Name: Tavily 101: A Real-Time Look Into AI-Powered Search for Developers
Uploaded: 2025-12-10T18:16:12.763Z
Duration: 43 min 52 s
Description: Tavily 101: A Real-Time Look Into AI-Powered Search for Developers

Transcript for "Tavily 101: A Real-Time Look Into AI-Powered Search for Developers": Hello. We are going live. Hello? We are live. Hi, everyone. If you can hear us and see us, write in the chat where you're tuning in from today. Nice. It worked. Wow. Raleigh, USA. Windy city. Aon. Is it windy? Great. London, Germany. Wow. Peru, Italy. That's exciting. I wonder how the weather the change in weather between all these places. Here I am in New York. It's getting cold here. Yeah. Yep. We'll give everyone about another minute. So in the meantime, drop what the weather is in your area right now. Snowy in Toronto. Oh my gosh. I love the debate between Celsius and Fahrenheit. I never know which one to use. Yeah. Well, welcome, everyone. We're super excited to do our first webinar. Gonna let Jackie kick it off if we're ready to go. What do you think, Jackie? Yeah. I think let's take it away. We appreciate everyone joining today. This is our first live virtual event and definitely not our last. We want to bring these about more often, more frequent, right, so everyone could kind of understand what Tavili is and the power of Tavili. So with that. said, let's kick it to the next slide. So your hosts, you see them right here. So my name is Jackie. I just joined about a month ago, and I'm heading up our marketing team here at Tavilly. Super excited to be here. It's been a wonderful experience so far. The product is so powerful, and the people on the team are just so easy and fun to work with. And then, Dean, I'll let you introduce yourself. For sure. Hey, everyone. I'm Dean. I lead our FTE team. I've been here for nine months, so since, I guess, the earlier days, and super excited to do this webinar and show you guys what we've been up to. Oh, David, the good how do you pronounce Tivoli? We were actually just talking about that before. Dean, how do you say. it? So the the true pronunciation is Tavili, but we're we're flexible here. We get all different pronunciations. And I think it's a Hebrew term that means to retrieve. Yeah. In Hebrew, it means, like, bring me. Mhmm. A little fun fact of the day. Yeah. Okay. Cool. Let's get into it. So here's our agenda. Right? It's pretty short and simple, and we want that to be the case. It's intentional. We wanna make this very casual. We want to learn together, ask questions. Dean will be answering your questions. So we'll go through some quick housekeeping items. We'll learn high level what Tivili is. Then Dean's gonna take it away with a live coding demo session. Right? And that's kind of where you can ask your questions. He can answer them live, get everything that you need to. We'll run through some use case slides, and then we'll save time for q and a. But as we said, this is fun casual session. So if you have questions, you know, during this time with us too, please ask it, and we'll make sure that it gets answered. Yeah. For sure. Hey, Henry from Houston. That's about three hours from me. Love to see another Texan in the chat. Quick housekeeping items. So, yes, this session will be recorded and it will be sent to you afterwards. But being live is 10 times better in my experience, but, yes, you can watch it again if you'd like. You could send it to your coworkers, your friends, your family for Christmas, all of that. You can ask questions in the chat as we mentioned before. Resources will be shared after as well, and then please stay until the end. Right? Because we wanna make sure that we're in this together and then if there's, again, any questions or that you have, we'll get to those towards the end and the middle of the session. Okay. Alright. I'm gonna kick it off to our expert, Dean, and then I'll be here for the fun and the hype, and I'll see y'all in the chat. Okay. Let's get to it. So for those of you who are new and aren't so familiar, Tavili is an AI native search and research platform, which gives LLMs and agents access to fast, reliable, and up to date information. So this is at a high level what we do. We're gonna go into detail. And if you know us well, we're gonna tell you about everything that's changed in the last few months, and we've been up to a lot. So you might ask, why would an LLM or an agent need access to real time information? And let's imagine here what would happen if you didn't have access to real time information. So assume you have some chatbot or LLM and you ask a question that requires up to date knowledge. So you could say, what's the weather right now in New York? And what's gonna happen in this scenario is the LLM or the agent will refuse to answer the question because it does not have access to real time information. There is an inherent knowledge gap or knowledge cutoff with LMS. They're static, but the world around us is dynamic, and we need to be able to bridge the gap so that we can answer questions about what's going on in the world day to day. And the Internet has been around for decades, but in the last two years, it's undergone a significant shift. So if we think traditionally, many years ago, the main user of the web was a human. You would go to Google. You'd search, scroll through