AI analytics, on your terms, on your infrastructure
AI Analytics Week, Metabase
Transcript of the session recording: https://www.metabase.com/events/ai-analytics-on-your-terms-on-your-infrastructure

Welcome to our tour of, uh... welcome to day two of AI Analytics Week. I'm Abed, and,
uh, we are going to talk about analytics and AI in your infrastructure. Before we get
started, let us know from where you are joining. I'm joining from Budapest. Um, it's
Budapest, Hungary. The weather is really great here. This is the second of the three
sessions. Yesterday was the first one with Perina. If you didn't attend it, we will
upload it on YouTube and, uh, you can watch it. And tomorrow we will have another
session with Matt. It will be more cooler than mine, uh, so make sure you attend it.
Today is really less flashy side of AI. We will be talking about permissions, cost
controls, audit logs, model choice, and where everything runs.

And honestly, uh, that's usually the stuff that decide whether an AI project gets
approved or stuck, uh, as a demo. Um, a couple of, uh, quick things before we get
started. As always, we will have a few polls throughout this webinar, so please take a
moment to answer them. And if you have any questions, please feel to, to drop them in
the chat at any point. We will leave some time for the Q&A at the end of the w-- uh, uh,
end the session. Okay. So the main point I want to, um... Let me change picture in
picture. Okay. It's okay. So the main point I want to start is pretty simple: AI didn't
create the governance gap. Companies have always had to figure out who can access their
analytics, what data different teams should be able to see, how people should sign in,
what happens when someone leaves a company or moves, uh, to another team.

What happened is AI changes the speed and the scale of access. Someone can now ask a
question in a plain English and directly, boop, you get the answer in, in, in seconds,
uh, without necessarily knowing where the data coming from or what's happening
underneath. Like you have a full trust that your admin that set up, uh, the platform for
you, so you can make sure you can answer a question and you get, uh, uh, uh, a better
results. So these existing controls still are matters, and they are now even matter more
with the AI and connecting AI with the data. So when you look at this control, they are
like authentications, group permissions, table permissions, row-level roles, database
roles, spending limits, logs to see where the application is running.

They are like technically boring stuff here, but they're really important, and the
companies are already dealing with them for years. And this is less visible side of the
AI, but it's very, uh... it make us like, uh, it make a real rollout possible. So once
you move beyond the demo, these are controls that really matter. Um, so this is today
we'll be talking about, uh, the conversation around the AI is not about like, hey, which
model can do or what, but it's also about like to bring, uh, your company how it already
works. Um, so, uh, I want to say it like in practice, for us, it comes down to five
pretty simple questions. The first one is, who ac- actually get access?

Do, uh, do you give it to everyone, to or just a few teams or maybe certain roles?
Because the moment AI can interact with company data, access become a governance
decision, not just like a feature toggle. Second is what, uh, you can actually see. It
should be, uh, should it be, uh, reachable everything in your analytical environment or
like certain data that a person already have access to? Third question would be what
happens when usage starts to grow? AI can be cheap when only few people are using it,
but once adoption spread across teams, the cost can scale quickly too. And, uh, so you
need to think about like how much usage you're comfortable with before it become a
surprise for you.

Fourth one is how do you make sure that the AI understand your business properly? This
is like the main thing actually here. Like, every company has its own language, its own
definitions, the way they talks about their a- uh, data. So if the AI doesn't understand
those things, it can give you an answer that sounds right but means completely different
from what your team intended or what you intended to talk to it. And this is the big
role of actually connecting a BI layer between your, uh, LLM and between your database.
And the last question is once people start using AI, how do you know what's actually
happening? Who's using it? What they are asking it?

What query is it generating? How much is being used? Because AI becomes part of how
people works with your data, not-- you need a visibility into that activity as well. So
we're going to put a poll for you in the chat. Um, please answer it. Is which AI
governance, uh, concern matters most for your organization? Uh, let's take a moment so
you can all answer it One is active Pronto, está tudo Can you see the poll? I'm just, um
Uh, probably it's lagging, but it will comes. Just to double-check here, I have this,
like, uh, trust issues. Can you still hear me, everyone? Okay. So yes, you can hear me.
Perfect. Uh, the poll will come. Uh, so the software is just, uh, um, stuck.

