Agents By the Bay: Slava Tykhonov, Semantic Croissant

Agents By the Bay: Slava Tykhonov, Semantic Croissant

Recording: Agents By the Bay: Slava Tykhonov, Semantic Croissant

Hello everybody. My name is Alexey Khokhlov and I'm a community leader in open source science and graph AI in the Bay Area, founder and organizer of AI agent the self AI by the Bay and recently founder of the community research center for reliable AI at Northeastern University. And today on our podcast I have a great guest, Slava Tikhanov from CODATA. Welcome Slava. Hi everyone. Slava was the software engineer and an engineer at the Royal Dutch Academy, ambassador of Dataverse and he's also the co-author of the Crossref standard 1.0 for notation of machine learning data and also the creator of the semantic Crossref uh standard specification and multiple technologies following from it. So today we're covering topics that allow AI on graphs to be more reliable and using graph approaches to disambiguating, and tracing interactions with LLMs and various data sets. So we have a lot to cover so let's get into it

First of all, Slava, tell us you are now the head of AI and interoperability at CODATA. Tell us a little bit about CODATA, what kind of problems you are solving there and basically how Dataverse and others interact with with with these problems. Yeah. Thanks Alexey and thanks for for being here also and I just want to tell a little bit about CODATA. So CODATA is a part of International Science Council and we're providing advice for governmental bodies, for any kind of institutional organizations who are interested in in getting advice on data and on artificial intelligence. We're also supporting United Nations. For example, we're working with Hazard Information Profile Group on detecting hazards. and we are trying to use AI of course and we are training our own models and we are working also on multilingual framework that allows basically to translate all these hazards in any kind of languages and recognize automatically in detecting any text

Yeah, so that's something we do and we are quite busy. We recently got two different projects funded related to AI and by European Commission and it's about cross-on and it's about cross-domain interoperability framework and about bunch of stuff that I'm going to tell today. Great and I must also say for our audience that Slava is based in Paris, London and previously in Amsterdam. [clears throat] So the reason why all these languages need to be translated one to the other is because it's happening in Europe as well as around the world and as we all know there are multiple languages in which business is conducted in the European Union, in the UK and so forth and there is a great tradition of all of this data you know needed to be translated and and uh Codiate also serves scientific institutions across Europe right and a lot of this research is publicly funded in the EU. So this is a great evidence of support public support for science and and and open source and interoperability and open data. So maybe we'll start with Dataverse because Dataverse you're ambassador of Dataverse there is a Dataverse site at Harvard and you build innovative search for it which is which is which is very interesting. You talked about it AI Alliance knowledge group meetings and I think it was huge interest how you use already basically raw data for, um, for more precise search. So, maybe you can talk a little bit about what Dataverse is and, you know, what kind of, um, innovations you added to make search over it more reliable

Yeah, so Dataverse is is a, um, open-source data repository. It's originally maintained, uh, by Harvard University who created this repository about 20 years ago. I think 2006 is the official date. And, uh, basically the idea that, uh, there is, uh, platform, so every researcher can can get account and upload data and describe data, so produce some metadata and basically publish that and it will become citable. So, you can confirm all your evidence, uh, based on on you can provide all data sets confirming your evidence, uh, described in your paper and basically this is how research world works. So, Dataverse is not the only only platform, of course, but, uh, one of the most popular and also it's very popular now in in European Union, so a lot of countries already started to use it and, uh, what I actually did, uh, I'm doing that for for decade, even even like like 11 years, I think I I started 11 years ago. When I joined community, uh, we started to work on on a first on cloud platform, they tried to move Dataverse to the cloud and we tried to to, uh, produce Kubernetes infrastructure. Uh, well, of course, the crisis that and, uh, nowadays it's very easy just to basically to use one-liner to to run Dataverse and install it in on even on your computer

