EP. 263 Building Agents with Natural Language with guest

**(Note: Spotify listeners can also watch the screen sharing video accompanying the audio. Other podcast platforms offer the audio-only version.)**

In this episode of MongoDB Podcast Live, host Shane McAllister is joined by Sachin Hejip from Dataworkz. Sachin will showcase “Dataworkz Agent Builder” which is built with MongoDB Atlas Vector Search, and demonstrate how it can use Natural Language to create Agents and in turn, automate and simplify the creation of Agentic RAG applications. Sachin will demo the MongoDB Leafy Portal Chatbot Agent, which combines operational data with unstructured data for personalised customer experience and support, built using Dataworkz and MongoDB.

Struggling with millions of unstructured documents, legacy records, or scattered data formats? Discover how AI, Large Language Models (LLMs), and MongoDB are revolutionizing data management in this episode of the MongoDB Podcast.Join host Shane McAllister and the team as they delve into tackling complex data challenges using cutting-edge technology. Learn how MongoDB Atlas Vector Search enables powerful semantic search and Retrieval Augmented Generation (RAG) applications, transforming chaotic information into valuable insights. Explore integrations with popular frameworks like Langchain and Llama Index.Find out how to efficiently process and make sense of your unstructured data, potentially saving significant costs and unlocking new possibilities.Ready to dive deeper?

#MongoDB #AI #LLM #LargeLanguageModels #VectorSearch #AtlasVectorSearch #UnstructuredData #Podcast #DataManagement #Dataworkz #Observability #Developer #BigData #RAG

Episode Transcript

(00:06):
And then welcome to Mongo DB TV Podcast Live.
It's great to have you with us. I'm your host, Shane McAllister,
and I'm on the developer relations team here at Mongo DB.
Today we've got a fantastic showas ever lined up for you.
So joining me is Sachin Egypt from Data Works, and he's here
to show us something really exciting, the Data Works Agent

(00:27):
Builder, a product build with Mongo DB, Atlas, vector search
at its core. Sachin, you're very welcome to
the podcast live. How are you?
Very good. Thank you, Shane.
Thank you so much for having me.It's an honor and pleasure to be
here. Not at all.
It's great to have you on board.I'm very much looking forward to
the content of this show. Umm, I know Data Works have been
on before some of your colleagues, so it's lovely to

(00:49):
get you back. Each time we have Data Works
back on the podcast, there's something new and I'm very much
looking forward to what we're going to go through a little bit
later. As ever, Sachin, I do want for
the benefit of our audience to delve a little bit deeper into
you and your career path to dateand how you got here and how you
got the data work. So tell our audience a little

(01:11):
bit about the pathway that you have to where you are now as
Chief Architect for data. Works.
Sure, certainly. Um, so I've been around in the
industry for 25 years. That's a long time been through
a bunch of transformations in the, in the in the process, but
where we are now, I think it's it's just amazing and a magical
time and a professional timelineto be able to experience all the

(01:33):
transformation that's going on. So how did I get here?
So I was in a corporate job running engineering team about
10 years ago when the first AI wave hit, you know, not the Gen.
AI one, but the one with the AI ML stuff.
And we were doing, yeah, we're doing a bunch of AI acquisitions
and things like that and just didn't feel like we were tapping

(01:53):
enough of the AI potential. So I quit to Co found a startup
in the observability space wherewe were applying AI models for
very large scale anomaly detection of infrastructure and
business metrics. And we deployed that solution
and for a number of high growth unicorns.

(02:15):
And that was a phenomenal experience in terms of how to
how to enable practical adoptionof AI for enterprises AM AIM.
And then when the Gen. AI revolution took off, I wanted
to be a part of it. Data Works was doing some
fantastic work. Again, the thing that I liked
about it was the way Data Works was approaching it was about

(02:35):
very practical adoption of enterprises.
So that appealed to me. And then the other big advantage
I saw Data Works having was thatthey were building it on top of
a very robust data platform. So I have been focusing on RAG
and RAG pipelines at Data Works,figuring out how to make it easy
for enterprises to adopt it. And then last six months, my

(02:58):
focus has been entirely on enterprise agents.
How do we bring the same ease ofuse that we brought to RAG?
How do we bring that to enterprise agents?
How do we make the adoption of enterprise agents practical?
And you know, for enterprises that want to take that
transition and, you know, this opportunity to showcase what
we've done, I really appreciate it.

