Latest / Agentic Conversations (formally mlops.community) / Conversation with the MLflow Maintainers

Transcript

Conversation with the MLflow Maintainers · Agentic Conversations (formally mlops.community)

0:00
All right, folks, I'm super excited for this conversation
0:02
that I just had with Jules, Danny, and Corey.
0:05
They are on the ML Flow team. They are veterans in the space.
0:10
Jules literally wrote the book on Spark.
0:13
Now he is leading the developer relations side of ML Flow, and
0:17
Danny and Corey are both lead engineers crafting what we know
0:23
and love about this classic tool.
0:25
And they talked about how ML Flow is leading with open
0:31
source. They really want to go all in on
0:34
the open source ML Flow and bring it to this new agent
0:38
paradigm that we're in. Enough of me chatting.
0:41
Let's get into the conversation with it.
0:49
Not every day that I get to hang out with so many incredible
0:53
people, I am graced by not one, not 2, but three of the ML Flow
0:59
team. How y'all doing today?
1:01
I'm super stoked to chat about agents in production.
1:05
You guys have been doing a lot with it and I think each one of
1:09
you individually has been on this podcast before, just in
1:14
different phases of your lives. It's cool to see you all working
1:18
on ML Flow and doing things in this new.
1:22
Agent era, absolutely. I think all of us have come a
1:28
long way and at the same time, some of us are continuing to
1:32
focus on developer experience for AI in a radically different
1:38
domain, models to agents, feature stores to agents as it's
1:43
kind of exciting. Thanks for having us back.
1:45
Yes, we all, we all, we all, we all will come a long way from
1:49
from Spark to MO Flow and back to Spark and back to Emma flow.
1:53
So I'm delighted to be working again with Danny and with Corey
1:58
and obviously talking to you. We have had many interactions.
2:01
I just love you. Me serenading on, on, on on the
2:04
stage with your with your Kalela.
2:06
That was that was the highlight. Absolutely.
2:08
That was the highlight. Yeah, Well, I, I feel like a lot
2:13
of this is reflective just with, it's not just like what like
2:16
what the three of us have been doing.
2:17
It's also like MO OPS community is like been shifting as well.
2:20
Like we've we kind of all working in various versions of
2:24
the, the classic MO world before and then it kind of evolved
2:27
shifted over to this like NAI Asians kind of OPS space, right?
2:31
So. Yeah, it evolved real quick.
2:35
And so Speaking of evolving, like MO Flow has evolved a ton.
2:42
You guys have been seeing folks that are putting their agents
2:45
into production. I would love to start out with
2:49
just what have you been seeing out there?
2:51
Each one of you comes at it from a different perspective, so
2:53
maybe we can get your takes on the conversations you're having
2:59
each day. Yeah, 100%.
3:01
I think as we talked about agents in production, the first
3:06
natural question to ask is what are people trying to put in
3:10
production, what are they building and and what are they
3:12
deploying? And after talking to a lot of
3:16
our customers, open source users can pretty confidently declare
3:20
that the age of the chat bots is is still here.
3:25
It ain't going anywhere. The the bulk of our customers,
3:29
the bulk of our open source users, they're building chat
3:32
bots. Yes, you know, some of them that
3:34
started as text only bots a couple years ago have become
3:37
multimodal. And we do get the occasional
3:40
like voice application. Folks are doing speech to text
3:43
and then they get the text back and they go to text to speech.
3:46
Some are processing images, but the end of the day like that
3:49
chat interface remains pretty ubiquitous, like well over, you
3:55
know, 50 or 60% of use cases, which is, is massive.
3:59
The other thing that I'd say about chat bots and there there
4:02
are a bunch of other use cases that I'll let Danny and others
4:04
talk about is that they're pretty much all tool calling
4:07
chat bots now. And we, we have the age of Rag
4:11
up until about a year ago that everybody was really excited
4:15
about document retrieval specifically.
4:17
And now, yeah, folks are continuing to retrieve
4:20
documents. They're they're gathering
4:22
information from knowledge bases, they're indexing them,
4:25
but they're doing so much more around that as well.
4:27
You know, they're, they're making real time API calls to
4:29
other services, other services to fetch information.
4:33
They're even having their chat bots take various actions.
4:36
You know, I'm going to go ahead and save some information about
4:38
this user to a knowledge base. So there's not just read,
4:41
there's a little bit of rights coming in as well.
4:44
So this chatbot interface has remained somewhat consistent.
4:46
Maybe it's become multimodal, but then under the hood, the
4:50
architecture is has evolved quite a bit.
4:52
So thanks to the powerful models that seem to be being released
4:57
every week or two now, I think that's what you can probably
5:00
attribute to a lot of the upgrades behind the scenes.
5:02
So that that's my pitch on on chatbot use cases.
5:08
Danny, Jules, what else are we seeing?
5:10
Yeah, well just elaborate on that.
5:12
I think chatbots have all what I've seen is chatbots are
5:16
getting more and more sophisticated.
5:17
So maybe an initial just like text based chat chatbots now
5:20
we're seeing like they're there's a lot more like rich
5:24
kind of return types, like maybe like rich HTML that's being
5:26
returned or like more human in the loops or within this agent.
5:30
Now you're adding like kind of multi agent systems or but it's
5:35
still like a chat interface to the user or there's a basically
5:39
yeah, the infrastructure underneath people are building a
5:41
like more complex like guardrails within the agent, you
5:43
know, just to as they get closer and closer to production or
5:46
working with higher stakes use cases, I think people are now
5:49
just beginning to get some of those like all you can't just
5:52
Willy nilly launch some of these production kind of use case on
5:54
the production now. So, so that's been cool to see
5:59
and. Yeah, I've seen.
6:02
Let me jump in there, Danny, because when you talk about the
6:05
advanced, the chatbot is a little bit more advanced.
6:09
I don't know if you guys saw this website called AIUX design
6:14
Patterns. It is brilliant.
6:17
It gives you all these different ways that folks are levelling up
6:22
chat. And right now a lot of them are
6:27
just kind of concepts. But you have things like you can
6:31
highlight certain parts of the response in a chat and then
6:36
expand on only that part, you know, which is so nice as a user
6:42
experience because a lot of times you want to say like, Oh
6:46
yeah, I like all of this, but I don't like that part.
6:49
Or I've also seen designs where they'll come back with a
6:55
carousel of items for you that you can scroll through like
6:59
we're natively used to using, as opposed to only chat in, chat
7:04
out, that type of thing. But yeah, you were going to say
7:06
something else, Danny, so hit me with what else you were.
7:08
Mentioning. Well, that kind of reminds me of
7:11
a different thing that I'll touch on as well.
7:13
But I was going to say, even though all the chat bots are
7:17
getting more and more complex, it's surprising to me how
7:21
nascent the observability and the eval space has been for for
7:25
these chat bots. Like you'll notice that the
7:28
current discourse is a lot about like multi turn evaluation
7:31
assimilation, how like basically every vendor or solution still
7:35
is like really just kind of be in to get the free wet.