Uh, I will continue, and then like you, you'll just... It appear to you. Okay. So the
nice thing is none of these question requires a completely new governance model.
Metabase already gives you the way to deal with each of them. So let me show you what
that looks in practice. Okay, I think the poll is ready and you're ready. Everyone is
answering it Okay, I can see it's coming all with the what it can see. This is the most,
uh, every time, even in my career when I'm bor- uh, talking with prospects and
customers, this is what they care. Like, "Hey, can AI see everything?" Or how it see the
authentication and everything. So let me share, uh, Metabase. Uh, right?

You can see my Metabase now Uh, let's make it bigger for you. It's good. Okay. So if I
go to Admin, and then I-- if I go to AI, the first question we said is, um, who actually
get access to the AI? So this is like the control access by group. So if I go to AI
settings, and then under Usage Controls, we have a section for access. And this access
is according to the groups of Metabase. If you're not familiar with permission in
Metabase, we assign the permissions to groups. So you invite someone to Metabase, they
sign in, and then you assign a group for them. I will show you this in a bit, but this
is like are the groups that we already have created them.

For example, we have a group for data analyst, we have a group for marketing team, we
have a group for operations team, and we can turn on the AI features. What are, what are
the AI features in Metabase? They are Meta Bot inside Metabase. Um, you can talk to him
inside it, uh, Metabase. The MCP server that you can connect Metabase to your MCP, uh,
using Claude or using, uh, Cursor or ChatGPT and, um, uh, other stuff that we will be
talking about during the, uh, um, the webinar. So these are the AI features we-- if we
turn them on for data analyst, and they can go granular like, hey, I want them to have
access for the chat and the, uh, natural language that they can talk to, um, Meta Bot.

I want the AI to give a SQL generation for them. And also I want to give them other
tools, which they are like fixing error in a SQL or analyzing a chart, uh, or having
inline, uh, inside the SQL as well. When you are writing in a SQL, you can have inline,
uh, chat with the, with the AI. So these are-- this is how Metabase started like this.
Hey, we need to give access. Who can have give access? Some groups we don't need to give
them access for this. Now, the second question was, what can the AI actually see? And
this is everyone is actually, uh, concerned about it. AI in Metabase works with the
access of the person using it. So if someone cannot see a table, collection or a certain
row of data, the AI should, uh, will not be seeing this st-stuff.

The AI will be seeing with the person who is signing in. So in Metabase, you don't need
to set up a separated permission model just for AI. It works within the same control you
already have. So in Metabase, as I said, you invite people to, uh, to Metabase, they
sign in, and then you assign them into groups, and then you go to the permission. And
here are two part of the permissions for these groups. You can set permissions on the
data itself From the databases or from the collections, which are the folders that you
create your dashboards and your reports, and you put them there. So in the groups, for
example, let's say embedded users, uh, these are the databases that we are having them.

We can give them the view data. Uh, you can block them. They can view or even you can
give them granular. Uh, for the creating queries, do you want to give them creating
queries, uh, with the query builder or even with the native query builder, which is the
SQL. So you have all this granularity. Same thing with download results, manage ta-table
metadata, manage databases. So all these, you assign them into, uh, the groups for the
people themselves. And even if you want to go more in details, if I go with the sample
database, then I can see each table, and then I can assign also viewing data, creating
queries, download results for each table. And here you assign for the row column
security.

So what does it mean? If I, I have access, if I am in the embedded user here, then I
will have these permission for myself. Like, so you don't need to be afraid that like,
oh, someone sign in, like how I can need to give him access. The permission works
according to what they have, they can see. And the same thing for the collections. The
collection, you go for each collection, and then you select which group can have access
either by curate or view or having no access at all Okay. Now, um, the third question we
said, what happens when the AI usage start to grow? And I think now this is something in
mid two thousand twenty-six everybody was talking about like, hey, uh, uh, paying for
Anthropic or OpenAI for the A-API usage, like it's going crazy.