So, it's universal technology which could be used everywhere and, uh, it's quite modest software, uh, not not software, more modest hardware which is required to to run Dataverse. You can just run it on your computer and it will be available in 5 minutes or something. So, um, related to AI, um, it started also a couple years ago when Harvard University decided like, you know, to to investigate what is possible with artificial intelligence and Google contacted Harvard and they decided to work together and it became basically cross-on standard. So, Google had this idea to create replacement of schema.org which is also suitable for artificial intelligence. So, Meta got invited NASA and as a partners and I was participating in this group on behalf of Harvard. So, I was I was also working on standard and I started to think how to use this how to apply this standard right away. So, first application I've developed was pretty straightforward. I connected Alama and I started to to ingest data from cross-on and I discovered that it's actually very useful to to get this as a JSON-LD a kind of navigation tool

So, inside of cross-on you you can you can define all all all places where AI should look for if someone is asking question. So, for example, if you're asking about authors or about topic or about keywords, it just goes through cross-on through the graph and it knows where to find it and give you back explanation. And it was quite efficient. So, it it was the first tool that I've created at that time a couple [snorts] years ago and after I decided why not to extend it to like question responding system. So, you can ask any question on any any stuff describing data set. It could be in metadata, could be in file, could be everywhere and basically I implemented this ask service. So, there is ask data service that allows judges to chat with data repository asking what kind of questions. And even more important that it's implemented in distributed way

So it's not ingested inside of the system. It's basically it just goes on the on the graph. So if you're asking question, for example, find something about climate change, I don't know, in US or in in other place. Or just give me a second. I need to respond. Yeah. Just second. Sorry, I'm back

Yeah. I'm sorry. So I was talking about ask service. So it allows to to ask question and basically it knows where to find information and it's croissant can can be seen as navigation system in your car. So if you want to go somewhere, you're defining city name and after address and after number of building and basically you go there. And this is exactly same purpose of croissant. So there is a service now called as data verse that allows to ask any questions. It just knows where to find, in which country, in which data verse, how to query data verse, how to get stuff

How to understand what what is what is in croissant and basically giving you all these answers. So it's based on the fact that underlying data sets are in croissant format and so the data verse is also a distributed collection right of these data sets and I think the number is approaching a million if I remember correctly, right? The different data I I think I think about 750,000. >> Yeah, yeah, under under a million. Yeah. So these are all basically hosted in different places, different countries. So you ask service routes the queries properly based on the croissant metadata in these data sets. Yes. Yes and and the So again because if you're using some some modern tools like croissant or cloud, it knows how to navigate through API

So you don't need to teach that. But the main point that it will get structured data back. So in this structured data is already properly described where to find information. And this is the main goal why why croissant was created to make it universal for any kind of models. Mhm. Mhm. And so croissant is basically annotation of data and it's a JSON metadata. It's a JSON blob describing various properties of the data set and so the standard basically defines what the fields are, right? And the necessary That's right

Yeah. There are few things that are missing in the in the original spec. So like in in 1.0 we we didn't cover control vocabulary support properly and ontology alignment. However, this is extremely important. So for example, if you want to use if you want to build multilingual search, so you're asking questions in different language, you you're searching information in Chinese or in in Italian, in English, it doesn't matter. So basically the idea that in croissant could be extended with reference to some control vocabulary uh and this control vocabulary It contain all variations, all translations of specific term. So, AI can understand and it can find it. And by using both relationships, you can actually get a graph

And by analyzing this graph, you can also get precise understanding what data is is describing. Mhm. Mhm. Right. So, I think logically, and you mentioned basically control vocabularies and you mentioned ontology. So, I think we're logically progressing to the CDIF network and framework and the semantic resource. So, can you tell us how how like basically in order to define semantic resource, you had to first define CDIF, right? And so, tell us a little bit about about that. Yeah, so so CDIF was created to to make this transition of data between different domains transparent and You you can think about CDIF as a kind of way how you can represent exactly same data set which should be used by different domain