(03:19):
I'm really excited about where we are, where data works is, and
where this technology trend itself is going.
OK, OK. I love that.
And look, as I alluded to the prior episodes, they with data
works, you know, very practical application.
We're all used to AI and the cool demos that we see from the
key companies in the AI space. But they're just that, they're

(03:41):
demos, they're very cool, they're very interesting, but
they're not, in my mind, very practical.
And I love the fact that Data Works brings that.
For those unfamiliar with Data Works section, before we get
into the meat of today's conversation, could you give us
the elevator pitch for Data Works?
What problems are you solving for your customers?
Oh, sure. Absolutely.
Um, so Databox is a platform forbuilding Genea applications.

(04:06):
So we want it's a full service platform in the sense that we
are involved from the inception,the prototyping, the
experimentation, the creation ofit and then taking it to
production, governing it, operationalizing it, all of it.
And so our goal is to make it simple and timely to adopt these
technologies. We want enterprises to be able

(04:28):
to take advantage of it without having to really invest in
expert AI resources or having tothrow a lot of build new teams
for it, just make it a lot more practical.
The ease of use is very important to us and so is
bringing economies of scale development.
So that's, that's fundamentally whatever we've done in the past
with respect to RAG, a RAG builder, a point and click

(04:48):
solution to build and deploy RAGpipelines and what we're doing
now, that's been our driving principle.
OK, OK. And I know we're going to get
into it in the in the demo lateron.
I had a sneak preview when we were prepping for our live
stream. So I appreciate that.
First of all, a quick shout out to those that are joining us in
the comments, Rio and Kenneth and Gustavo and Carolina and

(05:10):
Bill, etcetera. That's brilliant.
Do keep that up. Let us know where you're tuning
in and joining from. And as we go through our
conversation today with Sachin, if you have any pertinent
questions for what you're seeingon the screen or what we're
talking about, please throw theminto the comments and we'll try
and take care of them either at that time or wrap them up
towards the end of the show. You mentioned kindly on the way

(05:32):
you know, what appealed data works, appeal to you, etcetera,
what they were doing and be verypragmatic, but also the the
database that they had chosen. So, you know, let's not make
this it's all about Mongo DB. Obviously it is that it's our
live stream. But at the same time talk to us
a little bit because you were data works was essentially one
of the really early adopters of our Atlas vector search product

(05:54):
like pre public. You were in the preview
beforehand. And why was that the right
choice for yourselves or what pushed you towards choosing
Mongo for that, you know, early adopters, you're taking a leap
of faith, right, with the product and you guys were on
board very early before we came out with it publicly.
Yep. Oh, yeah, absolutely.
So that's actually an easy question to answer because there

(06:16):
are so many dimensions on which that's really worked out for us.
So all through this journey, Mongo DB and Data Works have
been great partners. We were one of the earliest
startups in the AI Innovators program.
From a technical dimension, I think Mongo DB is great on
multiple fronts. 1 is Data Worksin Mongo DB are very

(06:37):
complementary. So Mongo provides fantastic
infrastructure on which we can build the platform to actually
help enterprises build any applications and take advantage
of that. The fact that Mongo can be the
operational database and also provide the vector database in
the same instance or the same deployment massively simplifies
the architecture for everybody. That's, that's a big, big plus.