So we in Metabase, we have created an AI. If you go to users analytics, we created a
limitation for your team on how they are consuming these tokens, so they don't burn all
the tokens. You can go two ways, either you go by token usage or by message count. Like,
hey, you can have only ten messages or whatever. Uh, personally, I set it up for, for,
for, uh, our team is by token usage, and you can set like a weekly, uh, uh, uh, cadence
for the, uh, limits. Can be a daily one or a weekly one or a monthly one, and even you
can set up a limit for it, like two million tokens per week, for example. This is what
they can get. Uh, we have a nice thing that you can add, um, customizable message that,
"Hey, you have reached your AI usage limit." You can be software or you can be strict,
"Hey, stop burning tokens." So you can have, uh, you use your own language.

And also we went to the groups, the same thing. Remember the groups that I have told
you? You assign them to the people, you set the permission. You can assign them to the
groups. Like, hey, for example, uh, for, uh, data analysts, no, I don't want... I want
to give them, for example, ten millions or, um, m-more than the others peoples. Like,
for example, you can give them all one millions, but for the data analysts, like two
millions or whatever. They can just set up according to which teams you think they
deserve to have more tokens or they can, uh, you think like their, uh, usage is like
important. So that's how we can go with the group limits. And this is how Metabase
answered this question.

Like the usage is you set it up and you see, um, um, uh, how, how it grows. Now, the
fourth one, what we said is how do you guide the AI so it understand your business
properly? I always say it in our... Let me share this again. Uh, here we can see MetaBot
is explaining this to us. There are three layers with Metabase of how to make sure your
AI answer you correctly and understand your business. The first one is to set up correct
table metadata. Second one is to set up glossaries, and the last one is to set up system
prompts. Let's start with this, uh, table metadata. If I go back to Metabase and... Oops
Okay. And if I go to table metadata from the admin, and let's, for example, this is our
databases that connected to my database.

Let's select the first one. And for example, here I have doctors tables. Now, one
important things-- Two things you need to set them in the tables are description for the
table itself. For example, here, the doctor tables, I have set up a description, uh, for
it, like doctor... uh, directory of doctors in the healthcare system. One row represent
one doctor, includes their contact information, professional details or relate. So I put
all the information here. So when the AI, whoever, like, uh, if through an MCP or
through Metabot inside Metabase scan the tables, they will read the description, and
then they can ha-- knows if they want to select this table or not.

And then we go to each fields, like we have a doctor ID. Uh, we select... Also, we put,
uh, a description for this doctor ID, and we select the semantic type as well. And you
put it for all the tables that you are having it. So this is a very important things.
The better metadata you can do it, you can-- better metadata you have in your data, the
better results with the AI. And this is, like, a very important how Metabase can do
this, uh, for you. You can do it, uh, one by one through the UI. And also there are ways
through serialization that you can do it, uh, in a bulky way that like, "Hey, all my
tables just wanna set up the metadata for it." Okay?

So this is the first layer that I talked is about, um, table metadata. Oops. Oops, oops,
oops Okay? The second one is the glossary. So if I go back to Metabase, if I go to Data
Studio, and then if I see here, we have something called the glossary Glossary are terms
and definitions that you set them that your internal company knows about it, and then it
can be, uh, the AI can knows about it. For example, I added big fish. Big fish is for me
is any customer with the ARR that above fifty K, not an actual fish. So the AI doesn't
need like, "Oh, there is a big fish here." So if I ask AI, "Hey, who is our big fish?"
They can go there and talk about it. For example, I add zombie customers.

Hey, customer who's still paying us, but they are not using the product in the last t-
ninety days. So these are like internal terms that we use them, for example. These are
all examples, as I'm just tell- telling you. That, um, uh, the AI can read them, and
they can understand more context they can give them. Uh, for example, also activation
you can put or how your ARR is it about or the dependency. Um, our-- we call our MetaBot
is like the new intern in the company because, like, he is the person that you... sadly
to replace the intern. No need to have an intern. You can just ask MetaBot, and then
they can answer you. And they can read the docs as well.