So, basically, what it's doing, you're you're taking your data set, you applying different ontologies and different control vocabularies being used in a specific domain, and you are repackaging everything and you're delivering. And people or AI agents on that side, they're basically taking this and because it's in structured format, so it's also supplied with Cross Sun, you you can get all information about provenance, from where it's coming, and who has created that. And CDIF is basically semantics. It defines semantics. So, you can get precise understanding what is inside and what kind of relationships all variables have. We're also operating this concept called variable cascade, which is basically indicator. It's kind of complex indicator consisting from different variables. And you can represent any kind of phenomenon by using this approach

Mhm. >> [clears throat] >> Uh we're also using very So we're defining variables on very precise levels. So it's like you will have units of measurements, you will have some descriptive statistics, you will have also uh classes and hierarchies belonging to variables. So this information expressed as a graph as a subgraph allows to tell any AI model precisely what this variable is describing. Mhm. So people understand and because it's also multilingual, so we are using all these external control vocabularies like ELST and uh we are defining also linkage to to other domains. So we can basically transform data to any format and deliver quickly. Interesting

Can you give an example? I think there are sites where these mappings between uh domain ontologies are held and for folks who are not, you know, working on this everyday, can you give an example of when you need to map one domain to the other? What would a typical mapping look like? Okay, so I I think the most obvious example is is like GitHub. So if you think about software Mhm. and in software [clears throat] we we we have a couple of standards and the most popular is is code meta. It's called. >> Mhm. So basically it has like own fields names and own descriptions. And if you want to basically to to get this standard into Crossref, for example, you have to create mappings. So you have to map title to name and you have to map like field uh I don't know called license in into like digital properties

Mhm. So I I don't remember from my exactly how it's called in this standard. So, this is where I need to to create all these mappings and I have to describe properly publish mappings separately and after I need to apply transformation steps and I should I need to get this JSON-LD completely transformed and be ready to consume by other applications. Mhm. So, this is where transformation is really required. And we do this it's called this process called ontology alignment. We do it very often. So, basically we're creating all these semantic mappings between different standards to get after new graph and after graph will become queryable and you can find information

So, this is absolutely correct. >> And you mentioned also unit of measure and that comes from scientific domain, but you I think emphasize how units of measure uh make the meaning of variables much clearer, right? And that has a lot of value. Can you talk a little bit about it and how it helps scientists to define these variables and variable cascades with units of measure? Yeah, so let's talk about like specific example that that we have in one of the our projects. We're also working on a few projects related to climate change. So, obviously we're tracking temperature. So, temperature could could could be measured as Fahrenheit, could be Celsius, could be Kelvin, right? So, we need all these transformations to be done uh to do harmonization before we'll we'll start to consume uh all these data and we'll we'll get we'll have possibility to integrate data coming from different sources. So, this is specifically like this essential step before you will uh be able to integrate data coming from different sources and if you'll think about complex cases when you want to integrate, let's say, 10,000 data sets. So, basically you will recognize the same variable uh temperature everywhere

you will query, you will understand this is the same. So, after next step is just to understand what kind of units of measurements they're using because usually it's not even indicated. So, it could be like spreadsheet and you don't know what it represents. So, basically you're guessing. This is where you need to to go to human expert and to ask what do you think this units of measurements can represent. Basically, they're just showing a piece of data right from from spreadsheet. So, as soon as we'll have this human loop done and we're also building tools to provide this web interface where you'll get all variables, their descriptions also pre- sometimes predicted by artificial intelligence units of measurements. So, basically your human expert should be evaluating every field and correcting if it's necessary, also providing right units of measurements and we'll get really like verified data which are ready for integration

We call that ready data. Mhm. Mhm. And as soon as we'll get it, we can create any kind of like super data set on specific topic on specific assignment. Yeah. And deliver to researchers. And I'm thinking that you can also use AI to make reasonable assumptions. For instance, if the units of measure for temperature or scale is not shown and if it originates in the US, it may be Fahrenheit, but again depends if the scientific institution which was using Celsius

