AI analytics, with receipts
AI Analytics Week, Metabase
Transcript of the session recording: https://www.metabase.com/events/ai-analytics-with-receipts

So welcome everyone. I'm Pierina. I'm a product advocate here at Metabase, and we are
with, uh, our first, uh, webinar from the AI Analytics Week. Uh, it is the AI Analytics
with Receipts. So welcome. Um, we're happy to have you. First of all, uh, use... please
use the chat, use the reactions. Uh, we are really happy to just make this dynamic. So,
uh, feel free to put any comment, any, any question in the chat. We also have a Q&A
session, uh, so please, if you have questions, put it there. Uh, we will have some polls
throughout the, the presentation. Please help us, uh, filling them out. It will make
this session a little bit more dynamic. It will make sure that we are connected and I'm
not, uh, talking alone here.

And then, uh, at the end, we will have, uh, some space for Q&A too. So if you wanna, uh,
have some questions, we will have some space for that too. And with that, uh, let me
just get us started. So one of the things that I have here, this is actually our old
webpage, and the reason why I'm bringing our old webpage is because it gives you an idea
of what are the things that are changing. We, as Metabase, had this big tagline that was
"Don't be a bottleneck," and the idea was that, uh, you have a bottleneck at some place
that allows or that actually prevents you from having, uh, the creation of, uh,
analytics and data and information. So we wanted to be the tool that help you don't be a
bottleneck, help you unblock that piece.

Um, but now the block- the bottleneck is somewhere else. We have, uh, and we do used to
have an issue where, like, not everyone was able to understand SQL, not everyone was
able to create their own things, and the, the issue was at the top of that funnel. At
the beginning of that process. The creation was hard. Now, the bottleneck is somewhere
else. AI is helping you create easily, and it's making this process, uh, very accessible
for everyone. So the bottleneck is down the line, and the bottleneck for us, and we
believe that the bottleneck is at the verification process. And that's kind of like the,
the... what is setting the tone for what we are trying to build here at Metabase.

With that, the big tagline for this is that the cost of creation is going down. So it's
no longer something hard to do. Creation is something pretty straightforward, something
that everyone can do and that AI can help you do in a very easy and fast way. Um, so
yeah, with that, let's go into our first poll. We can bring the first poll, and you can
see that the poll is active. You are seeing this. We're just gonna give you a couple of,
uh, minutes to just fill that. Um, if you're not seeing it, uh, for, for the ones that
are not seeing this, the question is, what, if anything, had stopped you from trying
Metabase AI so far? So, um, most of you are saying that you are using it, which is
amazing.

Uh, some people, uh, are not using it. Um, some people weren't, uh, sure that that was
available, and some people hadn't had time. But, uh, overall, some of you are using it,
which is cool, and some people didn't know that was available. So that's, that's also
great because we are here to just try to understand, uh, that. Um, so yeah, some votes
still trickling in. Uh, but yeah, I think that we are... We have this split audience
where people are using it, and then some people didn't know that was available, so which
is great because we are gonna be able to cover that. Uh, so yeah, thank you for, for,
for answering our poll. Um, so the next thing, and if you have, um, joined our, our
today, you probably seen our quiz, which was about, um, knowing if your team was ready
for AI analytics.

And we have some results around that, that I wanted to share because I think it's
interesting, uh, what we have. So we just created this quiz, and we put it out there,
and some of you guys, uh, fill it out. Uh, so yeah, what is kind of like where, where
are you? Like, are you building? Are you curious? Are you scaling? And a lot of you are
building already with AI. You use it regularly. And y- the thing that concerns you the
most is security, who can see and who can do what. And you really want to control this
and self-host this on your own servers, which is something that is very important. And
also, you wanna connect things via an MCP. Some people also wanna use, um, the CLI, and
most of you are excited.

So just to kind of like share the things that, uh, people are telling us about, uh,
this. So in a sum up, you are using AI regularly or piloting. Security and governance is
the biggest blocker for broader adoption. Most of you wanna self-host these analytics
and these AI, uh, tools. And the MCP is the most preferred way to work with data
analytics, followed by, um, analytics as code. So this is just to understand a little
bit to put us in the, in the context of what we are gonna see today It is more than
obvious that creation is changing, and what we wanna kind of like establish today is
that we are part of that too. And how we are part of that too, this is what we are gonna
cover today.

