Christophe Blefari, nao Labs — Interview with Alexy

I'm Alexi Krabro, head of ecosystems at I'm Alexi Krabro, head of ecosystems at Lakes Sale and here I am at PI data Lakes Sale and here I am at PI data Amsterdam and with me I have Kristoff Amsterdam and with me I have Kristoff Bifford who is the founder of Now Labs. Bifford who is the founder of Now Labs. Exactly. Exactly. So he just gave a keynote uh recapping So he just gave a keynote uh recapping the history of analytics and uh how we the history of analytics and uh how we use data in the warehouse and the use data in the warehouse and the lakehouse and it was super interesting. lakehouse and it was super interesting. So if you know for those folks who So if you know for those folks who didn't see the keynote, can you just didn't see the keynote, can you just briefly recap like what did you talk briefly recap like what did you talk about and why I think it's important for about and why I think it's important for the attendees of PI data. the attendees of PI data. Yeah. So yeah the keynote was a bit of Yeah. So yeah the keynote was a bit of uh seeing where do we come from. So what uh seeing where do we come from. So what was analytics uh 50 60 years ago because was analytics uh 50 60 years ago because it started like a long time ago when it started like a long time ago when people invented the dimension the facts people invented the dimension the facts the warehouse and so on. So it was a bit the warehouse and so on. So it was a bit being nostalgic of all the stuff that we being nostalgic of all the stuff that we left along the road, you know, all these left along the road, you know, all these legacy system and then trying to get a legacy system and then trying to get a state-ofthe-art of what is uh an antics state-ofthe-art of what is uh an antics today and one what people are doing with today and one what people are doing with this agentic system and where it's going this agentic system and where it's going tomorrow. I tried also like to do some tomorrow. I tried also like to do some kind of fun experiment, you know, like kind of fun experiment, you know, like showing in a live demo what what is the showing in a live demo what what is the stuff that you can do already today, stuff that you can do already today, right? And what's really struck me as right? And what's really struck me as super fun that you basically said, you super fun that you basically said, you know, talk to your data. So you had some know, talk to your data. So you had some duct DB uh transactions and you could duct DB uh transactions and you could talk to the data and then you had a talk to the data and then you had a recording of your talk going on. So recording of your talk going on. So there was a little thing in the bottom there was a little thing in the bottom right corner which I didn't realize right corner which I didn't realize first what it was and then I saw that first what it was and then I saw that was actually the transcript and then you was actually the transcript and then you had the sentences as rows and then you had the sentences as rows and then you can actually talk to to them. Uh right can actually talk to to them. Uh right so like talking to your data live and so like talking to your data live and the whole presentation was super the whole presentation was super interactive. I think in the end you kind interactive. I think in the end you kind of said that you have your own like of said that you have your own like react app showing right like running in react app showing right like running in the background. So one thing I'm super the background. So one thing I'm super curious about so I come from um natural curious about so I come from um natural language processing and agentic AI and language processing and agentic AI and so when we talk about semantics we talk so when we talk about semantics we talk about the meaning of life like what is about the meaning of life like what is the meaning of this right so so what the meaning of this right so so what struck me when I talked to and started struck me when I talked to and started basically working back in the lakehouse basically working back in the lakehouse I did the very first spark meet up in I did the very first spark meet up in the world 15 years ago in Berkeley with the world 15 years ago in Berkeley with matear and so I came back now to rewrite matear and so I came back now to rewrite the whole thing in rust and so I'm not the whole thing in rust and so I'm not you know an analytics guy I'm a software you know an analytics guy I'm a software engineer so what struck me that uh uh engineer so what struck me that uh uh lakehouse people use semantic in a very lakehouse people use semantic in a very pedestrian way. It's like it's a SQL pedestrian way. It's like it's a SQL query, it's a metric. Yeah. query, it's a metric. Yeah. Right. And so what when you mentioned Right. And so what when you mentioned semantic layer I think you had semantic layer I think you had interesting observation that people have interesting observation that people have semantic layer but it's not it's semantic layer but it's not it's questionable whether it's useful or not questionable whether it's useful or not because and it was it was big news for because and it was it was big news for me that like you know I used to think of me that like you know I used to think of semantics is the ultimate like what is semantics is the ultimate like what is the meaning of this but in the lakehouse the meaning of this but in the lakehouse world it's a very simple thing it's a world it's a very simple thing it's a bunch of pre-arranged SQL queries right bunch of pre-arranged SQL queries right and so for instance bunduck people and so for instance bunduck people recently did a webinar where they say recently did a webinar where they say like we can put semantics in lighthouse like we can put semantics in lighthouse which means they put a markdown document which means they put a markdown document yeah with a standing SQL query And then yeah with a standing SQL query And then they have a little description. Hey, they have a little description. Hey, this is how many red coffee makers I this is how many red coffee makers I sold in New Jersey yesterday, right? So, sold in New Jersey yesterday, right? So, what do you think? Like, wouldn't you what do you think? Like, wouldn't you agree that like we should strive for agree that like we should strive for more for semantics? Like, because you more for semantics? Like, because you said it's even not doing what we want, said it's even not doing what we want, right? Like it's it's fixed like is it right? Like it's it's fixed like is it is it like I wonder how did this happen