(07:02):
And then we've had a number of mutual customers who've deployed
MongoDB and data works together to build some very interesting
business solutions on another dimension.
We've had great collaboration with various teams.
Christian, you've mentioned we've been on your show a few
Times Now. So that's been fantastic for us.
A big thank you to the Mongo DB Partners team for enabling us

(07:22):
with Atlas credits that powers our free tier on dataworks.com
and Mongo DB is our default vector database for powering RAG
applications. And we have fantastic
collaboration with the industry solutions team.
I personally have been working with Prashanth and Angie from
the retail solutions vertical. And in fact what I'm going to
show today is, is a built in partnership with all of us

(07:46):
together over the past few months.
So it's it's overall been fantastic.
The choice of Mongol DB's workedout for us from a technical
perspective, from a customer perspective, which has been
fantastic. Excellent.
We'll look very much Thank you for those kind words.
And yes, a big shout out to the the partner team and industry
team too, because I think they're, you know, they, they
love to get deep down involved with our clients and our

(08:10):
customers and our startups to tomake it a success.
You know, this whole build together notion and really is
part of what those teams do because it's you know, but they
want they want it to be a success.
They want, you know, to make sure that we create the easiest
path for that success as well too.
So that's it's thank you for shouting out to those teams.

(08:30):
I'm sure Prashant and the otherswould be very keen to hear that
as well. I appreciate that.
We're going to dig the title today, obviously was build
agents with natural language andwe're going to dig deep into
that now. And I know you've got some
slides really as by way of reference architecture to share,
but just before and for our audience who you know, they're

(08:51):
familiar maybe if they've tuned into a couple of these episodes.
We have done a lot of rag retrieval augmented generation
episodes. But agentic rag Sachin, what's
meant by that and why is it important in in the world of AI
applications today? Oh yeah.
I think the evolution of Rag to Agentic is a fantastic story

(09:12):
that talks about how this technology is progressing and
how as as it starts to solve problems, it sparks imagination
and people start to push it, push its boundaries.
And then the next set of evolution in the technology
starts to come in and it's it's a great example of that.
Shane, if you're OK, I do have acouple of slides that leads up

(09:32):
to this. Sure.
So we could switch to the slidesright now and I could walk
through that and. That makes sense.
Picture paints 1000 words as they say, Sachin.
So let's get that up on the screen because I think again,
for our audience who may or may not be familiar with this space,
I think it's, you know, we know we've got a technical audience,
a developer focused audience. So if we can share that the

(09:53):
slides and and kind of go through some of the
architectures before we head towards our our demo that we're
going to see you later as well. All right, so I just shared my
screen, Shane, I think that's looking all right to me.
Can you that's. Good.
Yeah, if you want to go into theslideshow mode, it should be
fine. Yeah, zoom wise it seems good.
Should anybody in the of our viewers have any issues, do

(10:14):
shout out in the comments, but Ithink that's legible and pretty
easy to read right away. OK.
That's, that's great. So, so Shane, you mentioned RAG
or retrieval admin generation. I thought we could start with
level setting and just talking about RAG and then build up from
there to identity RAG. So it will augmented generation

(10:35):
has been like the go to technique for enterprises to
leverage the power of LLMS over proprietary and private data.
It's a, it's a pretty simple effective architecture and it's,
it's been transformational in terms of information retrieval
for both users and employees. The picture on the left here on
the slide actually is a typical implementation or architecture

(10:58):
for what I'm going to call conventional RAG.
And it is characterized by a sequence of steps that are
executed. And so when a user asks a
question, a process of vectorization, the question is
performed as one step. And then you fetch relevant
chunks of data from like a vector database, like a Mongo

(11:18):
Atlas vector search. And then you put that into a
prompt for an LLM to answer. But the constraint the LLM has
is that the user's question can only be answered from the
context that's in the prompt. So, so that works well actually
for a lot of situations where you are dealing with relatively
static data or data that doesn'tchange very frequently.

(11:40):
It's largely text or images or, you know, PowerPoints or web
pages. That kind of stuff works really
well. But where it starts to struggle
is when users questions require accessing multiple data sources
or accessing dynamic data, or when the user asks a more

(12:02):
complex or compound question. That's when conventional rag
really starts to struggle. So let's take a look at the
example on the screen here on the left.
So the user asks the question, cancel the order with White Sox.
Now that's a very specific question about a specific order
in a conventional RAG system that has access to perhaps only
the policy document in it. The best that a conventional RAG

(12:24):
system can do is to answer the question by saying, hey, here's
the cancellation policy. And then now it's up to the user
to kind of apply it to their situation.
You can't really get a situationaware answer from the LLM
because the system doesn't actually have access to the
order to be able to apply the the rules onto the order.