7:38
But like multi turns, like conversations in chat bots are
7:40
naturally multi turns. So like, you know, why hasn't
7:42
there been kind of more robust tooling in that space?
7:44
And really a lot of customers have had to just work around
7:46
that and build their own tooling that's like super bespoke.
7:51
So yeah, that that's been interesting.
7:52
Yeah, I mean, I mean talking to Boatwood, Corey and and Danny
7:55
say they'd be actually begin to see my conversations having a
7:59
lot of developers out there. The chat boat is seems to be the
8:02
favorite one and the and and as Danny pointed out, more of these
8:06
evaluation judges are being written.
8:10
For example, deep eval and Ragas do provide a very specific board
8:14
built in as well as custom judges.
8:16
You can actually do multi turn evaluations.
8:18
So we begin to see more of that in that particular area where
8:23
chat ports are being written, but they begin to feel the
8:25
confident that they can actually write these judges to evaluate
8:28
the multi turn sessions, to evaluate the entire granular
8:30
session. And that somehow seems to be the
8:34
the trend, at least from what I've seen from developers trying
8:37
to ask questions or, or hit me with, you know, how are, how are
8:41
you actually evaluating your chat board in production?
8:43
I'm using my I'm using either ML flow, I'm using some framework
8:47
that actually has a host of built in judges that allow me to
8:52
gauge a particular metric, or I can write my custom judges along
8:55
with agent as a judge. So we we begin to see quite a
8:58
bit of that trend. But yeah, I I think it people
9:04
are beginning to talk about multi turn judges now, but not
9:07
because not because like they didn't need it.
9:10
Like I think as soon as we launched multi turn judges or we
9:12
started talking about other people were like, yeah, I need
9:14
that. That's like the first thing I
9:15
want to start with. It's like, you know, I want to
9:18
know is my chap is are humans like escalating to get like, you
9:22
know, people trying to get escalated to human support?
9:25
Are they, is the agents repeating it over and over again
9:28
like all these like common, like Asian problems you really start
9:31
seeing in the context of a conversation rather than in a
9:33
single turn? And really as soon as we
9:35
launched it, people are like, cool, yeah, we're going to use
9:37
that immediately. We needed that yesterday an.
9:40
Example of this that I like to use a fair bit is determining
9:45
whether an agent is being repetitive.
9:47
I can't figure that out just by looking at one input output.
9:51
No clue. But if I can see a, you know, 20
9:53
turn conversation, my God, this agent keeps saying the same
9:56
thing. It's giving me the same answer I
9:57
ask For more information and it continues to respond with
10:01
basically the same stuff rephrased, it becomes a lot more
10:04
obvious. And so that's where you know, I,
10:07
I can't figure that out If I have this little function, this
10:09
thing called a judge is, is looking turn by turn.
10:13
I have to have the bigger picture.
10:15
Same thing comes with human feedback though, that, you know,
10:18
a lot of times folks, rightfully so, are using their end users or
10:23
internal testers and whatnot to determine whether a chat bot
10:27
meets their needs and they need that conversational feedback.
10:33
Indicating that this particular message didn't make a ton of
10:36
sense is kind of helpful, but it's much more useful if I can
10:41
say this entire interaction should have been half as long
10:45
or, you know, I, I got this is what I was trying to achieve
10:49
with this interaction and this is where I got and what was
10:52
missing. So I think there's an element of
10:54
human feedback. There's there are these
10:56
questions about how I automate the process of finding issues in
11:00
these multi turn interactions and bugs at my agent.
11:04
And it's funny, it's taken us long this long to get here and
11:06
that everything started with sort of single turn input
11:08
output, because this chatbot interface is especially this
11:11
multi turn thing was the first thing launched, you know,
11:14
November of of the three years ago now, you know, ChatGPT was a
11:19
multi turn chat interface. And so it's kind of funny that
11:22
we've finally started to catch up in terms of the evaluation
11:26
and the quality measurement side.
11:29
Yeah, you touched you touched on 2 very selling .1 is the
11:32
repetition right? How do I know that if I had if I
11:36
had a 20 turn conversation, something was not being
11:38
repeated. And the third was the logical
11:41
consistency is, is there a logical flow between each and
11:44
every steps or am I being consistent or I've been
11:47
contradictory? And I think those are the
11:49
important elements where where the judges can actually help to
11:52
detect in the multi turn conversation 1 the coherence.
11:56
Is there a logical flow B is a context B retention.
11:59
I refer to something you know, 20 turns before is my agent or
12:04
is my conversation board has has the ability to remember that in
12:08
the context. So I think those are the
12:10
important thing and that actually collectively those met
12:13
make the the chat board more robust and and and and more
12:17
cohesive. Yeah.
12:18
So you, you touched on that important.
12:19
Point yeah, 'cause you kind of take it for granted us in this
12:23
field that we understand, oh, potentially a lot of times to
12:27
get the best result we should just start a new chat and clear
12:30
the context window. But that's not like I would
12:35
guess that your average user off the street doesn't know that.
12:40
And so how can you make sure that what you're talking about,
12:43
Jules, that stays true? Like if somebody wants to
12:47
reference back and you're not doing it properly, you know that
12:51
and you have a way to either be at the janitor of that context
12:58
or encourage them to start a new chat or do something because
13:02
that just leads to poor user experience. 100%, And I think it
13:08
dovetails nicely with this whole topic of memory management for
13:12
agents that that everybody seems very rightfully excited about
13:16
the idea that a very long message history gives an agent a
13:21
ton of rich context. If, if I've been talking with a
13:26
chatbot for 15 minutes and then I ask it to do something that
13:31
follows naturally from that 15 minutes conversation, I'm much
13:34
more likely, I would guess, to to get the result that I'm
13:36
looking for. At the same time, there is that
13:40
trapdoor pitfall that I think you rightly point out where if
13:43
I'm asking for something totally different, you know that 15
13:47
minutes of context can hurt more than it helps.
13:50
And so to me, I view it as a problem of memory and state
13:54
management in an agent across all the conversations that I've
13:59
had with this particular chatbot.
14:01
When I go and I ask for something, whether it's in the
14:03
middle of an active conversation or a brand new one, what
14:06
information about me or about my intense or previous interactions
14:11
is this agent drying from? And that's something that they
14:15
were taking a pretty close look at, at Databricks and then from
14:19
a research perspective as well. And, and I think it's an
14:22
important topic for the industry at large, just agent memory.
14:25
Yeah, yeah. And how do you weight memories?
14:28
Exactly. So Corey, Corey, you're
14:30
referring to the short term memory and long term memory
14:32
where if the 50 minute conversation has has has just
14:37
passed and nothing has been eluted in further conversation,
14:40
that 15 minute turn can be actually put in a long term
14:43
memory and can be retrieved on demand, right?