So this is the, uh, about, uh, the glossary. And the last thing is, uh, how to make sure
that your AI answer you better is system prompts. So let's go back to Meta- Metabase.
I'm going to the admin, and then AI, and then here you will see we have something called
system prompts. There are three sections for the system prompts. We have the AI chat,
the natural language queries, and the SQL generation. So here we give you a blank that
you can write anything, and this will be instructions for your MetaBot or for the AI
that is using it. Um, you can, like, personally, I do, like, how to answer. Ke- lead
with the answer and then with the number, and then one line of context helps.

Personally, I don't like when I'm talking to the LLMs, and they give me big paragraphs
that I have to, uh, read up after it to review it. Just I want to see the s- the, a
small context. And then, but I told keep it to three sentences and no, uh, apologies if
they are asking apologies. But, but I don't want it also to be shortened the context,
and then I'm not aware of what they are choosing. So I told never guess a number. It
doesn't ever g- it doesn't, uh, guess a number. And if the da- data doesn't answer it,
say what is missing and where it would actually come from. And if the term could mean
two things, check the glossary first. If it's still unclear, ask me questions.

And the same thing, you can add what to use or how to, what to hand back. So you-- it's
up to you to see how exactly when you are talking to the chat inside Metabase to set
instruction for it. The same thing with the natural language queries Um, you can just
like, for example, uh, for me, default time range is the last thirty days unless the
question names another one. If I exactly told like, "No, uh, give me different time
range." Exclude canceled and refunded orders unless I ask to include them. So the same
thing you can put, uh, when you're using, um, in the AI exploration. So if Metabase
here, let me show you the AI exploration main app and new AI exploration.

Here are the instruction for the AI, uh, that we are setting them up. So if we go back
to admin, AI system prompts, these are the natural queries. And the last one is for the
sequels. Um, you can also put, uh, uh, uh, instructions for the sequels. Um, so it's up
to you if you have, like, your own way on doing the sequels, like this is something that
you can do it. Add something in correctness if you're using the MetaBot inside Metabase
to correct you. So this is like something you can add it, uh, for alliance as well and
other stuff you can do it. So this is something in the fourth, uh, f- third thing that
you need to take care about table metadata, uh, about the glossary, and about the system
prompts Um, okay.

Okay. Uh, there is something, the last question is about how to see what are y- uh, what
you wanna see the usage to people, what they are using. So we have here something called
usage, uh, auditing. If I come here, we have a section for stats. It says statistics to
see, uh, what are the conversation happening during the day. I can see here, for
example, uh, yesterday we had like nine conversations. These are the users, uh, these
are the profile for them, how many tokens they have been used, if there is a cache
tokens as well, how many queries, how many search. Uh, other than that, we have here
conversation by source. So it's like these are breakdowns for the data, conversation by
profile.

Um, groups, like these are the groups that they have like most conversation. For
example, administrator are using a lot, user with the most conversation. So we have here
a set of all statistics of the data they're using. We can al- And we also added tokens,
so you can also manage your token to see your input tokens and your output tokens. Uh,
you can see also the analytics here. You can also apply groups, uh, filtration for the
groups or the users, also for, for the times. Uh, or it's also breakdown by tokens and
people. And the last one is, uh, messages. You can see the amount of messages per day,
uh, or by profile, by source, and so on. The second one is the conversation, and this is
also important when you are like, uh, you have faced a problem with them.

Um, for example, here, it's me, I, I ask for messa- uh, messages. I can open these
messages. I can see, "Hey, do you have any customer with the error?" And then it went
there, it searched. And then I ask, "Hey, what's the big fish?" It told me, "Big fish,
an, any customer with the error above fifty K. It's not an actual fish." So this is
something I set it in the, uh, in the glossary. And if I go back, I can see like all the
messages that I have, uh, text them. I can go and then I can see one by one, and then
these are the queries that it generated for me. I can also have a history for it. And
this is the really great part that here, uh, when you're working alone and, uh, you're
not working with team, it's, it's probably you'll not be taking care, but when you have
a big team, it's usually across the whole co- uh, across the whole company, this is very
good like to go back, uh, to see who is using what, how much they are using.