is it like I wonder how did this happen that people call semantics just a that people call semantics just a collection of queries? collection of queries? Yeah. So I I think the the issue and the Yeah. So I I think the the issue and the misunderstanding about all the these misunderstanding about all the these words is that they are words is that they are let's go. let's go. Sorry. Um the the thing is u there is an Sorry. Um the the thing is u there is an issue of just understanding because we issue of just understanding because we have a lot of people saying semantic have a lot of people saying semantic layer or semantics layer or semantics but it means a lot of different stuff but it means a lot of different stuff you know. you know. That's right. Um and the way I frame it That's right. Um and the way I frame it and I I think it's I I think it's pretty and I I think it's I I think it's pretty uh right uh it's we say semantic layer uh right uh it's we say semantic layer and when we say semantic layer to me and when we say semantic layer to me it's it's the way we encode the business the the the way we encode the business the the you know the meaning of life inside some you know the meaning of life inside some files somewhere that explain what this files somewhere that explain what this metric mean what this dimension mean metric mean what this dimension mean what this measure mean and and so on. So what this measure mean and and so on. So that is to me what is the semantic layer that is to me what is the semantic layer and there is the other part which is and there is the other part which is like the metric store which is like the metric store which is how do you query a semantic how do you query a semantic right right so the metric store is like QGS metric so the metric store is like QGS metric flow or luke flow or luke or lucer in a sense more than lukel or lucer in a sense more than lukel because look is the way you you you you because look is the way you you you you define the semantics define the semantics and the thing is the AI is so good today and the thing is the AI is so good today that if you give to AI a definition of that if you give to AI a definition of semantics warehouse is a YAML warehouse semantics warehouse is a YAML warehouse it's semantics encoded in a markdown it's semantics encoded in a markdown warehouse it's just a txt warehouse it's warehouse it's just a txt warehouse it's whatever I don't care it's just text whatever I don't care it's just text that says revenue is defined this way uh that says revenue is defined this way uh country is defined this way my dimension country is defined this way my dimension are this and this the AI will be able to are this and this the AI will be able to do whatever you want to be honest the do whatever you want to be honest the the thing that you have to give to the the thing that you have to give to the AI is semantics that have unique AI is semantics that have unique definitions that are really not definitions that are really not ambiguous because this the ambiguity is ambiguous because this the ambiguity is that what kills the AI capabilities in a that what kills the AI capabilities in a sense. sense. Um, and so what you say about like the Um, and so what you say about like the semantics being some pedestrian way just semantics being some pedestrian way just having like one concept defined with having like one concept defined with SQL, right? SQL, right? Can be a solution. I I think it's not Can be a solution. I I think it's not the best one because the best one because if you give a lot of SQL queries to if you give a lot of SQL queries to explain semantics, you're going to lo explain semantics, you're going to lo because you for sure you're going to because you for sure you're going to have have different join that will contradict different join that will contradict themselves uh in the end. So I think themselves uh in the end. So I think having rather uh for each metric each having rather uh for each metric each measure is definition a human measure is definition a human description of it explaining what are description of it explaining what are the parameters all this goes inside a a the parameters all this goes inside a a bigger view which can be the entity bigger view which can be the entity depend on the semantics vocabulary can depend on the semantics vocabulary can be entity tables topic whatever and so be entity tables topic whatever and so on I think that's what it should be on I think that's what it should be right right um and there is another thing that is a um and there is another thing that is a bit difficult in the current ecosystem bit difficult in the current ecosystem is there are a lot of uh AI agent that is there are a lot of uh AI agent that works without semantic layer works without semantic layer right right and that's what I do with now as well uh and that's what I do with now as well uh we support semantic layer but if you we support semantic layer but if you don't have a semantic layer it works as don't have a semantic layer it works as well because we think that well because we think that the business knowledge that people have the business knowledge that people have have to be unc have to be uncoded have to be unc have to be uncoded somewhere somewhere right right you can encode it in your data model you can encode it in your data model you can encode it in the silver layer you can encode it in the silver layer you can encode it in the gold layer you you can encode it in the gold layer you can encode it in the marked you can can encode it in the marked you can encode it in the semantic layer if you encode it in the semantic layer if you Uh so it's something super flexible you Uh so it's something super flexible you know as long as you uncut it you uncut know as long as you uncut it you uncut this business knowledge somewhere the AI this business knowledge somewhere the AI going to find the path to find what is going to find the path to find what is the right context to answer the right the right context to answer the right the work the work you should have it somewhere and agree you should have it somewhere and agree with the agent and other people here is with the agent and other people here is where my business logic where my business logic exactly and we've seen also a limitation exactly and we've seen also a limitation the issue with the semantic layer is the issue with the semantic layer is that the semantic layer is something that the semantic layer is something that data people are are doing where that data people are are doing where they say this is the 50 metrics of my they say this is the 50 metrics of my company. company. This is the 10 dimension of my company. This