(12:45):
So so how do we fix that so agentic track can help us with
this? So if we in we add agent, sorry
for that. If we add an agent to a RAG
architecture, what does that give us?
Firstly, we are not going to trans, we are not changing RAG.
It's still the same concept as before.
We are still going to populate acontext that can leverage to

(13:06):
answer the question. But the the quality of a RAGS
response depends very heavily onthe the completeness and the
accuracy of the context that is provided to them.
So what an agent gives us is that it can iteratively build
out the context. It can have access to data
sources apart from what is typically provided to a

(13:27):
conventional Rack system. So an agent can have access to
databases, it can have access toweb services and many other data
sources. So when a user asks a question,
an agent can basically understand the users intent and
decompose that question into subqueries and then tap into the
various data sources to populatethe context.
So that gives you a context that's significantly richer than

(13:50):
what you would get with conventional Rack.
So now and again, do a much better answer, provide a much
better answer, much more complete answer than it could
before. So let's apply it back to the
previous example, right? So the user asks the same
question, cancel the order with white socks.
But let's say now we have an agent system and the agent has
access to the same policy document as before, but it also

(14:14):
has access to say the operational orders collection.
It says it's in Mongo DB, right?And we have access to, let's say
we have access to the real time status of the order that comes
from REST service. When the agent looks at the
query, it basically gets the cancellation policy like the

(14:36):
conventional track system. But it also realizes that it
would help to add the order information into the context as
well as add the real time statusinformation into the context.
Now the LLM has a lot more information and it can provide a
much more situationally aware answer where it can say, yes,
this order can, can be cancelledor no, this order cannot be

(14:58):
cancelled because it's already been shipped.
So this allows us to create a significantly better, you know,
RAG systems. And what Agentic basically
allows us to do is to create transformational experiences by
extending RAG to be personalizedto the user and to be
situationally aware because it has access to operational data.

(15:19):
Which is incredibly valuable both on both sides of the table
here. So you know, obviously for the
company involved, having access to the operational data as well
as the, you know, the unstructured data, the policies,
etcetera is is really key because you know that does that
make sure you have everything tohand to involve, to be involved

(15:41):
in that decision making process as to whether in this example
the order can be cancelled. But obviously it's incredibly
valuable as a customer. We've all been there, Sachin,
right? We're trying to do these types
of things and the level of frustration that you have when
you you go down a dead end maybewith some of the chat bots that
are taking you somewhere. And then you get the human and

(16:02):
the humans asking you to Add allthe information you just gave
the chat bot and you know, and then the human says I'll be back
to you, etcetera. So it's an incredibly valuable
process that we're automating for the, in my perspective, the
good of both sides of the table,the customer and also the
business, right? Oh, absolutely.
So we have an. I can give you an example from

(16:24):
the customer who we are helping with this.
It's it's a customer in the mobility space.
And one of the things that they wanted to do was, you know,
improve the meantime to resolution.
When one of the customers calls into the customer service agent,
right? And they, they provide
troubleshooting, they'll get like some error code and they

(16:44):
can basically ask, you know, canyou help me fix this, resolve
this problem. And the, the customer service
agent has a RAG application now that allows them to basically
ask the exact same question thatthe customer's asking into it
and then get a response, you know, from all their internal
documentation and all the troubleshooting guides, it can
pull exactly what the customer wants.

(17:05):
The only caveat is that you haveto 1st know what product the
customer has already ordered from you.
So that's an additional step that needs to happen because the
conventional RAG system cannot cannot connect to that.
But the moment you plug in an agentic rack system, what that
can do is fetch the exact products that a customer has
ordered, then ask a question forthose specific troubleshooting

(17:27):
requests and provide a response back.
If the the result is that you need to order a replacement
part, even the part numbers can be fetched based on what you've
already gotten. So the meantime to the resolve
the customer satisfaction, you know to your example Shane it
just it's the number of things they can achieve with this is
significant. Yeah, it's great, right?
And I know we're going to go into more.