We're gonna see MetaBot. We have MetaBot in product and we have MetaBot in Slack. We
have the MCP server and we have the CLI. Today we're gonna cover both MetaBot and the
MCP server, which is great because some of you or most of you wanted to go over the MCP
server. For the CLI, we have a great webinar, uh, that we did maybe a month ago or
something like that, uh, with a partner. I will definitely suggest that you go to that.
The CLI is such a big topic that we wouldn't have enough time for everything. So, um, I
would definitely suggest that you, that you go and check it out. And with that, uh, I'm
just gonna go and go over, uh, what we have, uh, in, in Metabase and we have in MetaBot
So you're probably seeing my Metabase, uh, screen right now.

This is, uh, my Metabase instance. This is where we do things in Metabase. Uh, for those
of you that haven't used Metabase before, Metabase will connect directly to your
database. You will have this browser-like interface, and you will be able to see
something very similar to this. Uh, and what you will have once you connect to your AI
of preference, and that is important because you can bring your own token, you can bring
your own model, you can use Metabase service, uh, you will have this chat with MetaBot.
And what I'm gonna do here is I'm gonna... I have a question. I actually just... I'm
gonna copy-paste it just to make it a little bit faster. But basically, uh, I'm just
saying, "I wanna know how much revenue we did last month," which is a very common
question, and the main product categories that drive the revenue, as well as the
location of the biggest purchases.

So I'm asking a lot. I'm asking, uh, how much revenue we did, what are the main product
categories that I, that I, um, I'm selling, and what are the location of the biggest
purchases. So I'm asking three different things, and I wanna know, uh, how MetaBot can
help me do this. What is happening underneath? Metabase is gonna go and find, with AI,
of course, we're gonna try to find, uh, things that are related to this. And it's
actually identifying that I am asking three queries, three questions in parallel, and
it's gonna go and try and create all queries. First of all, it's gonna show me what are
the things it's using. So for example, it's using an orders table because I wanna know
how much I made as revenue.

Then it's using our products because I wanna know which products. And then it's using
the people table to just bring the location of this pers- of these people. And here you
have the answer. It says like, "Here you have your full revenue picture of last month."
And it's actually showing me three different, uh, questions. So I have total revenue of
last month, which makes sense. Like it's telling me, "Hey, this is what you made last
month." Um, and, and you can see here that it has a, a filter. Revenue by product
category, so it's giving me how much revenue I had by product category. I know that
widget is the most, uh, is the, the most successful one.

And then I have the biggest purchase by location, and this is actually what it showed me
as the first option. Uh, and, and it's telling me that Texas is the, the most
interesting one at this point. Um- It's telling me that location is based by the billing
state because that's what they had, and that is kind of like the, the, the result I'm
having. I think this is pretty, pretty decent. It's telling me all the things I have. Of
course, would have been nice if I had a way for, like, having these three dimensions
into one specific question, but, um, MetaBot decided that having three different
questions will be the best way to play it. Um, so yeah, that's how MetaBot will work.

Couple of things that are important around MetaBot. It will respect the permissions that
you have. It will respect, uh, what you've created as, uh, as the groups that you have
in Metabase. So the data that MetaBot will see is exactly the same data that you will
see in any other way of creating things in Metabase. You have the option to put limits.
You have the option to say who has access to what, uh, and that is something that is
very important. And as I mentioned, you can bring your own model With the same idea, we
can go and I have this demo space, um, that I also use. It's a Slack demo space. I'm
just gonna move this thing. Uh, but basically, what I can do here is I can tag MetaBot,
MetaBot, and I just paste the same question.

So it's gonna, uh, create, and it's gonna start thinking about the same thing. So why is
this relevant or important? Um, maybe these are the kind of things, and these are the
kind of discussions that you're having with your team inside the Slack. So you are not
having these kind of things, these kind of conversations, uh, is- isolated. You are,
uh... you wanna get these numbers, uh, somewhere else. So this is actually what you can
get inside of a Slack. Uh, so this is a bit, a little bit different. It did a search,
and then it tried to build the questions in the same way. Uh, I think that this is
interesting because it will go and said, like, "Okay, I wanna know how much is the total
revenue per month," and it show me the revenue per month, which is interesting.