So, you can use some AI logic to make assumptions, but that can already make it easier for the expert then to approve it, right? But you know, maybe semi-automatic or automatic. You know, I think it's very interesting because you know, coming from functional programming languages, I noticed that there are like F sharp has units of measure which is a rarely used feature, but when it's used, it saves dramatically a lot of problems. And as we all know, you know, the spacecraft failed to fly to Mars because of this mismatch. So, like the the errors in units of measure can lead to a lot of errors, especially if given to AI. Right? So, so that's very interesting. Uh right. So, So, I think the heart of semantic croissant comes from the idea that you need these domain ontologies to understand various domains, and the best people to build these ontologies are actually domain experts, right? You cannot really try to do the whole work for the world. You cannot boil the ocean, and you need to figure out what is collaborative process by which you delegate ontology curation to the domain experts, right? Who will be responsible uh for for them

And uh and the croissant itself in its current setup doesn't care. It's just adjacent. So, if you want to add an ontology, you basically can add any kind of nested field, right? Because you can endlessly add JSON, but it's not just that because that field will have to resolve against some external server, right? And then eventually it will need to be routed to some ontology maintained by the expert. And so, I have an example in my kind of in my mind when I talk to people, if you have a BMW, you know, uh model X7, and you have you know, X3, uh and you have um Bosch repair facility, uh and you have uh you know, a provider and supplier of of the parts somewhere in in China, for instance, right? So, all the three of them need to refer to specific part or assembly by the same name. And like all the subcomponents of this will be actually have to map identically because if, you know, Bosch had this example as far back as a 2019, you know, there is a smart car, a part of it has failed, it automatically connects to Bosch, figures out what is the next appointment available, and tells its owner, "Drive me." Right? Maybe it'll be self-driving, so it will drive itself and and and the Bosch repairman already orders the part from the provider from the supply, right? So, all of them need to be on the same page and the part is ready and it'll it'll be inserted properly. So, so I'm curious like what is the vision for Semantic Web to make this, let's say I'm writing an agent which will which will create a plan for the repair of my car and it will have to find this part and disambiguate it. Uh what would be the architecture uh of this agent? And so, Semantic Web will need to connect effectively an ontology which is like I BMW is using some ontology internally. It may be, you know, in its own format, but it will have to be aligned to the one used by Bosch which has probably its own large ontology of all the parts it can have, not just for BMW, but also for Audi and other cars, right? And then the provider engine has a giant ontology, maybe it's building sparks for all kind of agents around the world for all kind of cars, right? But like So, they have a bunch of ontologies which need to be aligned as relates to uh a given part of a BMW

So, uh what would Semantic Web for this look like? Where this ontology would be hosted and how would you route this query, right? We need to Yeah. uh identify a part and and and get it, order it, ship it, have it ready. So, we uh we're discussing this topic and we this is very tough topic for us for us because uh knowledge partly is missing, so we don't have access to to the whole knowledge and sometimes knowledge knowledge could be proprietary. So, we came with this idea to use uh DID uh which stands for place identifier. And this decentralized identifier also it should be connected to ODRL which means uh open digital rights language. Mhm. >> So, basically in combination it allows also to connect like proprietary sources, also to to get them in inside of ontology, and also keep them as a part of proprietary source. So, basically these the ID which protected by permissions Mhm

it could be stored somewhere, but without having granted access, you will not able to to to get this information. So, this kind of ontology still still be staying at the original place, but we can query as soon as we'll have permission. Mhm. The ID is a basically glue in in our knowledge graph. So, we can connect any kind of knowledge with uh concepts, and also with ontologies, and I must admit that this is also moving knowledge. So, it's as soon as you have some concept, and it's becoming also fair, and it will get it will receive the ID, but still it it will can be moving because what is changing. So, we need to keep all this graph with all changes, all revisions, and this is where we are also using IDs. So, basically you you can see uh some concept could be translated in all languages, and for your example, like if there are some some parts for cars, you can also define all all this compatibility inside, so you can get all the relationships, and still um it's information which could be adjusted and changed