(17:48):
I just want to bring some of thecomments on screen.
So Kenneth came to the right live stream at the right times.
He says it fits right in with the class he's taking right now.
Database integrations with othersystems.
And so you know where MongoDB sits here is the operation of
the transactional database, which is which is superb.
Nathan is coming to the realization that many customer

(18:10):
service bots aren't agentic rag,just conventional rag perhaps
for most. So which leads me really to the
next next step. So this obviously makes sense.
If you put this flow chart in front of any potential customer,
Sachin, they will go, yeah, thisis going to help my business.
This is going to help my customers.
But the developers in that company are going to go, what's

(18:31):
this mean to me? Well, how much work is involved
here? So talk to us a little bit about
how we've ended up here with, you know, essentially data work
day to Grag as a service that we're going to be showing.
Yeah, absolutely. So 2025 is supposed to be the
the year for AI agents and we think it's going to be the year
for agentic rag as well. So every time we talk to

(18:53):
customers, agentics parts their imagination.
You know, we've had a couple of very interesting comments
already when they saw the the flow chart.
A lot of the enterprises or the,the organizations, they, they
start with a DIY approach. This they would like to start
the DIY approach. And that's, that's fine because
you need to build up some experience.
But the, we quickly realized that this is the complexity of

(19:18):
doing this and the iterative nature of it makes it expensive
and time consuming. And the issue with the time it
takes to put together a solutionlike this is that you dilute
your competitive edge because it's a very transformative
technology and the story you have to adopt it the further
behind you go. And we've been in the business
of trying to make adoption of Gen.

(19:38):
AI applications significantly simpler, make it make it timely
to adopt. And our goal has been like I
said earlier has been to kind ofenable the adoption of these
technologies with ease of use and economies of skill in mind.
So we would like enterprises to be able to do this without
having to invest in expert AI resources or build large teams

(19:59):
to do this. We would we have We were invited
to an RFP actually, and one of the.
The key points there the the company that invited us was
making was that the question was, you know, do you have, do
you want technical resources or not technical resources to
access to build engineer applications.
The key aspect of the answer wasthat the closer we can get to

(20:22):
the business users building this, the the better it would be
because that would be the fast. I can't see that.
I mean, it's sometimes we've allbeen there, you know, that, that
maybe perhaps the business use case isn't translated, you know,
completely or complexity enough down for developers and and kind
of there's always a disconnect in the middle there somewhere.

(20:42):
So I can see that, that if if the business use case and those
folks putting that together can actually use or manipulate the
system. That makes a ton of sense,
right? Right.
And, and with the, the existing customers we have, we've, uh,
we've seen their existing technical resources dive into
this because they understand it so much better.
And they can, they can create transformative experiences and

(21:04):
workflows without having to actually go and learn a ton of
stuff. They already know their domain,
they know their business and they can leverage that.
So to enable this, we'll have anidentical service offering,
which is based on three pillars.Now I'm going to provide an
overview because I think showingthis would make a lot more
sense. So I'm not going to go too deep
into each of these, but I'll talk about them so we have an

(21:26):
idea and then I can refer to this when I'm actually showing
it to you. Brilliant.
So, yeah, so the first part is the agent hub, which is all
things agents in data works. So that's where you have, that's
where you build agents and that's where you build these
sophisticated experiences and then you can take them to
production and you can operationalize them.

(21:48):
So is there a playground there, Sachin?
Is there some way you can load sample data etcetera and kind of
get your proof of concept up andrunning before you take it
further? Definitely, yes, absolutely.
So Data Works itself is built ona data studio that allows you to
connect to a real number of datasources, ingest data, transform

(22:09):
it, and it's optimized for AI pipelines.
And when you connect to these data sources, you can make them
available to agents. Now the way you do that is you
kind of interface the system you're connecting through
something called a tool. So the agent can talk to that
system via this tool interface. And if you, if you, if you think

(22:30):
about this, we're talking about creating these workflows that
span the enterprise. So we are pulling in data where
these tools from like different departments and different teams
and kind of bring it all together in an agent.
So having a centralized management capability where you
discover these tools and you create them and you manage them
and you audit them would make sense.
So that's, that's what we call the tools repository.