Oh, um, you're not seeing my screen. Okay. I'm sorry for this. Can you see my screen
now? Wait, the whole time you didn't see my screen? So- Here's stats. Um, okay, now, now
can you see my screen? Okay, perfect. Sorry for this. So this is what I'm saying here.
In the users auditing, we have stats. This is what I'm talking about, the statistics.
You can see graph, conversation by day, and breakdown by the people. Uh, we can also see
by tokens. We can see the input tokens, the output tokens. You can filter for the
groups, you can filter according to the people, and you can see also the tokens
breakdown by source or by profile and so on. And also, you can see the number of
messages.

Like, hey, how many messages have been texted? Uh, also breakdown by source, by profile,
and groups with the most messages, or user with the most messages as well. Also IP
addresses. And I'm talking here about the conversations. These are the conversation we
have done in this Metabase instance. Like, if I come to my Meta-- I, I already sent
this, so I can see this is what I did, uh, four messages between me and MetaBot. This is
the total tokens, input, output, uh, zero query runs, but two search. It did for me two
search. Like for example, here I asked a question, and these are the results of this
question. So you have a history of everything that you talk with MetaBot inside, uh, uh,
in-inside Metabase.

And if you run, uh, into, uh, a query, it will show you the query. Let me show you for
example here. I did a query. It show me the query that like, hey, I, it connect, um... I
was counting how many doctors we have, so it went, it-- I can see the query itself that
I have run it. Okay? So for now, this user-usage analytics is about MetaBot, but it's
coming soon. Um, now this is what I want to share, that MCP usage, so w-everything you
type with MCP, we will have access for it and with the CLI usage that, uh, anything you
text through your CLI to Metabase, you will have also a usage for it. It will be the
same thing. You will have statistics, and you will see, uh, the messages there.

And this will take us to our second poll, which will be, um... It's active, the second
poll? Yep. Uh, about bringing your own AI models. Do you wanna use, uh, AI models, like
it's, it's nice to have, or do you have any hard requirement? Let's take a few moments
here so you can answer it. Okay, so I can see it's preferred. Fifty, fifty-three percent
of you have preferred to, uh, to bring your o- uh, AI model. And, um, in Metabase,
actually, you can... Let me share my screen, make sure that you'll see it So in
Metabase, this is currently also for our open source and for the paid users. Um, you can
connect to any AI provider. So for example, here If I come, uh, these are the options.

I can connect it to MetaBa- Metabase. It's our AI, uh, service that you can connect it
directly to us, you pay for us. Or if you don't want to take our, uh, service, you can
go with Anthropic or Microsoft Azure or Amazon Bedrock, uh, or even Mistral or OpenAI,
OpenRouter. So, uh, you have more options, and you don't pay for us. Technically, you
pay for the usage with these providers. And even when you put your API key and you
connect it, you can select which model you want to collect, uh, select. So you can
select, hey, wanna have Fable five. This will burn all the tokens, so of course, don't
use it. Uh, or you can go with, uh, all the options that y- your API, uh, can provide it
for you, so you can have this option.

And also this is coming soon, you can bring, uh, different models. So you can connect to
different models at the same time. Uh, so make sure that you always see our re- releases
in the upcoming weeks and months. So we will announce about a new feature for, for this
one. And when you select also one of the things you have them, uh, for MetaBot exactly,
uh, these are, uh, you can enable MetaBot or you can disable it, like I don't-- if you
don't want. Uh, there are collection that you can select it for the natural language
query that preferred, uh, co- uh, uh, collections. This doesn't mean they will not see
the other stuff that you don't see, but like you can say, uh, that I can select, uh,
uh...

It's an empty Metabase here. I can show you, um, this one here. AI setting. Here, for
example. Like in this instance, we selected the people, we selected Metabase to always
go to library and then to select the data from the library. If you're not familiar with
the library, I will just say it in a very quick. We have a data studio that you clean up
your data, and then you publish the tables, the clean tables in the library, and then it
will be the starting point whenever you are asking any question, which means that you're
telling all your, uh, users, "These are a clean table, so please use them." Like for
example, everybody have accounts table, 10 accounts tables, and whenever any user comes
and they want to do a query, they don't know which account table to use.