However, by by because it's already has the ID, uh all this new information also is becoming findable. So, you can query in your knowledge graph, you can see if someone will add this uh part is is also compatible compatible with another car, you will immediately see it because you can get this information from the graph coming into system. So, we basically came with this idea to to use as a this the ID uh identifiers as a glue. So, glue which connecting all these pieces together, and uh this open digital rights language, we can actually define who will get access. It could be also not human. It could be agent Mhm. principal. A agent which has specific access

So, it can request information from from some dealer and get this information and bring it into a system to extend context. And after this expert system based on AI, it will understand precisely what is available, what kind of replacement replacements are available for to repair car. And it can be also managed in really efficient way. Yeah. No, this is like the idea of the IDs is very interesting. And it really to me it's really amazing how you kind of mapped this whole area step-by-step because we talked months ago, and you mentioned the ID is an exploratory, you know, concept. And then now you already implemented it, and it gets traction. So, to recap for folks who did not know, and I actually did not come across this specific implementation

So, DID is the digital identifier, which is both unique globally, and you can verify it. So, this is effectively some kind of a blockchain ID in a good sense, in the sense that you can track its its provenance. You can you can ascertain so they issued by an authority, and and they signed. So, basically you can verify any specific DID that it exists, and validate it. But it's not just an ID. It's actually carries a payload, which can be, you know, like a random, you know, payload. And so, you came with an idea, and that existed for instance, you know, it's a part of the Azure infrastructure, right? So, it's actually because obviously cloud providers need something like this, right? Across applications to track some unique IDs which are verifiable. And so, and you found basically an innovative use case where you use the IDs to trace LLM interactions, right? Which to me is very important

A lot of folks don't understand this that everything that we do with LLMs is changing every moment because we query LLM, it's a non-deterministic process to begin with because LLM data set is changing every time, right? And there is a lot of randomness inherent in any communication. So if you have any hope to understand performance of your queries or why drift is happening which right we should dealt with in dataverse, you need to basically snapshot the world, right? And and so so the IDs give you this ability because they are unique IDs and they can be time stamped, right? So they can give you the trace of this interaction and then you can also have a chain of them, right? If you engineer, you know, the IDs as a sequence that your consumer. So so I find this and and like ODRL, I think it's again something I learned from you. It follows very naturally because obviously if you work with proprietary data, proprietary ontologies and you give them to LLM for context, you want to be able to debug it like Bosch, BMW and they, you know, change and supplier will agree of course that they see each others supply chain data, but nobody else hopefully will be able to see that. So within this group they will want to have a shared context that they feed into LLM to answer proper queries and they will debug if something is going wrong by looking at the sequence of the IDs and payload and queries for instance because you can wrap the queries, right? And so ODRL can be used to to basically constrain this access to these three specific, you know, entities, but also like they can can be probably limited to what kind of databases that can be using, right? Like you can always only access, you know, shipping data or part data, right? And not like HR data in this company, which it doesn't consider. So, so I mean, I find this extremely like it's falling like coming together, right? As as an architecture. And and you mentioned now there are teams, the projects funded by the EU to basically uh build on top of it. So, so how do you see the next steps? So, you you have the basic building blocks, right? You have CDIF, DID, and a DRL

Uh and like Dataverse is one of the examples. So, what kind of are the next steps for these projects? What would be the deliverables, uh the milestones? What are you going to build next and who's going to build it? How do you envision this ecosystem uh development? Yeah, so um I must admit that first step that and we already started to to follow this direction. So, we started to assign decentralized identifiers to every LM prompt. So, you can really make every prompt globally resolvable and shareable. So, let's say if you're doing something with Cursor, you you you can just assign some some identifier to the whole thing and it's basically compressed in your uh decentralized identifier. Which could be also used to, for example, to reduce uh context window for smaller models. Right. So, basically you you can query your graph, you can get um you you can ask to to to deliver some prompts and basically it's delivering you the DID