links, and read websites. But since the explosion of AI a few years ago, we now have a different way that we interact with the web. We now have an intermediary, which is our AI system, And this AI system will then communicate directly to the web, and then the information will be passed back to the human user. And since we have this new inter intermediary player, a whole new infrastructure gateway is required to enable this web access, and that's where Tavili comes into play. So we bridge the gap between the models or the AI agents and the Internet. So how exactly does this work? How can your you as a developer build an LLM or, an agent that connects to the web? The answer is our product, which is the Tavili API. And our API has a few different endpoints, which allow the LLMs to interact with the web in a variety of different ways. To better explain these different endpoints, which are search, extract, and crawl, let's think about how you as a human would interact with the web. So you might first, perform a Google search, then you might find an article that's interesting and open that article and read it. And then you also could click on links within articles and do deeper dives. And our APIs are analogous to these three operations you would do on the web. So first, we have search where you input a query and you get a list of ranked URLs. But here, the results are tailored to LOMs and not humans. So it's no longer a human clicking links, but this is a search that can take in high volume of data and requires refined snippets of information for each URL. So that's search. Extract, you could think of it as when you are clicking into a link and reading further. So you as a human, you you would find a very interesting article and click into it. Extract can take in a batch of many URLs simultaneously. So the human is limited to only click and read one URL at a time, whereas the extract endpoint can take in a batch of up to 20. So this is efficient, scalable processing of web pages. And the output will be the text, the images, and markdown of what actually exists within the articles. And then crawl is when you might wanna dive deeper. And given one web page, you might wanna find all the links that exist within it. So you can get this deep, drastic coverage of a web page with our crawl API. And later on in the live coding session, we're gonna show exactly how you do this seamlessly. So at a very high level, there are a few pillars that we at Tavili really live by. The first, is a unified model agnostic infrastructure layer. So when we say unified, what we mean is it's one API with a variety of access points to the web. And model agnostic means that you as a developer can bring your own model and connect it to Tavili to interact with the web. So you're not locked into a specific model provider. You can take advantage of the fact that new models and new providers are released on a monthly basis at this point, and you can just swap out the models and plug it into Tivoli. Then we have the firewall between agents and the Internet. So Tivoli really is serious about security. And what we do is we're the firewall between your agents and the Internet. So all information leaving your agent and going to the Internet and then all information being returned from the Internet to your agent. So we protect the data flow and secure it in both directions. Then the horizontal layer across verticals means that we, we can apply our technology to many different domains, such as legal, finance, health care, and we're gonna go into that further later on. And then developer focused and enterprise safe. So we care a lot about security, scalability, and production grade readiness for enterprises, but we're also, developer centric. So we care about the developer community, and we want a product that serves them. And later, we'll discuss all the integrations we have, which make it easy to build with our API. Hey, Dean. There is a question in the chat that we should answer before. diving into live coding. Let me know if you don't see it. For sure. Yeah. I see it. I see Tobias asked, how is the search conducted? Google, or do you crawl the web by yourself? So we at Tavili, have a sophisticated search system that is comprised of both third party providers and our own internally built index. And the way we think about things is we actually let the user define what URLs or domains we should crawl and include in our index. So we have all of our users that are hitting our API, and they're telling us exactly which domains to focus more on and to apply more compute to. And that's how we handle building our own index. Hopefully, that answers your question, Tobias. So we're gonna move on to the live coding component. So the idea here is I'm gonna go into the nitty gritty details of how to hit our different APIs. But before I do so, I wanna zoom out. Everything that I'm gonna show you is the fundamental building blocks on how you can build larger applications that have business value. So the goal is to use Tavili to enhance your to enhance your chatbots, to build company research agents, market researchers, research assistants, etcetera. And in later sessions, we're gonna go into these more advanced use cases. So an example would be if you wanna build a chatbot that can return answers on real time information and be grounded