example a Google sheet you literally just clicked on the tabs and pressed extract that's actually crazy so I don't want to I we kind of we're kind of going unfortunately with the with the the lag on the first one I I just want to make sure

we hone back in on like what it actually did so literally he all arjent did was picked a couple of tabs and it took all of the what I would imagine a majority of the relevant data uh and then made it comparable on one Google sheet now I do have a specific question is there any way for you to ask it to have specific schema right like it it just assumed schema but like this is the prompt section I'm seeing that I'm guessing yeah yeah so if you just want your own specific thing like maybe color uh price delivery date uh and those tabs are still selected and from those tabs and uh the our model like our AI agent can then extract that information from the page itself just using like the natural want because llms have like a quadratic scaling with the output

tokens so yeah so got the exact things I wanted and immediately yeah MH yeah so you it could even be something more complicated that on this page click on like the reviews like the uh what is it like get the most recent review instead and get the like do this action on the page and see that'd be very let's let's keep going with this this Amazon example because I find I find it kind of uh even applicable for personal use me okay um so I have those tab selected and maybe I could just say uh extract the okay color price and also click on uh top on most recent reviews and get the most recent review and review time maybe so honestly like it's the world your oyster with these prompts right like uh you could ask for anything any combination I'm I also have

never done this before so let's hope the demo gods are with us and it can do those uh actions across the tab so let's see if it clicked most yeah and um it clicked on yeah on all of them across all the tabs thank you like okay thank goodness that worked yeah yeah yeah because you never know I just did it in front of our very eyes I saw it just yeah it and so you couldn't do this with operator because with operator it it would have to go to each page at a time and then maybe it's like a 90% success rate with each Tab and so if you have like 10 rows of this it's like an exponential failure rate but if you paralyze it and distribute it like this um that failure rate is much much less so it got all of

these things and we most recent three years of good use out of yeah I saw I remember seeing three years of good on the um Google buds I recall for this one something about six mics yeah yeah that's exactly it yeah so yeah with the other tools you would need to Def find automation show some recordings and um you don't need to demo the recording for this but you didn't even try that hard on that prompt to be to be blunt you didn't even try that hard on that prompt that was pretty impressive I'm kind of that's kind of that's kind of breathtaking so can you explain again what that meant by the way with a failure rate and operator and by the way for those who are unaware uh chpt operator is the recently released um AI agent uh that's available for only the

Pro Plan which costs $200 a month yeah ours only cost 9.99 a month and you have a free tier but uh yeah with the failure rate I meant uh with each step that you take the llm has x% chance of failure sure it's maybe it's only right 90% of the time yeah but if you start uh doing like step you need three-step plan that's actually 90% times 90 so9 time9 time9 yeah so the total task completion rate is actually like something like it's like 72% 72% yeah yeah but if I split it out like this into parallel tasks right like so this was like a two-step thing like do this action extract this data now it's only per row it's only 80% chance but if you 80% chance per row but with operator it would have been like these four rows and two steps per

row so it's like 0.9 to The Power 8 which actually is like probably like close to 50% or wow yeah it's lower than pretty good yeah wow okay and the reason that just so I can understand one more time what is the reason that yours is Ste uh stepped out uh separately is it basically because like the in order for operator to not fail it has to go through the entire process for the execution to be valid because it's not in the browser it's in its own ecosystem or am I misstating that so they're actually setting up a cloud browser over there it's not your browser it's a yeah but the key reason is they can only do actions on the current active tab but mine can do tab like actions on tabs in parallel so these are four separate tabs and I'm doing actions

on those four separate tabs independently of each other but operator would like go visit the first URL do those actions go visit so it's like a sequential process understood yeah that makes sense that's why you went down row by row which would be a different okay yeah that makes sense cool that is that is awesome um so we talked through your explore capabilities your oh that was capil your extraction capabilities yep I could show the explore really quickly so you were actually the inspiration for this deep crawling I saw your video on Bine and I was like oh that's super cool I get it like you have like paginated response is and you maybe you just want these first five PDFs here and like extract this data from the first five PDFs and this explore capability is that uh I can open okay so it