It gave me a different answer. It then show me the revenue breakdown by category, so it
show me, like, all the categories, and then it show the product by state and order. So
it didn't get me the same answer. It didn't get me exactly the same answer, but it went
and tried a different angle, which is something that is very common with AI. But you can
see that you can arrive to the same type of result with this. Uh, I will probably would
wanna kind of, like, tweak this a little bit because in this case, the revenue by
category, it seems like it's showing me everything, so this is not that accurate because
it doesn't apply the filter. So I would probably wanna go and say, "Okay, now show me,
uh, the previous month," and this will probably give me a little bit more of what I
want, you know?

Uh, it's not the greatest graph because it's a pie chart. I'd probably wanna use a, a
bar chart for this, but you can see that it gives you a starting point. For the other
ones, I think it did a, a pretty decent job because it did everything else. But as you
know, like, you need to kind of, like, check and see, uh, what will work. But this is
how it will work inside of Slack. Let's go and see how it would look like in the MCP
server. So you saw this, you saw what I created here. I'm just gonna go and open
another, uh, instance of the same, uh, Metabase, and I'm gonna show you what we have
here. So in the admin, you have the option to go to the AI tab, and you have the MCP
server.

Uh, here you can enable the MCP server. This is available for all plans and editions,
and you can enable the ones that you are gonna be using. So you can use Claude, Cursor,
you can use ChatGPT. Whatever you use in your team, you just need to enable it here. And
then what you are able to do, and this is what I'm gonna show you, is how that will look
like into, um, into Claude. In this case, I'm using Claude, and I'm just... You saw I'm
just using exactly the same question. Uh, of course, this is gonna, uh, try to get this,
um, this information. Actually, I made a small mistake here. I didn't tell it to use
Metabase. I usually try to do that. I tell them, "Hey, this is a question.

Use your Metabase MCP server and use this database." But in this case, it was smart
enough to know that I was using, uh, Metabase or that I wanted to use Metabase. And now
what it's doing is, okay, you have this e-commerce dataset. Let me load the queries, let
me pull the schemas, and let me just try to find the right columns to do this. Of
course, this is going to take a second to just load so we can see what is happening. I
had, uh, another one here, so we can maybe see that if this takes a little bit longer
But, um, basically what it's doing, it's just, you know, you can see that it's trying to
bring all the data and, like, getting, getting you this, those reports.

We're gonna give it a second. And while this happens, I wanna also show you that, uh,
you have the option to do controls and access to the different tools. So if you don't
want everyone to have access to everything, you can control that with the permissions.
Uh, let's see. Okay. So it actually gave me all this, and what I can do is ask it, it...
So this is an important thing. What it did is it bring everything from Metabase, and it
brought, like, all the data that it had, and it's showing me that inside Claude, which
could be something that works for me because I just wanna get the answer, and that's
good enough for me. Like what I did in Slack, I get the answer, and that's good enough
for me.

But maybe what you want is to see this inside Metabase. Give me the Metabase charts. And
I can actually ask that as a second thing. And what is this going to do is it's gonna
take all the results. It already had, uh, the results, so it's gonna construct the
queries and give me the charts. It might give me some, some authentication. Let's see
It, it gave some, some error there. Let's go and see the, the previous one that we had.
Okay, let's reload this. Maybe it's gonna allow us to see it now. Okay, perfect. There
it is. So you can see here, um, it built exactly the same charts. I can actually go and
explore these charts inside, uh, Metabase. You can see it's exactly the same thing that
MetaBot built for us.

In this case, it actually built a chart, uh, with the, uh, with the different states. So
you can see it's very similar to what we were seeing at the MetaBot level. Um, so yeah,
you can see that these are the same ways of creating exactly the same thing. You can go
into MetaBot and have those reports in the same way. You can go into Slack and create
those kind of things, uh, in the same way or you can use your MCP server, the MCP server
of your choice, and just basically use them. Um, these are very different ways of
building and getting the same thing. I wanna touch a little bit on the CLI. So the CLI,
uh, works, uh, great if you have or if you wanna do a lot of things, uh, at the same
time, if you wanna use agents.