in citations. So this is zooming out, but now we're gonna zoom in and really, learn the fundamentals of our API. So I've pre prepared a set of Jupyter notebooks, which are tutorials. They're click through. We're gonna share them, afterwards so that everyone on on your own time, you can learn and replicate all of this functionality and build and hit our APIs yourself. But, essentially, what we're these tutorials start from the basics of signing up on our platform, creating an API key, which you would then have to paste into, this Jupyter notebook, and then running it. And we're gonna discuss the different endpoints. So first, we set up our Tivoli client, which is our Python, SDK client. And let's start off with search. So in order to build these research agents and chatbots, you need to be able to search the web as I showed at a high level before. Let's say you had the query NYC news. So what we do here is we hit the Tavili search API. We can set the number of results to five. And then on the output side, we're gonna get our five results with the page title, the URL, just a snippet of the content, and a similarity score representing the relevancy of this source. So this is a very basic search. This is your first Tavili search. But now let's look at, a slightly more detailed search where we're leveraging some of our API features. Let's say you wanted to find the new anthropic model releases. So one feature we have is you can set a time range. So since we're looking for new model releases, we might set the time range to month. And maybe we wanna check, only TechCrunch because that's our trusted reliable source, as an example. And then we can also leverage Tavile's news topic, which is a specific index that we have prepared, which has credible news sources. So if we perform this search, then we're gonna get a list of URLs. The first one is all of these URLs are from TechCrunch, you'll notice, due to the include domains feature that we leverage. And we find that anthropic released Opus 4.5, which is in fact true. This is their newest model that was released just a few weeks ago. So this is a more advanced search leveraging our parameters. Now let's move on to extract. So in the previous section with search, what we're retrieving is just snippets from web pages. You can think of it as a quick lookup, but sometimes there's a specific page that is really significant, and you wanna read it in its entirety. It's as if you're clicking into the page and reading the full web page. So that's when we would use our extract endpoint. And what we're feeding into the extract endpoint is a list of URLs, which is the search results from the previous step because it can take up to 20 URLs simultaneously. And then what we're gonna get back is the raw content. And when we say raw content, we mean the full, content from the web page as opposed to just a snippet from search. And then we get a lot more information. So this is what an entire web page might look like in markdown as opposed to the snippets from search, which are much smaller. So that's extract. Hopefully, you guys can start to imagine where you might, use extract in your workflows as opposed to search. Finally, let's move into crawl. So for crawl, you take some URL such as tvlead.com, and we're gonna crawl it and discover all the links that exist within tvlead.com. So we have our terms, our careers page, enterprise information, use cases. These are all the pages that exist within tivoli.com. And through the crawl API, we're able to discover all of them and scrape them simultaneously. So we also have access to the raw content from these pages. So if you think about it, in just one and a bit seconds, we were able to discover an entire web page, an entire domain, map it out, and scrape all of those pages and have all the content. That's, like, a lot of information discovery in a short period of time. Our other endpoint is called map, which is a subset of crawl. So what map does is it discovers the URLs but does not scrape the content. So maybe you just want to, kind of map out a web page, but you don't want the content from all these websites. That's when you would use our map endpoint. A final feature that I wanna showcase is our AI native instructions feature. So let's say we want to map or crawl to billy.com. But rather than finding all the pages, we want to specifically look or we wanna narrow our crawling to some instruction. So we might say find only the developer docs. We run, map, and it takes a little bit longer because now we're actually using AI to guide the site mapping. And the results we see are all from Tivili's documentation. So they're all nested within tivili.com, but it's all semantically relevant to the instruction that we've inputted. So now this is a whole other level of intelligent crawling, which can be used to build deep research applications. So this is the one zero one walk through of how to use our APIs. However, we have a whole tutorial set. Our second tutorial is how to build a web research agent, which makes the use of these APIs. That will be the topic of our next, webinar, but that kinda wraps up the live coding. So I see a lot of questions in the chat, which I'm we keep quickly looking through. I'm gonna try and handle them at the end during the q and a. So super exciting announcement from the Tivoli