didn't work this time uh let me just try again uh but the explor capability is I can spin off x subtasks on a page as a new tab and do those independently like similar like you don't have to have them opened already so I gave this prompt of for the first five papers extract this information and so it opened these links themselves and they were uh from the main page so it distributed these subtabs opened them as new tabs and did those in parallel and extracted back to Google Sheets so and you could have just said for all of these pages for all the pages here go through all the PDFs and extract this information so it would have opened all of these uh PDFs as tab batches extracted the information wrote that row and eventually like made its way through all of these multiple

pages so the explore feature kind of leverages that deep crawling uh of like paginated uh sites and it open so from this parent website that was that list it opened this like actual website that was the Paper website like and extracted the top author summary and all of these and another use case could be on LinkedIn and you have a bunch of jobs and you could say apply to all these jobs in parallel and it would open those jobs as a new tab and apply them in parallel yeah very cool okay so this this is another great example of it though because I just I noticed some issues when I was using when I've used bardine and other products it seems like you guys have the capability to have multiple actions even happening at once and obvious and also spurting it across um um multiple

tabs when you're doing one execution of something um what's able to like make what was this one of the things that you noticed originally that was kind of an issue with some of those other products because for me I've always kind of been frustrated whenever I'm doing something RPM or uh scraping related and it's like stuck on one thing at a time almost right um am I incorrect in saying that you can run multiple different uh retriever AI uh requests at the same time since it's tab oriented so um you could have multiple windows open and in so this Chrome extension is like per window so the side panel opens per window and you could have multiple windows open and execute things in parallel and you could have it running in the background so you have like a schedule feature like it will okay this

is interesting too I was going to ask that next yeah so you could you could schedule something and do something and you could just be on your own thing doing something and not even pay attention to it and it won't take um like it won't uh open new tabs actively that will like change the focus of the Chrome browser but um so where would that be just where would that be running in in practicality uh when it's doing the background scheduler so it's still all on your local laptop and browser okay so you would need the laptop open and you would need Chrome open but we actually think this is like a positive because eventually we'll have to set up like a cloud browser as well of course similar to operator but for now we think this is a positive because you know if you

have a AI browser agent like it can do anything imagine the page changes and now it's like a scammer asking you the AI web agent to like donate money or like make a purchase and you're not no one's like monitoring it right and if it's just running on like open ai's like infrastructure and you forgot about it and it's just been running for like a month and now like it had because it has your credentials of your credit card and Etc and it just made a purchase and then you're like checking your credit card so it's like these browser agents because they're autonomous and can do actions independently you kind of do want to monitor them and make like have like a running idea what what all is like ex you have going on like you have executing in the background so it making sure

like it has to be on your local browser for now like should make people more secure feel more secure yeah that's fair um last pass H one p one password all those sorts of thing plus something like this now creep me out um but maybe want to have it installed on a browser that doesn't have the uh password thing that also includes bank account information um yeah actually saw that there was like a tweet going on that like open AI passwords got leaked I got scared I like immediately like went to and like reset my passwords but and then imagine like having it in a cloud browser somewhere and like maybe have like a couple like oh my God so much maintenance yeah with this one it's just within your own browser you don't have to it doesn't even have access to it like you're

it's assumed like these sites are logged in right like Amazon is like already on my account okay yeah it's already on my account I just happened to be logged out but uh no yeah or yeah even LinkedIn I'm logged in so it doesn't uh you don't need to share anything with us we don't even take passwords for the extension itself just do Google sign in so when it comes to this this product obviously there is a wealth of you know industries that would benefit from this because it is pretty broad but what what do you think some of the main uh use not use cases but like industry uh segments do you think could benefit from this because I know with Bine we me sales sales and scraping kind of go hand inand um and what's interesting to me is that could still be the

case you know scraping for that same reason and it actually kind of it's just so interesting because I mean you've been I'm sure around tech for a long time strifing was such a manual process like I remember it was like magical to be able to go onto somebody's LinkedIn and get their their email right for something like that but you'd have to go tab by Tab and tab by Tab and click it do you think that there and and this this is this is something that I've I've used before there's like products called um overloop for example this is a LinkedIn browser tool or sorry uh Chrome browser tool that allows you to go on somebody's LinkedIn and get the remount or the phone number right if they have it there does this would this tool be able to basically take uh a two Chrome