And by querying on DID, you can get sub prompt that you allow which allows you to to implement some skills, for example. So, this is very efficient and you can compress any knowledge by using that. And because it's globally resolvable, you can also construct graph and put everything in linked data. So, by apparently assigning the ID both to input and output, you can create your reference model. And you can show this reference model to to AI as well and create annotation and you can train new models on those IDs. So, I see unlimited potential of this technology and again, so we are just experimenting right now, but as I said, we are coming into implementation in two different projects. And we also have have couple of projects I already mentioned related to climate change research. So, we have a teams from different countries and we're going to experiment further how to extend it and how to use it as to create kind of sovereign technology to build sovereign AI which has also protected data sets layer and yeah, we are pretty optimistic that it's possible by using the whole by using this combination of both technologies

Yeah, and I find that DID wrapped prompt since it's globally resolvable and verifiable, you can publish it to the world. So, I can easily imagine marketplace of prompts or repository of prompts, right? >> And skills and skills, this is important. So, you you can basically produce any kind of skills and you can publish them and you can also prove that you are owner of this skill. So, if someone wants to use it or you wants to buy it something from you. Right. >> Or already started to use because I don't know if some somehow got published for public, but you want to to restrict access. You can always confirm that this is basically you, you know, because your own digital verification allows to do that. Right

>> This is I think extremely important. So, basically in any models we can now recognize uh some pieces created by specific companies or some people, right? And they can even they can show how to to prove authorship. Mhm. So, it's for software world, it can be really game-changer. Right. Right. Right. No, this is So, this is very generalizable

And uh I think another thing you recently published uh is the Pale Fire project. And uh you donated it to Linux Foundation, specifically to agricultural stack, which is very interesting. You know, I'm going to the Linux Foundation Member Summit tomorrow uh in Napa. And and I'm really curious, you know, can you tell us a little bit about this project and uh how it became of interest to the AgStack uh project? Okay. So, um couple of years ago, I created another AI-powered uh framework. I call it Ghostwriter. So, basically, Ghostwriter engine it became uh kind of engine for us Dataverse, uh what I already mentioned. And uh this uh engine uh is created on top of Quadrant

And uh you can use any kind of uh vector store. But apparently, it's integration of graph and uh vector storage. So, it's it's a kind of way I've I've tried to find the way how both uh both technologies can work together uh in in really like complementary way. So, if you find something in in in graph, and after we can find something in a vector store, we can also take take a similarity measure, and we can understand if it's the same thing. Or it's kind of it could be the same thing for 90%. Or can be not the same thing because it it doesn't pass threshold. But still, if uh if you'll add a little bit more information coming from knowledge graph, coming from from those the idea identifiers that which I I just mentioned, we can we can we can compute that we can recalculate and finally we we can identify that this is the same kind of concept. So, why it's important? So, for example, we are working also on hazard information profiles for United Nations and we're doing translations

Mhm. >> And those translations I mean, there's no problem that just to use any LM model that just to translate something in specific context, but the problem actually is to find some official terminology, some official concept which is accepted as a you know, like standard. Mhm. So, you'll get translation. It's also quite interesting how it's being translated by LM models because sometimes they do job even more precisely than than human experts. I will give you example. We did translation, for example, like catastrophe Mhm. or disaster

I think it was concept called disaster. Official Official name was disaster, but in French it was translated as catastrophe. Mhm. Okay? But we use different models, so it's kind of arbitrage system. We ask GPT also from Open AI and Gemma, Deep Seek, Lama and all of them said, "No, no, this is disaster." Mhm. So, so this is where you you have to to be really creative and you you should understand what is official name because it could be also catastrophe Right. that terminology which is accepted by United Nations. Okay? So, this is where you you have to to also going from graph