So we can definitely, uh, use the CLI. We can definitely, um, rely on the APIs that the
CLI has. Uh, as I mentioned, there's a great, uh, webinar that you can check out, um,
and we can definitely, uh, understand a little bit more. But as I mentioned today, maybe
it's a little bit of an overview for the time that we have. So yeah, with that, uh, let
me just go back into my presentation and my slideshow. Uh, as I mentioned, MetaBotting
product, MCP server, and CLI. With all that, you saw how easy it was for me to create
everything. So I just like put the same prompt in MetaBot, in Slack, in Claude, and it
was matter of like maybe, I don't know, like seven minutes because I wait for everything
to happen, and you see me doing it live, but it's very straightforward.

I just use my, uh, basic knowledge of what I wanted, and it created everything. So this
is very clear that, uh, creation is fast, but verification is more important than ever.
So if I don't know enough about the data, then I am not sure of what is happening. So
the cost of verification is actually the thing that it's going up, and it is basically
what we wanna talk about today. So with that, let's go into our second poll today. Uh,
so for you, for those of you who are following, it's what worries you the most about AI
generated answers? So wrong answers, wrong numbers, n- no control over what it can see.
Uh, it can tell how it got the, the answer. So I think that those are very valid issues.

Uh, not worried, curious how it... to prevent. So I think that most of you are around
wrong numbers and can't tell how it got the answer, which I think are very valid
concerns and important things. Let's, let's give it a couple more seconds to see how it
goes I think that most of you are worried about wrong numbers, which I think it's pretty
valid. I think that it's very common. And yeah, can't tell how it got the answer. I
think those are the main ones based on the, on the votes. Uh, so yeah, thank you so much
for, for your opinions there We, we have the same, the same concerns. We all have the
same concerns, I think, and that is why, um, we, we wanna have this verification using
Metabase.

Uh, in order to try to, to work and kind of, like, try to reduce that fear of wrong
numbers and, like, not knowing where the answer is actually coming from and what is it
using, uh, we have all these features. We have, uh, Data Studio, uh, which has
Transforms, which has table metadata, metrics, and glossary. We have a set of, uh,
tools, a set of features that, uh, ca- target verified content with library official
collections and system prompts, and we also have remote sync. And that's what we are
gonna cover now as the next, uh, piece of our, of our demo today. So, uh, we just go
back here into Metabase, and we are gonna start with, uh, Data Studio.

So the first thing that we have in Data Studio is Transforms, and I think it's a really
interesting way of you, uh, to kind of, like, control and have a little bit more control
of what's happening. Um, some of you might use, uh, some other tools to do
transformations and to get the data ready. But if you don't have that or if you wanna do
some, uh, light, uh, next step transformations and Transforms inside Metabase are a
great way to do that. You have the option to do SQL Transforms, to run them on a
schedule, and to use that as the starting point for everything that Metabase and MetaBot
and the AI will be creating inside Metabase. You can also create Python Transforms if
that's something that you want, and this is gonna be, like, the starting building block,
the beginning of what you wanna build inside Metabase.

The other thing that it's about curating and about making sure that, uh, everything is
correct is table metadata. Um, so inside Metabase, and this is something that you have
in all plans and editions of Metabase, you have, um, the option to provide as much, uh,
details on the data as you want. So you can go into a specific table, you can go into a
specific, uh, field and provide descriptions. You can provide semantic types. You can,
uh, provide more detail on what's happening, and all of this will nurture the AI and
will provide a lot more, uh, curated results for it to actually build things that make
sense. Um, and this is gonna be coming from someone that understands the data, not just
assumptions or, um Intuition from, from the AI and just like, um, like the statistics
of, like the results of the AI that could happen.

Um, so yeah, those are two things that you have. The other thing that you have, uh, is
metrics. So metrics actually live inside and outside of library, so we're gonna touch a
little bit more on, on the library. But metrics are these curated, uh, aggregations,
curated numbers that you can create, that you can say, "Okay, this is how we calculate
the revenue in this company. This is how we do it, and this is what should be referenced
as the main thing." Um, and you can have, uh, something like this. You can have a
revenue like this. You can have and see this, how this looks like by different
dimensions, which is really cool. You can have the definition, and you can see what,
what dependencies are coming from this.