is the 10 dimension of my company. And if you give to the semantic layer, And if you give to the semantic layer, you you're going to create some kind of you you're going to create some kind of frustration because there is no way that frustration because there is no way that people are just asking for this metrics. people are just asking for this metrics. Stakeholders, they're going to ask a lot Stakeholders, they're going to ask a lot of stuff and all the stuff they're going of stuff and all the stuff they're going to ask is not necessarily only within to ask is not necessarily only within the semantic layers. the semantic layers. So they're going to ask something and So they're going to ask something and the will say I don't have the answer to the will say I don't have the answer to this question or the willinate something this question or the willinate something that is a bit wrong. Right? So I think that is a bit wrong. Right? So I think what you have to do is like a two layer what you have to do is like a two layer uh thing where you first eventually have uh thing where you first eventually have the semantic layer that answer fast uh the semantic layer that answer fast uh answer something that is grounded driven answer something that is grounded driven and super strict and if there is no and super strict and if there is no answer in this or if the AI decide that answer in this or if the AI decide that the semantic layer is not the right the semantic layer is not the right place to ask the question they should be place to ask the question they should be able to query a broader spectre of able to query a broader spectre of context context to find the thing that is necessary to to find the thing that is necessary to answer the question and and write the answer the question and and write the SQL query, answer the question and so SQL query, answer the question and so on. on. Right. No, that's great. And so uh the Right. No, that's great. And so uh the last question is you mentioned now that last question is you mentioned now that is your labs and there is a project on is your labs and there is a project on GitHub with 1600 stars and counting. GitHub with 1600 stars and counting. Yeah. Yeah. Uh what is now and why should folks use Uh what is now and why should folks use it? it? Yeah. So uh I'm building so now uh I'm Yeah. So uh I'm building so now uh I'm the co-founder of the company called Now the co-founder of the company called Now Labs. Now is an open-source analytics Labs. Now is an open-source analytics agent. What we provide is um a tool that agent. What we provide is um a tool that as a data person you can deploy to your as a data person you can deploy to your company. So everyone at the company can company. So everyone at the company can chat with the data that you have in the chat with the data that you have in the warehouse in the lakehouse everywhere uh warehouse in the lakehouse everywhere uh in the company and the thing that we in the company and the thing that we think is data people uh are key to the think is data people uh are key to the transformation with the chat with the transformation with the chat with the data. data. Yes. Yes. Um because they are the people knowing Um because they are the people knowing where is the data. They are the people where is the data. They are the people that will document the context. So they that will document the context. So they are the context engineer like I have on are the context engineer like I have on my on my shirt. my on my shirt. Yes. Yes. Yeah. Context engineer. Yeah. Context engineer. Context engineer. Yes. Context engineer. Yes. This is like the the new job. you know This is like the the new job. you know the analysis engineer is a bit rebranded the analysis engineer is a bit rebranded as the context engineer. as the context engineer. So key people on this uh on this So key people on this uh on this transformation are data people. So we transformation are data people. So we give this tool to data people. So they give this tool to data people. So they deploy it. So everyone at the company deploy it. So everyone at the company can chat with the data and we give the can chat with the data and we give the tools to the data people to administrate tools to the data people to administrate this with uh a feedback loop where you this with uh a feedback loop where you get recommendation to improve the get recommendation to improve the context later on to improve the data context later on to improve the data modeling. Uh we give the observability modeling. Uh we give the observability on everything. We gave this as code. So on everything. We gave this as code. So you can manage the context as code and you can manage the context as code and it's fully open source. it's fully open source. Uh so you can use it without paying Uh so you can use it without paying because it's a product or you can pay because it's a product or you can pay like the enterprise version like the enterprise version and the cool thing is you can use the and the cool thing is you can use the model that you want. You we don't sell model that you want. You we don't sell any token. So you can use every endpoint any token. So you can use every endpoint of bring your own model. of bring your own model. Yeah. Bring your own model and also Yeah. Bring your own model and also bring your own database. So we support bring your own database. So we support all the standard database at the moment. all the standard database at the moment. But But bring your own model. Bring your own bring your own model. Bring your own database. Bring your own beer. database. Bring your own beer. Exactly. bring everything that you want Exactly. bring everything that you want and we have like a slack community. We and we have like a slack community. We organize a lot of events. Uh and yeah, organize a lot of events. Uh and yeah, thank you very much. I'm excited about thank you very much. I'm excited about it because as a software engineer it because as a software engineer building houses for the people I really building houses for the people I really want to understand what my users like want to understand what my users like and if analysts and context engineers and if analysts and context engineers are the users then we want to have the are the users then we want to have the best practices. So you basically encode best practices. So you basically encode best practice of analysts in now. Yeah. best practice of analysts in now. Yeah. And now we can use it to do what the And now we can use it to do what the best folks like you do. So this is best folks like you do. So this is awesome. Thank you very much. awesome. Thank you very much. Love it. I'm going to check out now.