But with the library, it is like a source of truth. The data, uh, analyst or the data
engineers are telling the users, the business users, "Hey, this is the accounts that
you're using it." And that's why in our, uh, AI, if I go back to the AI, and this is
like a collection, I said like, "Hey, choose the library. Choose the data inside the
collection." But doesn't mean it doesn't have access for the other stuff for the user
already having it. They already have an-- Metabot have access to everybody what they are
seeing, but these are like, uh, the first thing appear for them when they are talking in
the AI exploration. Also have the option to, uh, generate, um, prompt, uh, for the
users.

So, uh, now it's not running prompt because it's connecting. Um, but the prompts are
technically just like, uh, shortcuts when you're talking to Metabase-- uh, Metabot, it
can give you the shortcuts. And you have the option to disable all the AI features. So
as I said, coming soon for the MCP, you will have, uh, an ax-- uh, we will have a usage
analytics for the MCP and also with the, uh, UI, uh, CLI as well And this will take us
to Polls three. Let's, uh, share Poll three Okay. Uh, people are answering it Okay, the
answer is, um, you don't want your data to leave your network. Um, so the best way is
you self-host, uh, you're using the self-hosting, you self-host Metabase inside, uh, in
your infrastructure.

And, um, uh, for now, Amazon Bedrock is the best, uh, option that we are having it. If
you already have Amazon Bedrock that you can connect it there in the own infrastructure,
uh, so you don't need to, uh, have any, uh... So your data will not be leaving your
network in this case. All right. Um, let's see. Q&A I'm just-- I will be reading this
one. Is there any way in Metabase where I can specifically preferences regarding which
table are best for searching for specific data? I have already populated the data So,
uh, to answer you, uh, Maria, is as same what is about I'm talking about the library.
Uh, we don't have any current option to go to tell it specifically to go to table A if
I'm telling you about this.

Uh, this can be something in the description that you can set it up for each table that
like if this the table that, uh, talking about this topic, you're saying about like if
you need balance data, go to table A. So this is something you need to add it to table A
description. So whenever the Meta bot is reading your de- descriptions, they will know
about the topic. So they will know like, "Oh, we're talking about balance data, so I
will go there." This is now the current, uh, option that we are having it. Uh, we don't
have like a dedicated option that you can tell it, uh, like pick and choose. Uh, but it
would be nice actually, actually, a feature request Um, let's see the other Okay.

We'll be supporting open source model. Um, for now, these are the, uh, connection that
we are having them. Um, uh, we don't connect to support any like Ollama yet for the
MetaBase, but these are actually, uh, probably we have opened the feature request for
it. So, uh, I think Andrew already shared this feature request with you, so you can, um,
uh, pop it up, and then we can also sh-share it with our internal team to tell them
about it as well. Okay. Does the data stay in library or in the DB connection? Is there
any refresh timing or real-time connecting? All right, that's good, really good
question. So the data actually, Metabase is a very thin layer, sit on the top of your,
uh, metab- in your database.

Your data doesn't move anywhere. Metabase run queries against your database, and then
you will get the results in your database. So when you connect, if I go to databases,
and here are the name of the databases that I have. Uh, you can sync the, uh, sync for
you the database schema or re-rescan, uh, the s- uh, fields. But here you can set the
advanced option is the cadencey. Uh, not this one actually Okay. So here you can choose
when it sync and scan happen. So you can set a cadency for syncing the database, uh, on
a daily basis or an hourly basis. And, uh, this will scan be, uh, scanning your schema,
and that's how actually Metabase works. It knows your schema.