Basically, you are defining graph, right? With all translations and all information should be also ingested in in a vector store. And uh, you should take this similarity measure to understand if concept which is discovered by AI is synonym of what is uh, accepted. Mhm. And uh, so uh, it's it's a quite difficult task, I must say. So there is also like human work involved and uh, we can't trust completely AI. So it could be that like like if I ask in 10 models, all of them we are trained on the same source with the same mistake, right? Mhm. So this is exactly where problem actually started. So as a part of Pale Fire, I implemented kind of transition between knowledge graph and vector store because I really wanted to understand if you can can kind of compress data coming from vector storage into graph and >> Mhm

going back like like from graph to vector storage for some specific use cases. So for example, if you have some data set about earthquakes and you have some observations and this is just spreadsheet, right? So like data points. Can we turn those data points in something nicely describing, you know, it was bad weather and it was somehow like large waves started in the ocean and temperature started to to rise or stopped started to grow. Just just to basically to describe the whole picture and make it available for researchers. And make it for every observation. This idea I had around of Pale Fire and there is like first implementation which is already donated to Linux Foundation and we're going to test it. We're also going to use it for quite interesting use cases. So for example, they want to investigate and understand like how coffee prices are changing and why

Mhm. So by doing that you you have to to collect all information from news and from reports and from any kind of sources. And basically also to create knowledge graph, so ingest this knowledge graph and ask questions. What will be price of coffee I don't know, tomorrow if something will happen in Colombia. Mhm. Ethiopia. And you're giving example what what what can happen. And basically this intelligent system she should should get all information from knowledge graph and should should connect all all things, should apply all these ontologies and control vocabularies and give back to LM models which should give you precisely what should happen

Mhm. And predict how it's possible. Nice. That's nice. That's I think we will start with coffee but I think it can be used for any kind of prediction system. So this intelligent approach can be used basically for any kind of research. You can also do like deep research by trying by trying to connect all these dots together which are not obvious, you know, for human but somehow they can make sense because it can connect different control vocabularies and can find all these relationships between entities that we don't catch but but somehow they also connected and should be related. So we we can probably get all these kind of interesting discoveries by using this approach

I find it very interesting that Ag Tech is the one project where that became a home for this because coffee, they saw the coffee use case and apparently coffee use case is so economically important, right? There are billions of dollars in coffee and obviously uncertainties, right? Farmers basically live and die by weather and market change and whatever, right? So I find this very interesting that you know, they figured out that this technology is a good, you know, one for to predict stuff. I'm also very curious about Quadrant because Quadrant is recently emerging in the field of vector databases. Yeah. And very up-and-coming. So, what made you select Quadrant for this, you know, vector graph integration? Oh, it's it's very interesting story. So, I think it happened 1 year and roughly half a year ago. So, I was doing a talk for Harvard University and they recorded that thing and published. It was about first version of us dataverse

And suddenly, CEO of Quadrant he basically pinged me on LinkedIn and he said, "Look, there is something that you have you have to check. You don't know yet, but it's actually helpful." And I was a bit like, you know, like, well, I'm getting all these kind of requests quite often, you know, like, "Please check our stuff." But But suddenly, I started to read documentation and thought, "Well, probably it makes sense, you know, instead of like training model from scratch, probably we can use like vector storage where we can ingest something like the use vector database and we can also query with some some kind of threshold to get uh fastest search. And I tried that and I must say that I like that. So, I was thinking, "Okay, so instead of training process, we we need to replace with RAG." So, I started to move into RAG direction. I started also to to to build kind of distributed Quadrant things. Mhm. Different nodes. So, like very So, Quadrant is a database