side is that we just launched our fifth API endpoint, which is our research endpoint. So what I've shown you so far are the building blocks that you would use to build a larger solution such as an agent. But now we're releasing an out of the box agent in a single API, and this is our research API. We evaluated it on a standard benchmark on Hugging Face. So it's a neutral third party published benchmark. You can go look on the Hugging Face deep research bench, and you'll see that we're currently state of the art, outperforming all the other research labs, such as Gemini, OpenAI, Claude, Perplexity, and Kimmy. So we're super excited about this new API that is in private data currently. If you want to sign up and get early access, reach out to the team, and we can give your account, special access. We also work with some of the best companies in the world, some AI native start ups, more classic trusted companies who have been around for ages. So we're super excited about who we work with. And now let's discuss some common use cases for Tivoli. We have we we like to think of it in three main buckets. There is research agents, enrichment agents, and AI assistants. AI assistants are what you would think of as most classic AI consumption. So chat GPT, different chat vaults where you're connecting, to the agent in real time, and you want very low latency. So this is one use case. We have enrichment agents. You could think of this as if you want to enhance some dataset with fresh web data. An example of this would be a CRM. So your CRM is a database, and maybe you have information on companies and people, but you want to enrich it. You wanna know what are these companies up to, how have they performed in the last quarter, have there been any new roles at the company. So we can enrich we can help you enrich your databases with, enrichment agents. And then there's research agents. So you can think of this as deep research or some broad agent that will go to the web, do many searches, summarize information, refine, and synthesize. And this iterative process is, what we call a research agent, and that's exactly our new research API. We also as I alluded to earlier, we are horizontal. So we support many different companies and developers with across all sorts of use cases. You could think AI models, coding agents with JetBrains, go to market agents, so sales type agents, financial services, chatbots, cybersecurity, productivity tools like monday.com, legal tools, insurance, big pharma, deep research, mortgage. So we're really supporting people with a whole variety of use cases because at the end of the day, all these different fields of work require data enrichment, research, and grounded AI assistance. Previously, I mentioned that we're we care a lot about developers and the enterprises, and two ways in which we do that is our availability on hyperscaler cloud marketplaces. So for enterprises, you could think AWS marketplace, IBM marketplace, Azure Foundry, Snowflake, Databricks. And then for developers, we have low level integrations with all the different frameworks, whether you're a lang chain builder, LAMA index, n a n. If it's low code or pro code, we we've got you. Now I will pause for questions. Jackie, not sure if you're still there and have Of course, I'm here. okay. Maybe we we can tackle the questions now. Yes. Okay. Let's start with Chris. So he said, does Tavili foresee an evolution form, I think, from the single shot nature of the API in order to support future Multimodality. multi. multi modal yes. All the technical, terms. Chris, when I think there's kinda two questions here. One is, does Tavile foresee an evolution from single shot nature of the API where you just perform one search and provide the response back versus more of a multi shot with reasoning where we might search, iterate, perform more searches. So this, moving away from single shot to having more iterative process is exactly what we're tackling in the research endpoint. And then multimodality, which is totally separate. So how do we search over images, videos? Currently, our API supports image search. However, there's so much work to be done in this area, and this is something that's actually on our road map. So we're really digging into how we can understand images, use it to enrich context of web pages, and that that's actually something we're working on right now behind the scenes. And, Chris, if you have any ideas, right, or you can always shoot us an email. We like to build alongside people as well. Yeah. Okay. For sure. Frank said, what is the advantage of using Tivoli instead of custom combination of Bing search slash Google search for grounding. and LLM? Yeah. So the main difference between Tivili and Bing or Google with grounding is we decouple retrieval from generation, which prevents model lock in. So we're fully model agnostic. Bing and Google, packages together the web retrieval with the LLM generation. So you're getting a AI response grounded in web data. However, we provide access to the raw web data so you can bring your own model. You might wanna build a system that might have access to 10 different models depending on the, context limit or modality. So we want to provide developers with maximum configurability so that they can build the best agents that meet their use cases, and that's