14:46
And the short, short term memory is just like, OK, I'm going to
14:48
flush this out and I'm going to start a new session.
14:51
But I'll keep that in short term memory.
14:52
If there is a reference to certain thing that I talked
14:56
about previously that doesn't exist in my short term memory,
14:58
I'm going to go and refer to the long term memory to see if the
15:01
context is there. And that way, as you say it, it
15:05
maintains a good flow and you as a user has a better experience
15:09
to say, OK, you know, this guy remembers everything for the
15:11
last half an hour that I actually had.
15:13
So it's like a human really be paying attention to you and, and
15:17
referring to something that you had asked about 15 to 20 minutes
15:20
half an hour ago, that is. And that is like a moment.
15:23
Oh, wow, this is this is brilliant.
15:25
I can't, I can't believe half an hour, half an hour ago, I asked
15:28
it to do something and I referred it to it.
15:30
And now he actually comes back and gives me a full answer.
15:33
That is definitely aha moment. Magical I'll tell you what
15:38
though, when I encounter chat bots in the wild and they don't
15:42
work that well, I am very adamant caps locks and
15:47
everything saying how bad this damn bot is and I just like I'm
15:53
trying to flag it so whoever is doing evals on the other side,
15:57
they see that like oh maybe we should try and make it better in
16:02
this area. You know it's.
16:04
Funny because 1 customer that we worked with, they were kind of
16:09
like they lost the production without really doing too much
16:11
for us Evo. And like at some point they were
16:14
looking at all their like responses there.
16:15
It's just like like I think most of their cast were just people
16:19
swearing at the chat bot. And so really, it was like
16:22
trying to serve a different purpose, but ended up just being
16:24
a therapy chat bot for all of all their users.
16:28
But but interesting, Danny, when you, when you become abusive to
16:31
the Chatbone, if you have a judge that actually that, that,
16:34
that has the frustration, like for example, in Nemo Flow, we
16:36
have the built in evaluation called user frustration, right?
16:41
And so it detects, it detects your language and tonality and
16:45
based on that, it would actually respond.
16:47
Sometimes it surprises you when you know, you, you were
16:49
frustrated about this topic and you, you would come back with a
16:52
nice polite answer to mitigate the frustration.
16:55
I think that's an interesting way how how chatbots can somehow
16:58
appear like human interaction because of the fact that we
17:01
actually have this old both built in as well as custom
17:04
judges that you can actually write and you can specify and
17:07
tailor it to your to your chat bot, right.
17:10
That the chat bot is talking about insects then, and
17:12
everything dealing with insects would be sort of more tailored
17:15
to that. The other piece of this, beyond
17:19
yelling at the chat bot, is the degree to which folks are
17:23
collecting like thumbs up, thumbs down, categorical and and
17:28
free form feedback from their chat bots.
17:31
I I noticed that a lot of folks, the first time they deploy to
17:35
production, haven't really thought about collecting those
17:38
signals as much as you might expect.
17:41
As a result, you know, like Dimitrios, you resort to caps
17:44
locking and yelling at an agent in the chat window.
17:48
But it's often high signal for developers if they can get a
17:53
more structured set of feedback, if they can get thumbs up,
17:56
thumbs down, or they can get a 1 to 5 rating and they can, they
18:00
can carve out a separate text window or field for an
18:04
explanation about why things are going well or where that they or
18:07
where they can improve. I think a lot of folks treat
18:10
that as an afterthought. And it's a little bit
18:13
surprising, at least to me, because it's pretty fundamental
18:15
to A, figuring out if you deployed something that's
18:19
working reasonably well, and then B, automating that process
18:23
of fixing and identifying quality issues in the future.
18:26
Like it's hard to sort of automate the detection of a real
18:30
quality problem without having some early signal from your
18:33
users about what the quality problems are.
18:36
So the The upshot here would be feedback collection is massively
18:40
important and to some extent under invested in in the chat
18:44
bot ecosystem. And the interesting thing is
18:46
it's very easy to build all the all the stuff that folks are
18:49
doing around stitching together, you know, tool calling chat bots
18:53
with knowledge databases and external APIs and putting that
18:57
in an awesome web front end and storing data securely.
19:00
All that stuff is hard. You know, collecting a little
19:03
bit of end user feedback is pretty straightforward and all
19:05
shamelessly, you know, plug the ML flow feedback collection AP
19:09
is as being a particularly low friction and an easy product
19:13
surface to use. But the the point is like, spend
19:17
that extra hour that's that's really all it should take to
19:20
wire up some end user feedback to and understand the quality of
19:23
your agents. Corey, you, you, you touch on a
19:27
very important point. How do we actually take the
19:29
human feedback and what, and, and operationalize it from your,
19:33
you know, from your experience talking to customers, even
19:37
developers, how do you think they're actually doing it?
19:39
So when they have the MO flow deployed and, and they're going
19:42
through the UI to add additional feedback, how does it get
19:45
injected into, into the, into the traces so they can improve
19:50
that? Yeah.
19:52
It's a great question. So what you have with most
19:56
deployed agents where you've integrated some amount of
19:58
tracing is effectively this massive log statements, whether
20:03
you collected them with them all flow or a variety of other
20:05
solutions. You you have thousands, hundreds
20:08
of thousands, even millions of these time stamped logs.
20:12
You can open them up and you can see the statement by statement
20:14
stacked trace. What feedback helps you do is
20:17
filter those. So I want to.
20:20
You know, find the needles in this haystack in most of my
20:23
hundreds of thousands of interactions aren't necessarily
20:26
going to be labeled. The quality might be reasonably
20:29
good if folks didn't take the time to express their
20:32
frustration or, you know, provide some feedback.
20:35
And So what I would like to quickly do is ignore all the
20:38
stuff that I know is working well and I don't have to look at
20:41
closely. I want to, you know, click in
20:43
this UI, apply a filter for certain types of negative
20:48
feedback. You know, for example, user
20:50
frustration is is high. OK, I'm going to apply that
20:53
filter. Or the completeness of whatever
20:57
task the agent was being asked to perform is is low.
21:01
It's only providing partial information or not fully meeting
21:04
the users objective that condenses this log of thousands
21:09
and hundreds of thousands of stack traces into a few things
21:14
that I can then analyze more deeply and start to root cause.
21:18
It's great to understand that the user was frustrated why?
21:21
And then now I have to actually take a look at those particular
21:23
log statements. Then in the future, once I've
21:27
uncovered that root cause and I've started to implement a fix,
21:30
I don't want to have to go back to my end user and say, is this
21:32
better now I should be able to like take their their feedback
21:36
and turn it into some sort of unit test.
21:38
And that's where these like evals, you know, automated
21:42
techniques really help. So that that's sort of the
21:46
pipeline that I see. It starts with the feedback that
21:47
collection and proceeds to filtering.