browser step sort of thing by telling it to use the other Chrome extension to scrap a bunch of uh to scrape a bunch of emails I I think that specific use case you gave uh it uh it should be able to do like uh but like no it unfortunately uh this since this is a Chrome extension as well it's kind of hard it can't interact with other Chrome extension got it okay so chro like Google set up a really good sandbox within Chrome for Chrome extension so they can be very secure and you don't to worry about anything but I could do that use case right now like click uh contact info and get arin's email or oh my God okay I hope this doesn't go wrong uh I mean I guess people would know the email they could just do it themselves so like

fair enough I see okay I see what you meant yeah yeah so it was able to do that action itself and get it and you could have it as part of so we have this other feature of like sheets workflows so you can get like a Google sheet of URL so maybe like a bunch of LinkedIn profiles and you could say like extract could you do a two stepper could you do a two- stepper where you you say you scrape like a an industry and for and get that on a Google sheet and then tell the Google sheet or tell the agent then to extract everything from that Google sheet like I okay yeah exactly so you can set like a input URL column it could be anything I could just do the one I have here and give the prompt that could in include

actions to take on the page and yeah you could do do that and we actually even we took we actually thought much deeper on this as well and I don't know if you know about clay which allows you to enrich like um information from other sources so we kind of have this deep research uh sheets workflows where say you have like a column of company URLs and it would and you want to enrich that information with like pitchbook or crunch bit and you could give the prompt for uh every single like company on this um in this column go get their headcount uh latest headcount latest fundraise uh what like current series like of like series a and it would actually research on these websites in parallel and extract that as a long row I could give a quick example um I had I had

heard of clay but i' never used it prior I think it yeah it's a little expensive so that's that's probably why so I had already done this a little bit earlier like I had a list of companies and I wanted their LinkedIn employee account I wanted like some features from their main web page and use cases so it's able to similar to like per each row open like five websites like LinkedIn crunch Bas pitchbook and research those uh that information on that tab itself like search on LinkedIn for that company go to that profile company profile figure out how many employees they have and extract it so we I actually didn't mention this deep tab research so similar to so if you had like couple of these tabs open like retriever Bine browse and you want to compare them by like companies by strategy features

pricing that involves you like trying to navigate these uh web pages uh independent like like take go to pricing page on browse go to pricing page on Bine go to pricing page here and do like this kind of research from those starting URLs and extract it to like a Google sheet of uh comparing these companies by these dimensions and this allows you to do like Supply like paywalled sites or sites that only you have a subscription for or like there's a login wall because all of those websites like Google deep research nor opening eyes deep research they can't crepe and so they don't have it as possible source for their deep research agent that does that make sense or yeah no it does absolutely so you could Supply you could since it's these Tabs are being open locally on your browser all of those websites

are now game for this kind of research and I thought like it's a much better representation to have uh This research represented as like like a matrix because like you know who who's reading all these reports that uh like six 10 page reports from Deep research what you really want is like a condensed uh representation and yeah let me try to make it okay maybe not this one I'm trying to do the wrapping okay so if you have like a condensed representation like this okay like I can just look at pricing for all of them over here and this is like much more useful to people that's our Theory than having like a 10-page report that from sources like Reddit like some person's random Reddit comment instead of sources that you can control and it was also relatively fast uh this this actually took a

little bit longer than expected uh okay but uh it should be faster if you try it out again but yeah so we think it's also cheaper and faster than those uh other Solutions interesting okay so uh question about this because obviously you keep you keep bringing it up and I think it's very interesting um you keep talking about the benefits of the fact that it's in the browser and I 100% agree with you that there's a huge benefits of it being in your browser you having the login capabilities but I have another question that obviously you're just a Chrome browser right now in the future do you have goals and aspiration aspirations for it to be a desktop app is that something that's feasible to kind of go above and beyond um the limit ation that exist within just being in the Chrome ecosystem yeah