And of course, you can see who created this and why is this important. So metrics are
another thing that are very important, and then that will help you with verification.
And then another one, which I think that helps with verification, but it's mainly for
providing more context, uh, for AI, is glossary. So one thing that is very common is
that for me, uh, ARR could be one specific thing, uh, in the success team, and maybe for
another team it means something a little bit different or it has a different angle that
needs to be taken care of. Or maybe I have an specific definition of a concept, and I
need that to be taken in consideration when I'm using AI.

So that kind of thing can be done inside of glossary, and you can create those terms
that will be referenced by AI. So that's something that is also very important The other
set of things that is, that is important for verification, so this part is... It is
about verification, but it's also about curation of the data, making sure that
everything that arrives into AI, it gets into the AI, it's actually good. Then you have
a second step that it's a little bit more about verifying things and making sure that
only the most important thing and only the correct thing, uh, gets fitted or gets, uh,
actually used. So for one side you have... In one side you have the library, and the
library is this place where, uh, you can say, "Okay, not everything is important.

Not everything that lives in my database should be accessible by everybody or AI." And
that is where you go and create a published table, where you create a metric, where you
create an snippet, and all of this will be predominant and will be the first thing that
appear and the first thing that should be referenced. Then you have two other things
that are important, uh, that actually live on the main application. One is official
collections, and this is important because you can configure AI to just look into a
specific collection. Uh, so you can create an official collection and say, "Okay, this
is a collection that is actually important, that is actually something that was either
generated by a user or generated by AI but it's actually verified.

I know this works. I know this is correct, so everything that lives there is gonna be
correct." And then you have verified items. So if this is something that was generated
by AI, then me as a human, I can come and say, "Okay, this is actually correct. This
invoice trend is correct." I can go and verify this question because I know that it's
now... it's gonna be helpful for someone else. So, you know, you have official
collections, you have verified items. It can be for verifying specific questions, a
specific dashboards model, so it really has, uh, that option for all the assets. And
then finally, another thing that helps on, on verification, and it actually helps on the
behavior of AI, is something that lives, uh, in the AI itself in the system prompt.

So system prompts can help you, uh, give instructions that are about the tone and the
way that AI responds of course, but also about the entities to prefer. If you have
specific models, specific collections, specific things that should be, uh, the things
that are, um, prefer when doing, when doing specific, uh, commands, then here's where
you will put those prompt instructions and where you will have this, uh, specific
instructions. Um, so yeah. And then- I, I, I, I thought a lot-- I talk a lot about
verification, but another, uh, part of this is making sure that there is a workflow
that, uh, people when they are creating, uh, they have a way of like a, a fail-back in a
way.

Uh, and what we suggest for this is to have a development instance, to have a space
where you can, uh, let everyone play around, let everyone work, and then have a
production instance where you push things that are specifically, uh, validated, that you
know are correct and are gonna be ready for everyone to use. So you have those, those
options, and the way that you do that is with, uh, remote sync in Metabase. So remote
sync allows you to have a development instance, have that connected to a repo. It could
a GitHub repo, a GitLab repo, you could even use Bitbucket, and then you can have that
into a production instance whenever that is ready. This might be an overkill for some
sp- small, uh, instances, so for that you could use maybe different, uh, collections and
just move things around between collections.

Uh, for bigger installations, uh, for internal use case, this might be a really nice way
of you to control things. Uh, but I think it's very, very important when you are facing
this for customer-facing, uh, analytics, when you're doing embedding or, or something
like that, 'cause you have a lot more control of what is happening and what you're
building for your end users. Um, so yeah, those are the main things that I wanted to
talk about on verification, and let's go back into this slide. I wanted to point out on,
on, on the why of this verification and why is this important, and I think that the main
thing is about consistent answers. Building everything for this pipeline to work, having
this semantic layer, making sure that you have verified items, making sure that, uh, you
have those system prompts in place will make sure that, uh, questions that are generated
by the API, by, uh, the AI all have consistent answers and all have this specific, uh,
line that they are following and you are not being faced with wrong numbers But then the
question is why in Metabase?