21:50
There is some human debugging and then you're fixing with
21:53
these automated unit tests that you can build.
21:56
And those, those unit tests somehow help us to improve the
21:59
the prompt as well, right? Not to not.
22:01
Not to get repeat the same rut again, not to get into the same
22:05
frustration loop and all that. Yeah, I think there's a variety
22:08
of of solutions out there. So prompt optimization is
22:11
certainly one that prompts and still remain, I think a major
22:17
determining factor of agent quality.
22:20
Two or three years ago, everybody was talking about
22:22
hiring professional prompt engineers.
22:24
A year after that it was, well, you know, this prompt
22:26
engineering thing is brittle. We should look at fine tuning.
22:30
We, yeah, there are other techniques that that may serve
22:33
us better and get us further. It seems like prompt
22:36
engineering, though, has sort of remained front and Center for a
22:39
lot of agent developers. So you're absolutely right.
22:41
In the process of fixing an issue, I may want to do a fair
22:45
bit of prompt engineering to get the agents to make better
22:48
decisions, retrieve information from a knowledge source more
22:52
proactively, issue queries against the web service somewhat
22:57
differently and beyond or, or to take on a different persona.
23:01
These are all valid use cases for prompt engineering.
23:03
And so something that we've been investing in quite a bit in them
23:07
all flow is making that quick and, and effective.
23:10
I've personally spent days prompt engineering Voice
23:15
Assistant that I build at home and it can be agonizing.
23:19
Wouldn't it be great if there was some automated technique
23:21
that I could use to to make that process a little more efficient?
23:25
And that's where, you know, the research folks at Databricks,
23:30
but also just more broadly in academia, I have been pushing
23:33
forward this concept of prompt optimization.
23:36
So given the collection of negative feedback, given some
23:41
information about what isn't working in an agent and given a
23:44
set of prompts, how can I come up with better the versions of
23:48
those prompts that that mitigate and fix those issues?
23:52
And, and can I treat this basically like a optimization
23:56
job and an automated automated solution rather than this human
24:02
in the loop English instruction modification thing where you
24:06
know, I, I don't know if I'm making deterministic progress
24:09
and I might be spending like hours or days coming up with
24:12
better versions of my prompt. You're talking about this where
24:15
the gap optimization comes into play, where you can actually
24:17
optimize your prompt engineering using the the the gap
24:21
optimization as part of the MO Flow library.
24:24
So one of the things as you mentioned that when you actually
24:27
want to automate that the prompt engineering part will be OK.
24:29
Here's a here's a set of prompts.
24:31
Run you through the gap Run you through the gap optimization to
24:34
get the best prompt to see how how you can actually do that.
24:38
One thing that I find interesting when it comes to
24:42
trying to optimize your prompts from when you're out there and
24:47
you're trying to Add all these edge cases, you can quickly get
24:50
into a place where you've got this massive prompt because
24:53
you've added all these edge cases in and then you end up
24:58
destroying your usability of the agent since you tried to cover
25:04
for all this. And I, I know I've seen folks
25:06
that will focus on like dynamically adding in or
25:11
injecting in the context. And that's why we have all
25:15
heard, you know, the rise of the context engineer, because you
25:17
got to throw the right context in at the right time and you
25:21
really want to. Have.
25:23
Not more than you need. It's If it's anything, it's like
25:27
just enough but not too much. Certainly.
25:32
And. And if you're not careful, like
25:33
you said, you end up rewriting a rule based system.
25:37
Except this time you're not writing it in Python or
25:39
TypeScript or or Java or whatever.
25:41
You're writing it in English and you know that that is
25:44
effectively becomes lower quality and more brittle almost
25:48
than just writing it in code. So you're 100% right.
25:51
And that's where some of these optimization techniques and
25:53
optimizer like EPPA frameworks like DS Pi that we try to
25:57
integrate closely with an ML flow can help you avoid that
26:01
because they're, you have to force them to generalize across
26:04
a whole bunch of samples at once.
26:06
So you can't necessarily overfit sample by sample by sample.
26:11
The ultimately that optimizer is not going to be able to produce
26:16
good quality that way across thousands of of test cases.
26:20
So that that's certainly one way to arrive at a prompt that a
26:23
fixes your problem, but B avoids that overfitting and then sort
26:27
of rule based prompting that can quickly really get out of hand.
26:31
So that really resonates. Absolutely.
26:35
Yeah, I, I don't know if we've seen our own kind of like
26:38
customers hit the case where they're just like adding so much
26:41
stuff into the prompt that it's like really polluting the the
26:44
conversation. I think they, they're still more
26:48
in the state of we've thrown something in there and we're
26:50
still trying to figure out like how good is the the agent in
26:53
general. But yeah, also I feel like
26:57
people have stopped short of that because they are cost
26:59
conscious as well. Like they don't want to just
27:01
blow their own kind of token budget, and that is something
27:04
they're tracking very closely. So you don't want to construct
27:06
this like massive prompts that just continuously sucks all your
27:09
budget every single turn. And you were you were talking
27:13
about something before gory that I think is fascinating when it
27:16
comes to the feedback collection and how like you normally need
27:22
to bring in the subject matter experts for that type of a
27:26
thing. When you get that feedback, it's
27:28
great from users if they raise their hand and they're kind
27:32
enough to take the time and tell you what's wrong with it.
27:34
But a lot of times when you're looking at these multi turn
27:37
conversations, you're just going to bring in a subject matter
27:39
expert and say, like, is it doing what we want it to be
27:42
doing? And I've heard a huge complaint
27:46
from folks about how subject matter experts don't want to be
27:51
looking at like Jason, they in fact they can't really be doing
27:55
that. So you have to figure out how
27:57
can I give these folks the data in a way that they can properly
28:02
annotate it and add value to it that isn't just like logs dump.
28:10
100%, absolutely. And that's something that we've
28:14
spent quite a bit of time at data breaks and in the ML flow
28:18
open source community investing in attempting to address that
28:23
challenge of as a domain expert, how do I quickly like recreate
28:28
that same view that that end user had, overlay some extra
28:32
information that helps the domain expert get to the bottom
28:35
of what went wrong and then very quickly provide their expertise.
28:41
And so that's a matter of if I have a chat bot out there and
28:45
user said, hey, this really isn't going well.
28:47
You know, I'm Dimitrios, I'm yelling at the thing.
28:49
I'm frustrated. Whatever.
28:51
I as I then call in my domain expert, Danny or Jules, and
28:55
they're reviewing this interaction.
28:56
They're trying to figure out whether this agents that's
29:00
designed to answer questions about like telecom customers or
29:03
something is what, what is it that is making Demetrios so
29:08
frustrated? They should be able to replay
29:11
that exact chat. They they should have access to
29:13
the turn by turn chat first and foremost, but then often times
29:17
that isn't 100% sufficient. They should also be able to see,
29:23
OK, this particular turn of the chat, the agent went and
29:26
accessed information. Where did it access?