so since I worked at Google I'm I'm a I'm a Google Fanboy I think Google Chrome is one of the most complicated pieces of software that the average person actually uses it has millions of lines of C++ that is highly performant and I actually enjoy being a browser extension because it's a Sandbox people trust it I I think I saw some other tools that were kind of like that hey install my app I'm like I don't even want to try it and I'm curious right like uh so curious to try it but like I don't want to install a random um desktop app that I don't know what it's doing but yep even it's someone technical that's I I have some hesitation so I don't have such an idea but I think there's some exciting things I'm really looking forward to like setting up Cloud

browsers that can execute things uh on your own schedule like independent of your laptop being on and one exciting thing I'm really thinking of is trying to do collaborative data sets right so I can maybe there's like a thousand URLs and I want to like with you like with a group of people try to like scrape these sites do actions on them and extract X data and with my own like laptop maybe it's uh not enough resources or like they're going to block me or it gets complicated but say you do 100 and I do a 100 rows and then together we construct a data set and like as you get more people it can uh scale up obviously but you can so that's the one thing I'm really excited for that I think is kind of the future because if you look at how

a lot of these scrapers work right they're kind of like using um having like a browser running the cloud but actually like websites easily block that like Cloud flare blocks those requ off the bat so what do they do is they hire they pay for residential proxies so they okay all these random vpns they're secretly probably like making requests for these uh so from the browser it's like making a request for this website but they route the request to these uh residential IP addresses and it's a little sketchy we don't no one knows where it's coming from my personal theory is like these random VPN providers everyone is installing they're making money on the side by providing these proxies but anyway so when it's that request is actually coming from someone's like home IP then the website actually like uh process that request and like

sends it over and it's kind of a long convoluted way to do this like scraping if you look look at the infrastructure it's also very expensive so it's like hey let's cut out all of these steps and do this collaborative scraping where I want to contribute to this data set that could be useful that could be useful for a lot of people but there's no like resources right uh currently that can support it okay so I didn't actually think of it until you you brought it forward with the Google Chrome being as complex as it is um and then obviously the your concerns about a desktop um agent um you know it's funny even a lot of people aren't aware of the uh uh chromium Foundation that exists on the majority of browsers um like for example I have a a pretty important question your

product is also available on the edge store right because I haven't looked into Edge or Arc yet uh it's on the road map thead okay this one of this is one of those unique situations where it actually probably be complicated um because most of the time it's pretty much a port from my understanding but since you guys specifically are dealing with the interface and you know like interacting with the interface it I could I could see where there's actually some issues right um yeah I I I have no idea I need to look into the what is different about listing on the edge store or uh Arc as well these other browsers is just we're trying to ramp up like it's it's like a whole new world on like these kind of things that we can do with like uh getting AI tasks on the

browser yeah that's yeah well actually no you actually no it is just a port still you're actually able to just I have it on my uh Edge right now oh basically the way that it work well the way that it works is there is no real uh Chrome there's no real like Microsoft Edge store um it's just since everything is chromium based you can install anything on the gr Google Chrome Store on edge right okay so I don't know if you you've ever dealt yeah I I personally I hate to Hate to I know you're a former googler um I actually am an edge user um really I think you're the first one I met yeah there was something that happened a couple years ago I used Brave which I don't know if you ever familiar with that one um that's generally used by uh

Gamers and crypto fanboys but then um I got annoyed with how it handled cookies and then I switched over to Edge um don't have that much of an issue with chrome but I guess now we've kind of settled that situation because I had a had a moment where I was like can I download this because I used it on Chrome because I'm you were very because when you said you're a googler and how important it was I Lear it so I'll try it on on Chrome but yes you can install it on any of the chroman based browsers it's something that is a little weird and complicated but um yeah no I I really like uh I really like what I'm seeing with this product I I had one specific question as well as I saw when you were doing the previous uh prompting for

the Deep research I noticed something that I'm hoping is something that I think it is there was a screen recording or a demonstration option like there was something was like what is that cuz I is this giving it context as to how to do something via screen recording so yeah we have this Advanced options where you can like select previous tasks or like um you could record uh you demonstrating an action like maybe on LinkedIn like the search thing because the search filters could get very complicated to search and with this when you click record it actually records the Dom interactions hey this element was clicked this we're not actually recording the screen itself so on like an RPA like an RPA Microsoft power automate extension but um you're leveraging your AI on top of that so it's yeah so the this the sequence of