the main difference would be in this model agnostic component. Awesome. Okay. Moving on. So how does Hivoli get rid of misleading content? For example, I have a topic of x y z, and there are different references with a few perspectives. Let's say three perspective read less. This is from Pavan. to okay. Thank you, Provan, for the question. I'm trying to understand a little better what you mean by misleading content. So we provide back end references and with content, and then it's really up to the agent or the model to decide which information is reliable. We provide similarity scores as well, but the source attribution component is done at the application or agent layer. So our search API just returns, the results. And then at the agent layer, you have to try and decipher what information is relevant and what might be misleading. Hopefully, that makes sense. If you have any more details, throw them in the chat, and I'll I'll try and provide a more direct answer. Awesome. How is Tivoli different from fetching the site map? So I would guess, Jaren, that this is, referring to our crawl and map endpoint. So the the whole idea is that we've handled all of the infrastructure. And with a simple API call, you can access the site map of a URL. And not only access it, but also retrieve all the scraped content from those web pages at low latency. So as I showed in just around two seconds, you can access the site map, get all of the content from all the web pages, and we have AI native features that allow you to guide your site map generation with a semantic instruction. So this is very different from anything else on the market. Great. And then I know we have some people helping us in the chat answer questions. In terms of the. deep research, right, is there anything that you have that you can show us or prompt any results, or is that still in production? So, I don't have anything prepared right now for this session. We do have a lot of material. We even developed a cookbook, which is a set of notebooks, Jupyter notebooks, that show you how to use the Deep Research API. Those are live and public. So I think when we send a follow-up, Jackie, we'll be able to share those resources so that you can get started and see it for yourself. I think in future webinars, we'll provide more content on the research API, and we'll also be publishing some videos on our LinkedIn and Twitter, showcasing this. Yep. And definitely, sign up for the early access on the form. I linked the blog and inside the blog, you can find that form and you can follow along with information there as well. Okay. How does Tivoli AI scoring and ranking. system decide on what sources are most relevant? What signals or factors does it prioritize priorities when filtering, the aggregated context? Mhmm. So we're looking at, variety of factors such as credibility, visit count. We're looking at the semantics or, like, the actual information within those URLs and how relevant they are to the query as well. And then we have, some other factors too that are going on under the hood, but, hopefully, that provides some insight at a high level that we're looking both at, semantic meaning, but also more traditional factors that you would use when building a search system. Awesome. And how does Tivoli handle query re rewriting result deduplication and source trust when multi agents with different goals hit the same topic simultaneously? Great question. I would guess that you're referring to our research agent, and we did a lot of work on deduplication, query generation, and how to spawn multi agents simultaneously. We wrote a blog, which maybe Jackie could share in the. comments, which discussed some of our approaches. But at a high level, you need LLMs you need to leverage models to write queries as opposed to humans because we're automating this process. You also wanna build a system that's super efficient. So when you're doing deep research over a long horizon, you get many redundant URLs, many duplicates, and you wanna deduplicate both at a global multi agent scale and a single agent scale as well. So we do different types of deduplication, and we're also monitoring the percentage of redundant results that are being returned by the agent. So if we and we call that search saturation. So if we find that, we're doing all these searches and we keep finding the same URLs, then our search is saturating, and we actually tell our agent that it might be time to change up, how your the types of queries you're you're using or how you're searching the web. So that's a high level of the technical approach we took to building that research agent. Awesome. Let's see. Okay. You mentioned the firewall layer between agents and the web. Can you share how you prevent prompt based exploits or injection attacks that attempt to manipulate the agent's browsing behavior? Yeah. That's a great question. So when I said that we're we act as a firewall and we protect the data that's being sent to the web and being returned, I really mean it. So we have a enhanced security layer for our enterprise customers, built in collaboration with a GenAI security company called Pillar. And this scans all of the web pages for prompt injections. And if we detect a prompt injection, we will block the result from being returned to the user. So we're ensuring to the end user