29:28
OK, it accessed information about the user's account rather
29:34
than about like the the line of products.
29:36
And Dimitrios was trying to get a new phone.
29:38
He's not asking about his account or billing related
29:41
questions. So if I can give Danny and Jules
29:43
that information in AUI and they can look through it and very
29:47
quickly go, OK, this thing is fetching the wrong info, then
29:50
that's very, it's very easy for them to come back back and say,
29:53
all right, I found the source of Demetrios's frustration.
29:56
It was, you know, the sort of knowledge retrieval problem.
30:00
And so data bricks, we've invested quite a bit of time
30:02
building this review application where I can collect a bunch of
30:06
these chat sessions and I can send them over to domain experts
30:10
in my organization. They get back that full chat
30:13
view with the overlay of what information was retrieved.
30:16
That way they can dive deeper and they get this awesome
30:19
feedback form where they can provide root cause analysis,
30:23
which then helps the developer go back and fix their agent.
30:26
They can then send that review app to any other member of their
30:29
organization. And we've been really invested
30:33
and actively working on bringing this into the open source
30:35
community as well. The vision is that anybody
30:39
running open source on all flow should be able to spin up
30:42
exactly the same UI, send it to any other member of their
30:45
organization, start collecting these labels.
30:47
And so that's a problem that it's home is deeply, I think,
30:53
personal to all of us on this call at this point.
30:55
Well, maybe it's interesting to talk for a second about how ML
31:00
flow. You traditionally were building
31:02
for data scientists, right? And now do you feel like the
31:07
persona has changed? And if so, who are you building
31:11
for? There's more stakeholders
31:12
involved. What does it look like now?
31:14
A. 100%, but we, we are focused on continuing to serve that
31:19
community of data scientists who want to build awesome high
31:22
quality models to, to make clear that's not going anywhere and is
31:26
is super important to us. At the same time, what we found
31:30
a couple of years ago is there's so much synergy between agent
31:33
developments and model developments.
31:35
You're going to this process of building an initial version of
31:41
an AI application, you're measuring its quality, you're
31:45
iterating on it, you're managing the deployments of that thing.
31:48
You're collecting signals from your end users once you've
31:52
deployed it. Yeah, that's true whether you're
31:54
building an agent or a model. And so we felt like, wow, we
31:56
can, we can take on this problem, we can start addressing
32:00
the needs of agent developers. But as you dive in the next
32:05
order of detail around, OK, how do I measure the quality?
32:08
How do I deploy this thing and and understand whether it's
32:11
operating well in in production, as we've talked, talked about,
32:14
looks a fair bit different. And so that's where you have to
32:17
serve your agent developer who's building that initial version of
32:22
an agent by giving them logs and observability and stack traces
32:25
so they can debug the same way that you serve the model
32:27
developer by giving them a way to track, train, test and valve
32:33
loss during the training process.
32:35
It's you're meeting similar needs, but you're doing it with
32:38
different tooling. And then at the point of I'm
32:41
trying to evaluate and improve quality.
32:44
OK, well, the, the for the model developer, again, they're going
32:48
to take a look at that test loss and they're going to go and do
32:51
feature engineering and things like that.
32:53
But for the agent developer to evaluate quality, they need to
32:58
go talk to a domain expert within their organization, or
33:01
they need to have these awesome out-of-the-box LLM judges that
33:07
can look at the natural language response and determine whether
33:11
it meets the needs of the user who sent the query or not.
33:15
So again, similar objectives, very different sort of sets of
33:19
tools underneath as we're still serving, you know, AI
33:23
developers, model developers and agent developers, but it's
33:26
required us to tap these brand new spaces of agent evaluations
33:31
and things like prompt optimization and build higher
33:34
fidelity feedback collection into the platform.
33:36
So what you'll see now when you go to mlflow.org, you install ML
33:40
Flow and open it up is you get this sort of structured
33:45
experimentation and deployment platform for ML and Gen.
33:50
AI. But you you pick the the
33:52
objective, am I building an AI agent or am I building an ML
33:56
model? We tailor the experience based
33:58
on that selection. I think Corey is absolutely
34:02
right in terms of this is sort of a natural progression that we
34:05
actually serving in two different audiences, right.
34:07
As as you pointed out, the data scientists deal with models
34:10
which are sort of deterministic and the evaluation criteria is,
34:13
is more quantitative. You know, you have, you have a
34:16
certain position that you're actually referring to, you have
34:18
certain recall. Whereas if you look at the
34:21
experimentation, they go through the same process, but the
34:23
evaluation process is quite different, right?
34:26
You have more a non deterministic model where you
34:28
actually now have this set of built in judges and set of
34:31
custom judges that evaluate based on the criteria.
34:35
And I see, I see a similarity between prompt optimization and
34:39
fine tuning a traditional model. They are sort of similar,
34:43
similar tools and similar techniques to use, but it's an
34:46
actual progression. And then we ML Flow currently
34:48
serves both needs. So with your traditional ML
34:51
engineer or ML data scientist, you can use ML Flow for that
34:54
particular need. And if you had the recent more
34:56
engineers, engineering developers with building agents,
35:00
as Corey pointed out, we have a set of techniques and tools and
35:04
judges that allow you to measure that in a way that that meets
35:09
your meets your demands and meets your needs.
35:11
Right, but also very often it's the same person doing both,
35:15
right. So I think a while back we were
35:18
like, you know, who is actually building agents and I'm actually
35:20
curious, Dimitris, like whether you've seen this as well.
35:24
We're like, oh, maybe it's like all these like new class of
35:26
engineers who are building agents and like we're not
35:27
talking about them at all. Then we went on and just talked
35:30
to like a bunch of bunch of people, including a lot of non
35:33
amicable users. And it seems like actually a lot
35:36
of people just shifted from being like, OK, I used to be a
35:38
data scientist or data engineer collaborating on the plastic
35:41
model. And now I've just rebranded into
35:43
like a Jenny I engineer. And so that's, that's still
35:47
probably like, you know, like half the market is like he's
35:50
like former data scientists or data engineers or ML ES
35:53
converted into these Jenny I people.
35:55
So that that's where they often then will come to us and be
35:58
like, hey, we've already used ML flow in the past.
36:00
Like we would love to continue using ML flow.
36:02
But certainly there are these kind of like new like engineers
36:05
who haven't really gone through this process before.
36:08
And then they're like, what is an eval?
36:10
What is? And so then just have an
36:11
education component there of like, you know, hey, we've
36:14
already built all these expertise over like decades of
36:17
how do like how do you iterate quickly to get high quality, you
36:20
know, outputs, right? Let us teach you kind of one of
36:25
the best practices there, so. Well, that the answer is my
36:29
question. That was going through my head
36:31
as you guys were talking, which is like, why not make a
36:34
different product? Why not separate the two?