actions that you did we add it to the models contact so it figures out when you it tries it again maybe I should uh click this thing instead of going off the rails so yeah and you could also give like a con sheet as context and uh do actions like that as well but yeah I I think what I uh I really wanted to show is that we also have function calling built in in that you can Define functions that uh you can like apis you can call so we have one that ships with like send a slack message so you could Define this uh function itself it's a little bit more complicated like we can uh for non-technical people I we appreciate that it's complicated but you could just type with natural language and like have a convert Button as well and since they're

sharable we're hoping people can uh it just takes one person to figure it out and set something up that is usable for other people but say I have this slack message function and I'm on this page and like I'm just saying send a slack message to the channel of the top headline using Slack function I just had to sometimes it's like you know with llms they it's iffy whether it like actually does it or not but say I have this um just with this natural language prompt the llm can figure out hey I need to make these arguments to this uh function and this function gets executed locally within the Chrome extension so nothing like those access tokens they never leave your Chrome extension so it doesn't leave and uh the locals um browser so I had this already set up so it's like time

is 450 and it just like uh sent this message of like Trump moves to like impose tariffs right so that was like the top headline right now and it yeah it called it by itself and executed it and you could think of use cases like I need to load this to my CRM database or load load this lead to my CRM database or it's like endless possibilities yep hey can this be utilized can these actions they're called actions right or functions sorry can these functions be utilized with the recurring um uh option recurring background back yeah the schedulers or no you uh so right now no but it's not that hard we just like slammed with work uh that's okay yeah yeah but it should be feasible within a month that uh they should be yeah you could have a prompt it could execute like

on my opinion that's that's ideal right that would be something where you can uh rather than it being limited to like Google Sheets or something like that it could um hopefully you know when it finds a lead or something right it could I use notion for like all my management of like work so it could maybe add use the API to to create pages instead um or something to that effect because I um I'm a very big uh user of apis I live on make.com um I love it it's like I I mean I spend lots on automation per month so perfect yeah no and so if for example this would uh have the ability to do you know research in such a way where it is kind of doing work that I would do in a browser um then I could for example populate

that information inside of a um inside of like you know a database that makes sense um even if it's uh you know a little bit complicated I would love to do it like for example you want to hear something crazy there is uh you can actually this this product I I guess for the public that's maybe unaware of this concept retriever in theory could be like a money saver for other products that charge too much for apis um like so for example if you found out a if you found a way to have you familiar with like sem rush you know what sem Rush is yeah vaguely is it like a SEO tool or correct yep yeah it's an SEO tool so sem Russia's pricing is like decently okay until you want to pay for the API it's this is not an exaggeration $500 a

month if you want to go to the plan that has an API the next closest price is $249 and then it's 139 you can basically with this tool if you trained it on how to figure out how to find low keyword difficulty High keyword volume keywords and you wanted it to populate in a database in a separate system you'd have a research on the user interface and you're logged in semrush account rather than needing to pay the 500 you know like 500 a month to to be able to utilize the API and that's that's some of the applications I'm seeing here a lot of people like um like price gouge for apis and it's really weird um and you can kind of like circumvent it in a way we want to meld the interface between uis and apis so like have retriever be like that

melding layer to do workflows with on the web pages as well as apis with just natural language with you just writing sentences and yeah we want to make make.com like similar level functionality just accessible to people with uh within their browser so one thing is that you can compose like maybe you have like uh you can compose a sequence of tasks right like do this action and then do like this other action and you could uh so you could kind of do like those make.com pipelines just all within the browser without like doing complicated drag and drop or like figuring out all these apis it could just have it as sentences and go run with it yeah yeah I mean because in some circumstances you're going to need like web hooks um and like instant web hooks for like for example we the reason I'm

big into it is I I deal a lot with content and a lot of contents like if then workflow stuff um for like status changing and stuff so that probably wouldn't need to be changed with with what you guys are doing but on the research side you know for sure it's like apis are limited by specific endpoints and actions right and if you can utilize the action of the API that is in add new pay for example Le in in notion or in Google Sheets like add new row in Google Sheets or add new page in database but then have everything pre that ad pre that ad step be a AI agent interacting on the browser and then extracting the data and putting it into the ad page that is like much simpler but a lot of people don't uh but but like for me