If you run the queries, the queries happens in your database, and then the results of it
will be appearing in Metabase. You can cache the results, uh, in Metabase. Is this
something, um, I answered you, H-Harshal? Okay, let's see the second question. Uh, do
you have an estimate when it will be possible to integrate, uh, Vertex AI with Metabot?
Uh, it's a good question. Uh, we need to check with the internal team. Uh, we don't have
any, um, we don't have any timing, uh, for it. As for now, we are, uh, these are the
new... What they are, let me share them again These are the, uh, supported AI providers.
We try- we started with actually only Anthropic and, uh, uh, OpenAI, but now we are
running more and more.

Uh, so keep your eyes with the upcoming releases or our, and our, uh, feature requests,
like people already asking for other AI, uh, providers. So we will try to accommodate
everything Okay. Jeffrey, you have, uh, Metabot is read-only. Is that correct? While MCP
and CLI have the full curd access. Um, Metabot create for you the query, and then you
just, uh, run it, uh, the results of the query, you will see the query itself, and then
you can save this query as a question or not. Um, the same thing with an MCP. The MCP
can be, uh, running through, um, running the queries for you, and then it will be, uh,
you can save them inside Metabase or, or not. Or do you-- are you asking about writing
to, to the database?

Let me know if I answered you, Jeffrey Um, let's see another questions. I like everyone
is asking all good questions Uh, from, uh, sorry for pronouncing your name wrongly. Um,
Jochen, uh, how do you recommend setting the meta, uh, data properly if we would like to
save our metadata in a repo which is applied in Metabase rather than setting everything
manually in Metabase UI? Good questions. Uh, you can use, uh, again, uh, serialization
that you can u- export a YAML file of your metadata, and you can use here any, uh, LLM
code agent that be setting it up for you, and then you can again import it, uh, to
Metabase. It will inherit all of this thing. And, uh, I can just like, uh, ask Andrew or
Perrina here, is remote sync can be done in this phase or, uh, uh, remote sync cannot
work in this case if they want to v- see version control in here?

Okay. Another question is: What is the use of library in Data Studio? So again, as I
said, here is the Data Studio and the library, uh, not this Metabase. Let me show you,
um, this Metabase, another instance Okay, so here's the library. You publish tables to
the library. It's, uh, like a workbench for you to set up all the, uh, see the table.
Uh, you can have observability on the fields. You can set up segments there. Uh, like
there are, uh, uh, a filtration that you can create it, and then you can, uh, to be
saving them. Same thing with the measures. Measures are saved aggregations that you set
them up. This is like your semantic, uh, one of the part to build up your semantic
types, and you can see the dependency of this table, what is already using it.

So again, as I said, let's say we have accounts from several tables. Uh, so when you go
to tables and then, uh, let's say for example, I have two sample databases. I have the
sample database coming... the orders table coming from this database, and I have another
order table that is coming from another table, uh, from this database. And here you
select which one you already have set up the semantic type for each, uh, fields, and
then you publish this table. You can see here, uh, there is a, a green tick that it's
published for you. So which mean if I go to the library, I can see these tables are
published for me here. Now, on a level, if I go to the main app and then I start a new
question, I will see here it's promoting, telling me, "Hey, select a data or select a
metric." If I select the data, it will shows me the me- tables that I have published
them in the library.

So like this, I'm telling the business user, "Hey, don't need... You don't need to go to
search for an- other tables. This is the source of truth for these tables." Like here
now, for example, we published two accounts, which is wrong in this case. So you need to
publish one account table so that user table knows or if you have different, of course,
uh, segments. But if they have access for the other stuff in your, uh, in the metabase,
they can browse all, and then they go and select, um, what, what they have permission
from the collections or from the databases. Now I have full permissions for all the
databases, but they c- they have the possibility to go there.

I hope I answered your question here. Okay, let's go back to AI. Um, could the AI do
filtering where it goes down through hundreds of thousands of rows and then provide
recommended KPI level on which product are outlier? Or can the AI virtually just see the
data as your dashboard see it, and you have manually to expose each filter item before
it has the same data to work with? Now, just to answer this, before it depends on the
model that you're using. Of course, you're using Fable 5 from Anthropic will be
different from using, um, GPT-3, for example. So this is number one. Number two is it
goes and read your, uh, fields. It reads, um... Now, inside Metabase, it can knows your
existing questions, so it will be a dashboard.