36:37
But it feels like if it is for almost this very similar
36:43
persona, then it makes sense because they're already familiar
36:47
with MO Flow. They're just going to extend the
36:49
capabilities. Absolutely the case.
36:53
I think that there are, from what I've seen, two temptations
36:58
that product developers, competitors out there have
37:04
they've had to grapple with. The first, like you say, is that
37:07
temptation to I'm going to build a brand new product, throw
37:10
everything out. You know that it's a brand new
37:12
world of Asian developments. We don't need all that ML stuff.
37:15
People can install a totally separate package, hope host a
37:18
separate service, manage the the work entirely separately.
37:24
Our agent developers who, as Danny points out, are often the
37:27
same folks, need a different platform.
37:28
They're going to install something else that we build and
37:31
we're going to have to maintain two different ecosystems that
37:33
that's sort of the, the first temptation, second temptation is
37:36
there's nothing different here. It's all just kind of, you know,
37:39
training jobs and, and, you know, quality measurements for
37:44
agents and, and for models. It all just comes down to
37:46
metrics anyway. You know what?
37:48
What's the difference? And that that has its own
37:51
problems, is that the Asian developer needs traces.
37:53
The model developer, yeah, I mean, they might want them like
37:57
way in the future after they deploy something, but it's not
37:59
even something they're thinking about as they start to build a
38:02
new use case. You know, the Asian developer
38:04
needs LLM as a judge. I don't know why the heck I
38:06
would use that. I thought I was building like a
38:08
traditional machine learning classifier.
38:10
So you definitely have to like understand the workflows are
38:12
different and the tools are different, but the, the
38:15
objectives and the, the persona, the person that's, that's trying
38:19
to meet those objectives remains the same.
38:21
They don't want to, you know, they have to learn and install a
38:24
separate platform and the the business stakeholders or folks
38:29
that are making purchasing and decisions within an organization
38:34
don't want to pay for multiple platforms and set them up and
38:36
figure out how to make them available.
38:39
So we believe we found a pretty good happy middle here.
38:42
You know, the one platform with the right tooling for those
38:46
Asian devs and model developers who were often just data
38:51
scientists. So fair bit of software
38:52
engineers as well. Yeah, the personas really had to
38:56
meet in the middle, in a way, like data scientists had to
39:00
learn how to become better software engineers.
39:03
And then software engineers had to learn how to become better
39:06
data scientists. And so you have this nice like
39:10
Venn diagram of, of both of those skill sets that you need
39:14
and you want to take advantage of.
39:17
And I do see like the agent building very much as lots of
39:23
software skills that we're talking about.
39:26
But being able to fall back on fundamental data science type
39:30
stuff is crucial. It really is, especially as the
39:35
actual software development part of it is increasingly automated.
39:39
The coding assistance 8 basically agents, building
39:43
agents are kind of becoming the de facto pattern for a new use
39:48
case. We talk to customers all the
39:49
time where, you know, despite the fact that Claude code and
39:53
codecs and other tools really only rose to prominence within
39:57
the last year or so, everybody has subscriptions throughout.
40:02
A lot of our customers, the entire dev team is using these
40:05
things to build their agents. And so it's easier than ever for
40:09
them to integrate new tools into their agent or connect a new
40:13
data source or, or even ask, you know, something like Clot or
40:17
Codex to prompt engineer. And so then because the software
40:20
development life cycle is so accelerated.
40:23
They have to fall back even more quickly in their workflow to OK,
40:26
like, what does the quality look like?
40:27
How well is this going? Like those data science skills,
40:30
to your point, become increasingly important earlier.
40:34
You know, I'm not thinking about it 3 or 4 weeks from now after
40:37
I've hand coded an entire agent. I'm thinking about it 3 or 4
40:40
hours from now because the the code is already written because
40:44
I, you know, I asked a coding assistant to do it.
40:48
And so where this is all leading, I believe is a,
40:52
ensuring that folks can do the data science properly
40:55
themselves, but the, B, why can't we enable coding
40:59
assistance to do some of this too?
41:00
So there there is, I think, an appetite out there for, you
41:04
know, if I'm asking Claude or Codex to build a, an awesome
41:08
agent, can I ask it to evaluate it as well?
41:11
Can I give it some example inputs or tell it more about
41:14
what I'm trying to build and let it come up with those inputs?
41:17
Can I then ask it to assess quality and iterate?
41:20
And so as a result, we've tried to make all of them, all flows,
41:24
judges, evaluation capabilities, feedback functionality available
41:28
to coding assistants through MCP.
41:30
So I, I can basically drop ML flow into my coding assistant
41:34
and have it act like the, the agent developer, which includes
41:38
the eval and the more data sciency work.
41:41
And that's something that folks seem to appreciate as well.
41:44
Joe, as you're off mute, I get the feeling you got something to
41:47
say. No, I mean, you know, Corey just
41:50
articulated very well that the the persona is now sort of
41:53
merging along And it's hard to say I'm a data scientist or I'm
41:56
a data engineer or I am an architect given the ability and
42:01
the extensibility of code assistant to say go ahead and
42:04
create a traditional model for me.
42:06
And I'm not even a data scientist, but I'm just saying
42:09
that the Rover might say, I'm not a data scientist, but I
42:11
heard about you can actually create it, a classical
42:13
traditional model and it will do it for me.
42:15
So I don't have to actually have the skill.
42:17
The skills comes in, how do you actually measure the metrics and
42:20
which metrics do you actually want to use for a classical
42:23
model or for an agent? I think that's where the skills
42:26
can be. That's where the background
42:27
knowledge comes in. But the coding assistance can do
42:30
one thing or the other is, as Corey pointed out, you drop in
42:33
an MCP server with code and you can say, go ahead and go ahead
42:36
and write, write me an agent to, to evaluate this particular chat
42:39
pod that I've written and it'll spill one out.
42:42
So I think we're entering, we're entering this new frontier where
42:45
where the personas are very Gray area, but you just need the
42:50
background to be able to say how I'm going to assess and what are
42:53
the things that I need to know to understand to, to get the
42:57
best out of this particular piece of software that the agent
43:00
has actually created. Yeah, I think one kind of cool
43:05
thing that's at least I'm also seeing is that, yes, these
43:10
personas are merging, but also a lot of the innovation seems to
43:13
be coming from the like more data science kind of persona or
43:17
like people who are wall versus net space.
43:20
A couple of examples of this would be I think we've seen on
43:24
our so like very good success with if you want to switch
43:27
between different. So like you know, you're using
43:29
like a gem, a model you want to switch to GBT or or vice versa.
43:33
How do you actually like how do you make that switch?
43:35
Your prompts are tied to your models.
43:37
I think there's the kind of the data science world kind of was
43:40
like, well, this kind of just looks like a model distillation.
43:42
How do I like teach a model to act like a different model?