I'm just attempting to do this the complicated step by steps without the uses of the you know the use of the browser but for me I'm like yeah this is this is great because I I would I don't have to deal with all that nonsense anymore because it's it's a little annoying when um apis are are limit in such a sense you know like I have so many complaints with so many different products for like oh you know like everyone's been requesting this one Endo for like a year you know or two years and they just haven't bothered but with this it's kind of a circumventing of that situation so it's it's very cool yeah and hopefully I think it's very really targeting people even less savvy than you are with these make.com pipelines and someone who has the need for some task to be

automated but they're not technically enough even to set up like a make.com pipeline and they don't want to but like hey I can just with the a sentence describe that task execute it see hey it works and then put it on a schedule or like make that a step in a multistep pipeline and just get it to run locally so we're hoping yeah to yeah like but just even people even less technically Savvy and I also I had this on the browser the whole time but like yeah we have a graph bot feature where you can to anthropics artifacts uh have self-contained artifacts that you can dis like get executed and run like so for example any like llms are not only great for data extraction but also like reformatting it like so maybe like you have this table and you're like I want to

set up a graph of this instead with this graph bot Fe feat and you can prompt it like please make a bar graph so the we're taking all of this website to the llm and the llm is writing the code to create this kind of graph artifact similar to anthropic uh Cloud artifacts and we're hoping eventually we'll figure out like other useful artifacts that maybe like every morning you have like a table of news from all these different websites that you want to execute every morning and you just want to see it or yeah we're open to ideas very cool no I mean there's a lot there's a lot you can do with this and there's a lot that it seems like you're planning on doing um but as we kind of I mean we're getting close to the end of the the length we

usually run episodes but I think this has been this has been very informative you guys got a lot of very cool stuff so let's um let's kind of bring it to to a close here with what your your final uh showcase is for for everyone yeah I think uh that's pretty much it we were just trying to compare it to the whole landscape I like iterated on like that we're basically taking a text approach can do multiple actions in the background targeting uh productivity compared to like operator and others targeting like booking flights right I that's very complex thing and it's not repetitive we think the productivity is perfect for like these manual and repetitive tasks and the benefits of being a Chrome extension like kept mentioning that as well and how we distribute subtasks instead of just doing like a huge like sequence of

steps on one tab is leads to like much smaller failure rates and that we have the way that we're setting up function calling allows us to do like user defines any third party integration but like for operator it's like hey they have partnership with Expedia like custom so it's not scalable yeah that's pretty much it but uh thank you for giving me some time and hopefully we'll run into each other again Demetri and I I usually follow your videos in podcast anyway so I'm going to see you but I'm all over the Internet um someone say to much much on the internet um but definitely but definitely but no I appreciate it and definitely everyone uh who's interested in having a product retrieve research and automate uh things on their day-to-day I I would recommend checking out retriever.com for a lot of people you're going

to not blow through that um and then the there is a pro and Enterprise option um as everybody in SAS has the same pricing model no no no offense to you it's the world we live in uh but no I think I honestly man this is very interesting you guys are so early on that I'm I'm I'm going to be keeping an eye on you closely because I'm an Automation and uh AI nerd so I'm going to be like all right let's finally get some uh let's finally get some automation scraping tools that I can I can stomach because for a while I've been like dude I'm I'm just I I I appreciate what everyone's trying to do this is actually this is actually impressive this is the first time I've ever seen someone just type a prompt and it's like oh yeah we can

make the columns for you like that's been something that's been frustrating the whole time that you have to like for lack of a better phrase like schema uh yeah I have to like be the I was trying to think of exactly I I I saw your videos and I was like dude this could be so much easier I know yeah and and and it's not like these you know we've come a long way from the RPA days of like Microsoft power automate like five six years ago where you'd attempt to have it click on the same well web elements over and over again and it would fail so we're a long way from that don't get me wrong but it's this is the first example of where I was like all right I type a prompt out it figures out contextually what would be some

good St on the page to scrape I don't even I don't have to hold its hand but I can hold its hand if I want and it'll still figure it out so yeah everyone check out retriever.com