(22:51):
And then finally in ease of use is our goal.
We've kind of built this agent builder, which we kind of spoke
about very briefly in the introduction and I'll, I'll
spend a lot more time on, which is our no code solution for
building. Not only is it no code, we've
kind of gone down the part whereyou can instruct it in just

(23:13):
plain English and I'll show examples of that.
And finally, the third part of this solution is our
conventional rag builder. It's a no code system again,
where you point and click. And we showed the rag builder on
your show, I think last year, Shane.
Yes, indeed. Remember that one?
Yeah, yeah. So so the the Rack Builder

(23:36):
allows you to build high qualityrack pipelines with just a few
clicks. We've had customers build these
things and deploy them in a matter of hours.
No need for any expert resources, no need for any
additional teams. And what this enables is that as
a tool, you can plug this into an agent and that gives you
access to unstructured data, whether that's you know from a

(23:56):
PowerPoint or APDF or images or you know, web pages or whatever.
And they can be on SharePoint onyour wiki, they can be on Google
Drive, and they can all be pulled into this.
OK, excellent. Sorry.
As a service offering. Yeah.
So all the tools at the disposalof the developers, depending on
how they want to use it on leverage, ETC, as well too.

(24:18):
And you're consistently trying to make it easier.
So as you said, it's, it's now not just no code, it's now
natural language. So talk to us a little bit about
that. Maybe it's part of the demo,
right? So talk to us a little bit about
how, you know, kind of that's going to help developers, how
easy it is to create those agents with natural language.
Sure. I would, I would love to do

(24:39):
that. So before I talk about the agent
builder, I just want to spend a slide and then and then we'll
we'll start showing stuff and talk the demo and the agent
builder. I wanted to speak about how we
are approaching the whole problem with enterprise agents.
And so that's going to eventually build into you know
how we are putting it all together using the agent
builder. So since we are building

(25:02):
enterprise agents, we aren't targeting general purpose
conversational agents. And the difference is
fundamentally that enterprise agents are purpose built for a
specific set of tasks. You want users to be able to
achieve something, but also equally importantly, you want
the agent to be able to identifywhen a user is asking you to do
something that it is not meant to do.

(25:23):
Because you don't want an enterprise agent to go off the
rails or be completely uncontrolled.
So it's very, very important that you know you have necessary
guard rails. The other issue that we should
be very aware of is that the, the tools repository has, you
know, has access to sensitive data all across the.

(25:43):
So if an agent has access to allof this, then the attack vector
and agent becomes very broad. So you want to contain that to
the extent that it makes sense. So, so any agent tech
architecture needs to kind of handle these.
There are other aspects to it, but these are some very
interesting issues that are worth fixing.
So, so we've, we've approached this problem in a particular

(26:04):
way. The the way we've thought of
this is that an agent solves a particular problem space and we
we partition the problem space into non overlapping segments
that we call scenarios. OK when a user asks a question.
Umm, based on the user's intent in the question, the agent
classifieds that into one of those scenarios and answers them

(26:25):
from within that scenario. I'll try and.
And so if it doesn't actually get classified into any of those
scenarios, then the users askingus something that the agent is
not meant to do, and then we canjust and respond.
So let me illustrate that with an example.
Let's say we have an agent that enables a customer to do a bunch
of things. And one set of those things is

(26:47):
to be able to query their profile and to update some parts
of their profile. Maybe they can update their own
shipping address, maybe they canupdate their preferred payment
methods. And so anything related to
profile updates can go into a scenario called profile.
And then if agent also allows you to provide feedback to, you
know, past orders and all that, then that can be in in a

(27:08):
separate scenario called feedback.
So now the agent can route the user's questions to the right
scenario. And anytime the user asks a
question that does not fall intoone of these, we can have the
agent do something specific. OK, OK.
That makes sense. Yeah, Yeah.
Because, yeah, you're still trying to have that good
customer experience, but you need to make sure you've got the