43:45
And I think we've seen that actually, if you combine, if you
43:48
try to do that and use DS Π in the mix to do that, you get like
43:51
really, really good results. And in fact, you know, we've,
43:55
we've even seen results where you can use like a, a smaller,
43:58
cheaper model and I can really accurately reflect the outputs
44:02
of like a much more expensive model.
44:03
So it's almost a cost optimization, right?
44:04
So that's kind of like 1 innovation.
44:07
I think that's like coming out of like more traditional like
44:09
researcher data science kind of spaces.
44:12
I just want to return some what Corey said, which was I think
44:16
the MLFO team has done a fabulous job of even though
44:19
having two distinct platform, one for the Asian developers and
44:22
Jenna developed and the other one for the for the MLFO
44:25
traditional machine learning scientists, the concepts are not
44:28
really dissimilar, right? Both need experimentation.
44:31
Board needs way to to fine tune, as I said earlier, which one one
44:36
does a prompt normalization, the other does the fine tuning.
44:39
Board needs a set of data sets that you can actually provision
44:41
over, you know, your training data set, you need a evaluation
44:44
data set, you need a test set. We do the same thing in, in, in
44:48
general where you actually have the data sets.
44:50
So I think the concepts are very similar and the underlying tools
44:52
and techniques used are, are, are pervasive across both both
44:57
this particular platforms. So somebody who's familiar with
45:01
with with ML and is coming into Gen.
45:04
AI can actually do that or somebody who's familiar with
45:06
Gen. AI and is going into ML can
45:08
easily transport those pretty good here.
45:10
So I think conceptually and visually and, and through
45:14
workflow, they're not very dissimilar.
45:16
I think the MO Flow team has done a very conscious job in
45:20
making sure that the efforts are flawless from one to the other.
45:24
It's not like you're you're in a completely different world.
45:26
There is some concepts that actually come from there.
45:28
There's a Venn diagram as as as dimension point.
45:31
There's kind of a Venn diagram of set of tools and concepts
45:33
that merge together. Absolutely the case.
45:37
I also wanted to dive deeper into that sort of model
45:41
distillation piece that that Danny was bringing up because I
45:43
think it's a great example of the capability that folks
45:47
resonate with, whether they're coming from a data science
45:50
background doing model development or they're building
45:53
agents. Everybody wants as high quality
45:57
and AI powered application as possible and they want it for
45:59
$0.00. And like obviously you trade off
46:02
cost and quality a fair bit. But the real upshot, I think a
46:06
lot of this work in model distillation is, you know, I can
46:10
build a great agent using frontier models.
46:12
And then over time, once I understand the quality of that
46:15
agent and the factors that determine quality, I can, I can
46:20
ramp that choice of model down to a less intelligent, less
46:25
expensive frontier without sacrificing quality.
46:27
And what I need is a structured way to measure it.
46:31
And, and what I need is a structured way to adjust my
46:34
agent's prompts. Actually, in particular if I'm
46:37
going to change models, so the the task of moving from like a G
46:42
PT52 down to a G PT-5 Mini, for example, or a Gemini Pro to, to
46:48
flash that, that process requires me to understand what
46:53
are the examples of great agent quality?
46:56
What are the examples of negative quality that I'm trying
46:59
to avoid? And how do I tailor my agents
47:04
instructions so that many can understand those instructions
47:08
just as well as as GPT 52 etcetera.
47:13
So if I have both of those pieces, which I'm awful
47:14
provides, I can get to a point of cutting my costs, you know,
47:18
by a factor of 10, even 20X just by virtue of, of switching the
47:23
model. And so that's pretty powerful as
47:25
well. One other area I wanted to hit
47:28
on Danny we talked about a little bit last time we chatted
47:32
was the idea of governance and how folks are really trying to
47:40
figure that out right now, especially with agents being
47:42
able to read and write and you're like, so who we going to
47:49
blame when they right the wrong thing?
47:52
How are we going to keep tabs on this, especially if we're at an
47:56
organization with 100,000 people and they all know are leveraging
48:01
agents? Yeah, I, I think governance is a
48:04
super hot topic for us right now at database.
48:07
And Corey's been actually in a lot of these conversations more
48:10
recently than I have. Obviously.
48:12
I guess database is what we're known for.
48:14
Governance. We, we think about governance a
48:16
lot. I think AI governance kind of
48:18
breaks down into a couple of components here.
48:20
Probably the first thing that people talk to us about is like,
48:23
well, I have like Pi and this and all my different like agents
48:27
and traces, like what do I do? We'll do with that?
48:29
I don't know, how do I kind of ensure only the right people can
48:31
have access to that? And then there's also, I guess
48:35
it if funnels out into well, what data does the agent have
48:39
access to as well, who can use what agents?
48:43
And also kind of really kind of gets into that like AI gateway
48:47
kind of realm as well like well kind of like fall back to
48:50
different models kind of control costs and we're not.
48:53
So I think we're thinking about all that stuff right now on our
48:56
side, what we do have an AI gateway kind of managed
49:00
component. I think Corey, we've now is it
49:03
controversial to say that we are kind of exploring open source
49:08
for. I would go beyond that and say
49:09
you know that MFL has and and. For we have an open source
49:12
gateway, but like kind of like we are, I think the plan is to
49:15
invest substantially more into kind of an open source
49:18
governance kind of gateway solution here. 100%, absolutely.
49:23
And so going a little deeper on on that axis and I think Danny
49:28
just did a fantastic job of discussing the the sort of
49:31
aspects of model access to data and then developer access to the
49:36
data that an agent produces. So going back to that, that
49:38
model access to data problem, we are invested pretty heavily in
49:43
providing cost controls for model access as part of open
49:48
source ML flow, also providing data access controls.
49:51
So I, I've built a bunch of tools I have like this awesome
49:55
tool catalog with MCP. I'm going to hook it up to my
49:58
agents. How do I make sure that, you
50:00
know, the right tools are exposed to the right agents for
50:04
different use cases and that you know, they, you don't have an
50:07
agent accessing data that it shouldn't be allowed to, for
50:09
example, about a completely different customer.
50:12
Like these knowledge bases are, are pretty sensitive.
50:14
And then so walking down tool access is really important.
50:17
But then as I mentioned, just the cost controls.
50:20
I have two teams building agents.
50:23
You know, 11 is building a low QPS chatbot.
50:26
The other one is processing billions of documents every day.
50:31
They require different budgets and blowing the budget out of
50:33
one the project and then sort of exceeding costs can mean that
50:37
the other developers don't get to do anything.
50:39
And so like folks definitely need what we call this AI
50:42
gateway that allows organizations to say, hey, all
50:45
right, for your use case, here's your budget.
50:47
Here are the models you have access to.
50:49
Here are the tools that you have access to.