It will be bringing this answer for you from the questions that they already have them
in your dashboard. But it also, if you don't have this already question, it goes and
read your metadata, it know the fields, and then it will build the query for you, and
then you can see the query live for you. So for example, if I see here, uh, how many
customers So lastly I give it like a filtering for it. Um, it will go search, uh, okay,
it will be viewing accounts Oh, okay. You're asking also more specific that you are, are
a outlier, um I'm just making-- I'm reading your, uh, questions, uh, again. Apology Can
the AI virtually see the data as your dashboard see it?

Um, so this question too, just like it doesn't see the data itself, it run the query for
you. Um, so you can just like apply... If you already know that there are outliers, then
you need to put like a option for it. Like, hey, the outliers, um, make sure to remove
them, but set a condition for them for, for it, so it can understand it on which field
it can be an outlier. If you just say outlier, uh, it might not understand you directly.
So if you put more, uh, instruction in your, uh, more conditions in your prompt, it will
give you the answer. But again, what I want to show you here, like here's the answer.
Like I can see editor, it just run for me the query.

It selected the account, uh, table, it applied the filter, and then it applied summarize
for me. Uh, it didn't know what is happening behind the scene. It doesn't know what is
the tables. Like it just knows this is how it should be. So if you want more conditions
in your prompts, you need to set it up for more, uh, for the prompts Um, let's see what
we have Okay, also from Jeffrey It helps understand that Metabase is a thin layer on
application since you would want MetaBot to edit source of database. Does MetaBot have
access to write application? Uh, no, MetaBot doesn't have any access to write. This is
something that you set up in Metabase level. Uh, um, when you're connecting it to your
database, you connect it either, uh, without, uh, writing permission, and we al- also
have an option that you can have writable connection when you're doing transformations
for your data, so you can write back to your database.

So MetaBot inside Metabase, it only runs the query for you here, and you have the option
to save them, and you don't save them. And, uh, as me as admin in Metabase, if I, um,
uh, if I have a write permissions in the data- in my database, the only thing I can do
is edit the tables from the database here, which is something that MetaBot cannot do it
for you. So anything writing to your database, uh, MetaBot is just, uh, can run the
queries for you Okay, the last question from Herschel. Does library data CD respect RLS?
I'm not talking about tenant isolation, but traditional RLS. Um, yes, RLS is set up
according to the groups that you set it them, uh, for, uh, at the beginning that I
explained.

Uh, for if I go to admin, permissions, so, uh, let's say here, row-level security, uh,
group, we call it like that. Now, this group, row-level security group, we call it the
row-level security group. They have a granule, uh, for the sample database, and we set
up for them for the people ty- uh, people, we set for them the row column security. So
if I publish this table in the library, it will every user, uh... It will respect every
user permissions. So if, if someone is blocked from this table, then they cannot see
this table. If they have the row column security in this table, then they will apply for
it. And then if they have the option to can view it, then they can view it.

So this is the permissions. We have the full page here. You set it up for each table. I
hope I answered you. And the last question is, can we get a transcript of this meeting?
Yes, this meeting is already recorded. You will have access. Uh, we will send you a
link, uh, through your email after half an hour or depends if this tool will not break.
And, uh, we will be uploading, uh, the, uh, the recorded in YouTube so you can watch it
later. And tomorrow, we r- also reminder that tomorrow we have, uh, another session with
Matthew. It will be more fun. Again, here it was more admin stuff setting up, which is
very, very, very important. Uh, but I really recommend everyone to go tomorrow to watch,
uh, Matthew, um, webinar.

It will be very fun. You will see it. Okay Um Any more last questions? Okay, go once
Wise. All right. Thank you so everyone for joining. I hope you enjoyed it. If you have
more questions, uh, you already can access through Discord and, um, as a PA, as a
product advocate, you'll always have-- can have access to meet with us to know your use
case more. Y- we can ask more questions. You can, uh, always send us an email and, uh,
we, we are happy to help you out. Thank you everyone. Enjoy your day. Bye bye.