(27:29):
right guardrails in place. You're dealing with a lot of
customer and sensitive data. So yeah, it's a complete, it's a
complete system you've thought of.
You've thought of most of the pitfalls anyway.
Right. And, and the thing that we do is
we, we classify the intent, not the actual text.
So, so even if the users like typed a follow up piece of text
that's, you know, that could be in isolation be assigned to any

(27:52):
scenario or none at all in the context of the conversation, the
agent would classify it to the right scenario.
So you, so you don't, you don't have a rigid conversation that
you're having. You can still have a very normal
conversational flow in the in the in the agent, but it can
still go to the right scenario. The other thing that scenarios
get us is that if you, if we have a profile scenario, it just

(28:13):
needs access to a couple of tools that help it fetch
profiles and update profile feeds.
We really don't need access to, you know, orders order history,
order status or feedback or anything else.
So, and a scenario for us can bea curated subset of tools from
the tools repository. So that gives us a very
diminished cast radius for when a scenario has been assigned for

(28:36):
a user's question, even if the agent is getting confused,
you're limited to the number of tools that are within the within
the scenario. OK, the yeah.
And there's a lot of research that you know, that shows that
the lesser number of tools you have that an agent needs to
create a plan with, the better the OR more reliable the plan

(28:56):
is. So we're trying to optimize the
ability for an agent to reason well as well.
There are a few more aspects to enterprise security access that
make this a little bit more interesting, but I think there
are some of the higher level concepts that that we will build
on when we when we do the demo. Perfect, perfect.
No, that's that's very clear. And again, should anyone have

(29:17):
any questions who's joining us on YouTube or LinkedIn to throw
them into the comments and we'lltry and take care of them as we
have done already and that will be great.
Thanks, Sachin. OK, so I think Shane, now's a
good time to probably jump into the demo.
Definitely. We know our audience love the
demo. We've got the slides, we have
the background there. We can see that, you know, the
architecture and the use case isvery clear, but there's really

(29:40):
nothing like seen it in action, Sachin.
So I'm very much looking forwardto this.
And I'm, I'm very excited to show this, this, this is, this
is the part I was waiting for aswell.
So, so we've built, like I mentioned earlier, we've, we've
built this in collaboration withthe, the Mongo DB retail
solutions team. And all of our work is available

(30:01):
as blogs and videos and and it'll also become available in
Mongo DB's fantastic solutions library resource.
That's that's on it's way. But the blog and the video is
already out and we can share thelinks for all of that.
There is also a GitHub, a GitHubrepo that has very detailed
instructions on for for anybody who wants to try to do it by

(30:22):
themselves. In fact, I would urge you to
write it out for yourself and incase you get any issues, reach
out to us on Slack or on GitHub itself.
So let me talk to you about whatwe are going to build.
So I'm going to build this agentfrom scratch, all of it on the
call here. This is going to be the back end

(30:43):
of the agent. This actually does that agentic
workflow using Agentic Rack. So the MongoDB Detail Solutions
team has a showcase MongoDB ecommerce site called that
highlights the application of MongoDB technology to that
specific domain. It's fantastic actually.
And then there was this idea to showcase a customer support chat

(31:07):
bot on that ecommerce store. So, so the typical how can I
help you bought on the bottom right corner.
So that lets you talk to it and this bot can do two things for
you. One is users can ask it
questions about store policies, you know, very, very similar to
what you could do with a rag application.
Basically, you ask things like, you know, how long the shipping

(31:28):
take, what kind of products cannot be returned, that kind of
stuff. So we can do that.
And the other thing that we can do is we can or users can do is
that they can ask specific questions about their orders,
You know, things like what is the status of my order?
Can I cancel it, that kind of stuff.
All right, so, so let's let's assume that I am the I am the

(31:48):
person who's being tasked with building the back end of this
agent in data works. And I have a product manager who
comes to me and says, hey, here's the business brief about
what we're going to build. It's it's text from the product
manager that says, writes the agent and it says that it's
going to be a customer support agent.
You know, it helps the portal customers.
There's some text about, you know, what kind of behaviour you