that there will not be prompt injections Awesome. And then do we, you, use underlying LLM in Tivoli? of So it depends on the endpoint and the configuration of the endpoint. So in our research agent, yes, we're using lots of LLM inference. In our map and crawl, if you're using that instruction, then we're using some AI inference to guide how the site map is generated. For search and extract, there's very little AI, but it depends if you're doing advanced search or basic search. So there it it depends on the parameter configuration. But I would say that on the search and extract component, it's more about information retrieval rather than LLM. But there are little components where they exist, and they're all configurable. So if you don't want any AI, we can shut it off. If you do, we can enable it. Awesome. Vikrant, it looks like you have some implementation issues. If you shoot me an email, I'll connect you with our support team to get those solved. Anything you can dive deeper into, Dean, regarding the MCP server? Yeah. I'm so curious what specifically you'd wanna know. But from a high level, we have an MCP server. We have both a open source repo as well as a remotely hosted MCP server. So we've deployed it. You can access it very easily. We have different tools within the server, all of our endpoints, and now we're gonna be adding the research agent as a tool in our MCP server. And, yeah, that's that's our MCP server. Awesome. Scott, that was for you. So if there's anything more specific, okay, Just let us know. Alright. Let's see if there's anything that I missed. These are great questions so far. It looks like okay. Pavan said, so the decipher is manual or automated? She's they said you mentioned about credibility, whatever we are evaluating that will have a reference to evaluate. Is the reference manually selected? So, Pavan, are you referring to our search API or our research API? And depending on which one you're referring to, I I'll be able to answer. Search. Search. So we have a search algorithm, that takes care of all the background work to get you the most relevant data. But, eventually, that data will be consumed by a language model or an AI agent downstream. And different agents have different requirements for what credibility means to them. So we wanna provide you, the developer, with all the information necessary to implement that yourself. Hopefully, that makes sense. Awesome. I think I covered the majority of the questions. Let me see. Yes. I think we're good. If there's any more that come in, I can read them, but a few last minute things, Dean. Thank you so much. That was awesome. I love anytime people can live learn and and see things coding. Oh, we have one more question. Okay. How do we handle sites like BBC with explicit no scraping policies? Right. So our scraping systems, only access publicly available information. So if there are paywalls, we will not go there. We also respect robots.txt. So if a certain domain doesn't want us to scrape, we won't do it. Perfect. Anything for the road map for location aware research queries? Interesting question. I'd love to hear your ideas. We are thinking about this. We're definitely thinking a lot about personalization and trying to make search more catered to different niches or different needs rather than having it general. So nothing that I can share yet, but we're definitely thinking about this a lot. So Henry asked, JSON output for Tavile search. So for our research agent, we do. We have a structured JSON output. So you can define a long complex, schema, and we will populate it. However, for the search API, the idea is we're just providing you with the the search building blocks. You'll then connect that to your model, which will have some structured JSON component to it. And I think maybe we'll create a cookbook, actually, and show developers exactly how to do this. But right now, it's not built into our search API because we're really more focused on the retrieval component and enabling the developers to build that JSON output on their side. With that, thank you for mentioning that. I think we'll we'll make a resource. Yeah. I think this live event was great, but it also gave us some good things. to put down for the road. map. So we appreciate you all. Taking notes, Jackie. Yeah. I think we can wrap there. So I put this in the chat, but we love to get big and loud on social media. Right? So if you make a post around your experience of this live webinar, tag us. Right? And then shoot me an email, and we can send some free credits your way. That's also a token and a thank you for joining us live today and staying with us for about forty five minutes. Other things, if you wanna talk to an expert on our team, you can hit the button up top, talk to an expert. It is in light blue up there. And I will also post I post this earlier, but if you wanna log in or sign up, you can do that in terms of the trial and get your hands a little bit dirty there. The recording will be sent, and we still have to plan the next session. Dean is a busy man, so we are working on that. It will be in January. But, Dean, any final words? Oh, thank you all for the great questions. I love the questions. It gives us great ideas of what to work on and what you appreciate, and, yeah, thanks for joining. Thanks so much, everyone. Bye. Bye.