50:51
If you need something else, you know, come and talk to us, talk
50:53
to your admins and we'll grant that access in a structured
50:57
manner. Otherwise it's, you know, API
51:00
keys flying around, access to data, unregulated it, it becomes
51:04
chaos. And so there's like real risks
51:06
there from a model access to data and costs of governance
51:10
perspective. And I think we could probably
51:12
talk about that all day. And as as Danny mentioned, in
51:16
terms of developer access to agent data, that that second
51:19
consideration, there's some big challenges around that as well.
51:22
PII ensuring that the only privileged eyes can look at that
51:29
certain slices of those traces. That's all super important.
51:34
And that's where Data Brick Community Catalog has been
51:38
crafted and tuned for years to provide governed access to model
51:45
data, agent data, BI data, all of all of basically this is no
51:52
different than any arbitrary set of sensitive data you're
51:56
ingesting into a platform. It's just that it's, it's coming
51:59
out of an agent. You need to govern and manage it
52:01
and share it just as securely. And so that's where like we've
52:05
been building data bricks and all flow on top of Unity Catalog
52:09
in this amazing secure enterprise grade and governed
52:12
data lake on data bricks. You know, yeah, I, I, I have a
52:17
separate life of database where I also think a lot about Pi
52:19
detection and, and how it relates to Unity catalog.
52:23
And I think one thing is similar to what we're seeing here, which
52:27
is the easiest thing that people will start with is they will
52:30
just redact all Pi. And so like nobody has access to
52:32
this or totally lock it down, but rarely is that the actual
52:35
goal. So you need to catalog and
52:38
there's been a lot of investments on like how do you
52:39
do, you know, a back attribute based access control?
52:42
So that and, and what people are trying to do now is like, OK,
52:45
well, if there's a scalable way to manage governance, then I
52:49
would actually open up data as much as possible to the people
52:51
who should have access to it. But in absence of that
52:54
technology, then they'll say, let me just fully lock it down.
52:58
Nobody can see this information. And that's kind of like now
53:00
you're, you're losing a lot of that business value that you
53:03
that you could be kind of getting access to, right?
53:06
I think that's also going to be playing out within the agent
53:09
space too, where if you want to really debug why is an agent
53:13
going wrong or you want to do like downstream, like BI
53:15
insights out of this stuff. Sometimes you are going to need
53:17
to capture some sort of like Pi information.
53:20
Just not everybody has access to that.
53:22
And so at least we're, we've seen some early kind of success
53:25
here is having all this traces get ingested directly into
53:32
adults table. And then now you have access to
53:34
all the governance controls you have with the Unity catalog.
53:37
And now, you know, you can go back to kind of having Open
53:41
Access and sharing and doing all your dash of insights, right?
53:44
Whereas a lot of, I think the other competitors or, or vendor
53:47
solutions out there, their default reaction is like, let's
53:50
give you a tool to just redact all API and have it never landed
53:53
to begin with. To me, that's that's a great
53:55
first step, but it means that you're you're locking away a lot
53:58
of that like super valuable data that you could be really
54:01
analysing and and parsing well. I've talked to people who can't
54:05
use memory tools because the memory tools just by default,
54:10
you know, the open source ones that you get off the shelf now
54:15
expose all of this conversation that's happening with their
54:21
users and and there's Pi all over.
54:25
It's just like littered throughout that even if you
54:28
extract away sensitive data that's obviously sensitive, like
54:32
a Social Security number, for example, inside of the
54:36
enterprise, if you're asking a chat bot to help you craft an
54:41
e-mail to fire someone, that's also sensitive.
54:44
And like, that's much harder to get when you're thinking of it
54:51
at it like in the traditional sense of PII, you know, it's not
54:55
like, oh, that's not my credit card number, that's not my
54:58
Social Security number, but that is very sensitive.
55:00
So you're talking about Dimitri, you're talking about, you know,
55:03
the the what Danny was talking about the PII that can be
55:06
redacted and and that comes up quite a bit in in meet up talks
55:10
about how we actually going to redact the traces.
55:12
But images. I think you we bring up a point
55:14
very important is I'm crafting an e-mail which is going to fire
55:17
someone or I'm crafting an e-mail that's going to promote
55:20
someone that's considered kind of, you know, private
55:23
information. How do you actually read out?
55:26
That, you know, what's the context, so that the context is
55:28
important. How do I actually take a piece
55:30
of trace or a piece of output that has come up and then
55:33
evaluate and categorize that as sensitive or nonsensitive?
55:37
Right. That's again an evaluation
55:39
criteria. I was going to say that actually
55:41
the process of redacting π is actually a hard process.
55:45
So I think a lot of people will start thinking like, let me just
55:47
write something like write on my jacks.
55:49
For people who are a little bit more sophisticated, they might
55:51
use like Microsoft Presidio. But I think everybody I've
55:54
talked to, including ourselves, who, who've tried it is like,
55:58
you just get so many false positives.
55:59
You, you, you miss everything. So you know what, what we end up
56:02
having to build for our own like Pi detection offering is we had
56:05
to build an agency essentially like a whole like kind of
56:08
complex like sequence of LLMS and heuristics and what not to
56:11
kind of actually figure out, you know, with relatively low cost,
56:15
what is PIL, what is not and that we've seen that is much,
56:18
much better than like kind of the off the off the shelf
56:21
solutions out there. And what makes this a lot easier
56:25
is actually having access to the raw data.
56:28
If I can treat this thing as a data frame or as a table and use
56:34
all the brilliance, you know, Sequel and Python And ecosystem
56:38
tools for scrubbing that data, analyzing it, tweaking and
56:42
feeding only subsets that I know are safe into a, a memory
56:45
database. So that when Demetrius tool
56:48
reads from and I know that it's, it's reading information that
56:52
you know is is safe and and secure.
56:54
That's really important. Unfortunately, a lot of
56:57
solutions out there don't actually give you access to the
56:59
raw data. Like typically these traces get
57:01
dumped into some shared like big multi tenant like Postgres
57:06
database. And if you ask for like, hey,
57:08
how can I get like a table or a data frame or the raw records
57:12
that I can do some sort of more complex analysis or find PII and
57:16
redact it. The answer is good luck.
57:19
You can't actually get that data back.
57:21
And so that's where again, this sort of data bricks, Unity
57:27
catalog delta table approach that stores all of your data as
57:32
the beautiful tabular formats and, and, and data frames makes
57:37
this kind of analysis a lot easy, easier.
57:39
You can plug in any tool from the ecosystem.
57:41
You can use Sequel, you can express any sort of
57:44
transformation or Pai redaction task that you'd like.
57:49
And so that's really powerful as well.
57:51
Yeah, I I'm excited for kind of what feature stuff we start
57:53
building with Lake base as well, because like base is kind of
57:57
tailored to this kind of use case that Dimitra you're talking
57:59
about, right. You, you want like a database
58:03
that has fast access and retrieval and whatnot, but is
58:06
well governed, right. And and that is I think kind of
58:09
the promise of lake based on database as well.