Like, you can have all this curation basically everywhere. You can make sure that all
these things happen in some other tools. That, that's, that's not, uh, something that,
uh, is, uh, that we are not aware of. So w- the question is why in Metabase? And I think
that the main thing will be the end-to-end visibility and the lineage. Metabase will
allow you to see everything that is happening in every step of the process. You have a
transform, and you make a change in the transform, you can see what is happening between
that transform, that specific published table, the specific dashboard that the CI
generated, the question that your team, uh, generated, and then all the verification
process.

You have all this lineage. You have the option to see what is happening throughout time,
uh, with Metabase, and I think that is the specific value that, that we find, uh, in
doing things in Metabase, and also doing the creation, uh, in Metabase. Um, I think we
have a third poll. We are, we're... We have a, a third poll there, uh, that is about the
curation. So let's see. I think the poll is already active. So the question is: Do you
already have a curated semantic layer or metrics layer? Uh, yes, something that is ad
hoc. Yes, something that is mature. Uh, not sure what that means. Um, so yeah, some
people... uh, most of you have something that is already ad hoc, so of course, this is
something that is, uh, uh, part of your workflow that you already, uh, have and that is
important for you Great.

I think that most of you are, are replying. You- most of you already know the value of
this and are already doing this Perfect. Perfect. Yeah, so I think that, yeah, you
have... Most of you have, uh, a semantic layer and a metrics layer that is a doc. Um,
some of you are not sure what it means. What it means is this, uh, specifically, uh,
verified, uh, layer where you define, define specific tables, define specific metrics,
define, uh, what each value means. A little bit like what I was showing with, uh, Data
Studio and what I was showing with the published table and the table metadata and the
glossary, all that, uh, will be around, uh, the semantic layer.

Uh, okay. So I think that is pretty much all we had, uh, as the specific, as this
specific part. So we wanna kind of like invite you to turn it on this week, invite you
to check it out, invite you to see how all these features work, and give us your
feedback, understand if this is something that will work for, for you. And with that,
uh, I also wanna invite you to our, uh, the rest of our webinars. We have one webinar
tomorrow, uh, for AI governance, uh, which is gonna be, uh, very, very, um- A complement
to what I've been talking about because, uh, all of this is really cool, but without
governance, then it really has some, uh, weak links. So we definitely wanna want you to
understand that that is there, and that we have all that information around.

Then, um, I also wanna invite you to the really cool webinar we will have about data
apps on, uh, Thursday, I think. I think it's at the same time. So remember that we have
two more webinars around this. We are just starting the AI week, so hopefully you can
join us, uh, tomorrow and the day after. And now, uh, let's go into some Q&A. I think we
have some, some questions in the chat, so let me s- stop sharing my screen, and now
you're gonna see me very big. But let's go and see the Q&A. I think that we have some
questions here. Uh, okay. Uh, perfect. Let's, let's get started. So can Metabot... Oh,
sorry, let me just add this to the stream. Can Metabot ground answers in native SQL
queries or only GUI build questions?

Most of my complexity are CTEs, window function, custom joins that live in native SQL.
Yes. So yes, Metabot can also write SQL. Uh, what I showed you today was pretty
straightforward. You saw the question I did. It was pretty, uh, pretty simple, which is
why it just relied on the GUI question, but it's gonna do, uh, more complex SQL. I would
be very honest, think that, um, when I've seen-- I haven't seen it build CTEs. Window
functions, yes. Um, in terms of custom joins, it doesn't do a lot of complex joins, so
what I would suggest is that, uh, you do it through the MCP server or the CLI because in
those, uh, two scenarios and with those two tools, it does CTEs, it does window
functions, it does the custom joins.

So, uh, maybe that, uh, that helps. Uh, uh, let us know if that, uh, answered your
question. Let me go into the next question. Let, let us know if, if, if you have any
other question there. Uh, okay. Let's see. What if Metabase already has the answer to
the question? Will it provide that first or will it default to building something new?
That's a great question, Ilse. Thank you so much. Yes, it will do a search first, and it
will show you that option. I, I gotta be very honest, it will do a semantic search. So
sometimes it will find it because you are using the specific, uh, words that will do
that, uh, that will show that, but sometimes it might not be great at finding that.