So, you have to think how to manage it in in your infrastructure. So, I just put everything in memory to make it um extremely fast. Mhm. So, like you you can also there are few approaches how you can use that, but for long-term storage you can just install it on on your hard drive or in the cloud, but if you want really like high performance, you have to to create RAM disk in in your memory and connect it up with Quadrant on it. And this is where it's extremely fast. So, you use it query and you you And this is very important for agents interaction. So, when you know, so the human can can wait for a couple of seconds or even 10 seconds even for half of minute if if there is some some important question. But for agents, it should be like like immediate response, you know, because they are operating in different paradigm

So, I think Quadrant is also optimized for GPUs nowadays. And it's making Quadrant is is very efficient, very fast and excellent solution for for AI agents if you want to build kind of multi-agent system where all agents interacting with each other. And what is interesting about Quadrant, you can also use the IDs as a kind of identifier of the block, so you can put it in Quadrant. That's something also used for to build like engine search engine on CrossEngage. So, I can now now ask questions any questions on on the whole dataverse stuff. And uh I can go like like in in traditional way with graph like like query on on graph on specific concepts, specific entities, but I can go also go to to Quadrant and say something like like can you give me some looking something like that, you know? Mhm. And apparently because in in vector space, it will find something interesting and it will bring me back. So, this is also kind of uh the way how you can connect the AI agents and in a way you can also find something unknown

That you don't know what you're looking for, but you know it should be looking something like that. Sometimes it's Yes. >> like nearest neighbor >> Neighborhood, yes. That's right, yeah. >> great. So, that was probably And there is the irony the CEO of who contacted >> That's right, yeah. >> Yeah, and I mean this is great, right? Because it's like it shows that they're very good at developers. They're reaching out to folks, yeah

Quadrant is in our community as well. So, this is probably >> I'm with them. So, so I was in in bio hackathon in April of last year. Actually, I've I've sent I've sent a message to Andre, and I went to to their office to just to talk to the team. And I I think they got surprised with all these kind of use cases that I had at that time. It was something very specific for them, and Yeah, it's very interesting. >> Indeed, yeah. Now, this is great

Well, so I think we covered a huge amount of information. So, this is probably a good point to wrap up, and probably you know, we'll end by saying that first of all, thank you for presenting all that. What would be the entry point where our viewers and listeners can find this information? We'll collect a bunch of links from you, and post it right under this video when it's published and the blog post. So, the CDIF I know like there is a Zenodo paper for CDIF, and there is a bunch of GitHub stuff. For your projects, and and Bale Fire is now under Ax stack work on GitHub. Is there any single like, you know, Slava Tikhonov entry point, or should we now like a data I think has a bunch of stuff. And we'll collect them. What would be the good like entry point to look for this information, and learn about it? Because there is a a to learn

Yeah, this is a bit of problem because we are working on a few things at the same time. So, it's quite difficult to get this information one place. However, I must admit that I recently created bookmark service. So, I'm I'm going also to to bookmark any pages with interesting materials and every page will get the ID will be analyzed by AI will produce croissant and it will become kind of entry point for any kind of research. So, >> Okay. it's coming soon hopefully and there are a lot of people already interested they are basically asking the same question. Can you give us like like knowledge center where we can not only like like read but but we can ask questions. Mhm

So, I'm going to combine all these technologies together and basically to bring kind of intelligence service where you can ask questions about all stuff and you can also get reasonable answers from AI. Mhm. Of course, if you don't know or you are not able to find it, you can always like contact me on LinkedIn and ask those questions and I'm I'm usually posting all these recent talks and recent developments on LinkedIn. So, feel free to to follow me. Sounds great. So, we'll publish you know all the links which we know Mhm. in this blog post and and definitely you know folks can contact Slava for more information and collaboration. Well, thank you Thank you

Thank you. Thank you. >> I appreciate it and looking forward to you know help build a system around all this wonderful projects. Thank you. Thanks a lot Alexy. Thank you. Okay, so