Um, you can ask it to find it, to find a question or to find something as specific also
in the MCP server. With the MCP server, you have a little bit more control because
MetaBot will do the search and then if it doesn't find the search, it will do, um, it
will do the, the creation. But with the MCP server you have more control because you can
just say, "Hey, search for this specific thing and tell me if it exists and show me
things that could work for that." So hopefully that answer your question. Uh, let me
just add the next one to the stream. When MetaBot gets it wrong and I correct it, does
that correction miss or does it learn for the next time, or is it, uh, every session is
stateless?

So when you correct it, you mean like, uh, maybe I need a little bit more, you can write
it in the, in the chat. When you mean you correct it is because like it did, uh,
something wrong and it didn't pick the, the correct data source. If you cor- if you do
that with the data source, it's gonna correct it for the next time. But maybe if it's
something else it might not because it doesn't remember all the, all the... It doesn't
have a memory to remember, uh, what is happening, uh, for different tab- different chats
Okay. I, if you, if you have, um, more that you wanna, that you wanna provide there,
just, just let me know and we'll put it there. Uh, okay.

Let's go into the next one. Is MetaBase a separate cost if our company already uses
MetaBase? We can use MetaBase in the site itself under chat, but not sure if the cost
increased with activating the Slack feature. So it is, um, it is a dif- it is a separate
cost if you are using, uh, our AI service. So how it works, you can bring your own
token, and you can bring whatever you use. If you use OpenAI, Anthropic, whatever, you
can bring it and put your own token. So that will consume your tokens, whatever you use.
Uh, if you enable it through MetaBot, uh, through MetaBase AI service, what we are doing
is we are just passing the charge of the cost, which is, uh, three and seventy-five,
three dollars and seventy-five cents for a million tokens.

So I don't know if that answered your questions, but it really depends on how you have
it configured. Um, if you are using your own tokens, then it's not something that will
increase your MetaBase bill. Uh, it's actually available in open source, so you wouldn't
have anything to pay to, to MetaBase there. It will be just your token Okay, let's go.
Um, next question from Ilse, does verification train the AI? Not at this moment. It
gives context for AI to do, do a better job, but it doesn't train the AI at this point.
That is a actually really good question, and I think that it might be a really good
feature request for us, which will make a lot of sense.

So I'll definitely take that one, Ilse. Thank you. Um- I think there is another one, uh,
from Aim. I'm sorry if I'm mispronouncing your name. Let me see. Uh, "Does MetaBot cache
answers or rerun every time? If 10 people ask similar question, is that 10 separate LLM
calls?" Yes. So MetaBa- it will rerun every time, like the way that MetaBase works in
general. We just, by default, we just go into the data- into the database every time,
and that's how MetaBot will work too. Um, so yes, it will be 10 separate LLM calls Mm.
Okay, we have another question from Igor. Is all data sent to LLM or does it only apply
to the specific structures, queries, and SQL code?

Uh, it only applies to a specific structures, queries, and SQL code, not all the data.
It's gonna be the metadata and the question that you did Mm, hopefully that answer your
question. And, uh, there is another one from Christian. Is this recorded so we can go
back, review and learn in more detail? Yes. Yes, uh, Christian, you will get a replay
after this, probably in half an hour after we finish. Um, so yeah, you will definitely
get it. Uh Cool. I don't know if there are any other questions. Hi, Ramiro Perfect.
Yeah. And thank you so much for A- Abid and, and, and Andrew that I saw that they
already replied some of these things, so awesome. Thank you so much.

Hopefully I didn't contradict any of those answers. They know better probably, so
anyway. Well, I think we're almost at time, so wanna thank everyone for joining. Thank
you so much, Ramiro. Thanks everyone. And yeah. Yeah, yeah. And hopefully we will see
you tomorrow, uh, with, with Abid. It's gonna be a great session, so you, you don't
wanna miss the, uh, the AI governance one. It's gonna be super cool. So yeah Yeah.
Perfect. Thank you, Ramiro. Take care. Bye, guys