(32:08):
want there. There's specific instructions
here about what users can do with the agent.
And then there's some text care about what the agent should do
if the users, you know, asking stuff that the agent can't do
right. So that's good.
And then the PM tells me that wehave access to two things.
One is APDF, that is the policies document that we share

(32:32):
with everybody on, on our web onour portal.
And that's you want to answer questions from there.
So everything is consistent thatthat's fine.
And then I have access to the the operational database for
orders, which is a Mongo collection, MongoDB collection.
And that will allow us to basically understand when a user

(32:54):
asks about a specific order, we can fetch that from there.
All right, so with this background, let's let's build
this from scratch. All of it.
And I have a working version in the next 10 or 15.
Minutes and using that exact brief there, Sachin, that you've
just gone. That exact brief, perfect,
perfect. We'll build exactly what the the
product manager wanted. That usually doesn't always

(33:14):
happen, but I'm looking forward to seeing it in this case.
For those tuning in, I've sharedthe blog links and also the
video links in the comments as well too on YouTube and
LinkedIn, so you can check thoseout and hopefully maybe not.
Now for watching the other video, stay tuned to the live
stream, but but afterwards, please, please check that out.
I'll put the GitHub repo up shortly as well too.

(33:35):
Thank you. So, so I urge everyone to come
to Airworks and sign up for a free account and give this all a
spin and you can basically just try it for free.
I already have an account, so I'm just going to log in and I
let's just confirm that the Zoomlevel's good as soon as we get
in here, that that looks OK, right?
That looks good to me. Yeah, it's coming through.

(33:55):
It's very light. No problem.
Perfect. OK so so here's here's what
we're gonna do. We have two data sources.
So we want our agent to connect to a RAG application so users
questions can be answered. So that's one.
And then we want to connect it to Mongo, so that's the other
one. So the way I figure we've got to
create 2 tools and one of those tools needs to be connected to

(34:16):
Mongo and one of those tools needs a RAG application.
So let's start with the RAG application first, because
that's that'll take us a minute or so and then we can build the
other one. So that's the policies etcetera
that you want that's. The policies 1.
Exactly. Yeah, that's the policies 1.
All right, so, OK and all right,so this is this is Dataworks.
This is where you can build, this is where you can launch the

(34:38):
the Rag builder. There's one piece of prep I have
done here for the Rag builder, which is that I have plugged in
Dataworks is open AIAPI key and configure a couple of LLMS
already. So if you are doing this, you
would bring your own and you would you would put that in.
So let's. Do you support a broad range of
LLMS in dataworks selection? Yes, we do.

(35:00):
We support Bedrock, we support Open AI and a bunch of others.
So yeah, so you could, you couldbring your own from any number
of them. Perfect.
Yeah. And, and I have, I don't have
anything in this account, so everything's going to be pretty
much empty. But if you had tag applications,
you'd see a list of them here. But now that we don't have any,
we fix that. We create the first one.

(35:22):
So let's say if this is B fee total policy, which is and seems
like a reasonable name for it, and let's pick GPD 4.
We could pick any, any one of these.
I'm going to pick that one. So like I was mentioning in the
slide, data works is built on the foundation of a data studio

(35:42):
that allows you to bring data from many different sources,
many different formats into the product and transform it into
all kinds of things in so So if I did that, I could use data set
that's already been transformed,but I'm not going to do that.
I'm just going to pick up the PDF from my local disk and this
is the no code rack builder. I'm pretty much just going to

(36:04):
pick the defaults. Shane, I'm just going to walk
through all the screens and justexcept for providing the PDF
file, everything's just going tobe default because that's going
to be just fine for us. But I will point out a couple of
interesting things as we go through it.
So, you know, we realize what ispossible.
So let's continue here. I know it's APDF.
I'm just going to pick it up from my computer here.

(36:25):
Sachin, you might go one zoom level more on the on the
platform again if you can. On the on the browser.
Yes, yeah, yeah. OK, I'll do that as soon as get
there. Perfect.
Yeah. Once you're back in, it's.
Yeah, some of the screens a little bit.
You're squinting to read some ofit.
And I know our audience will. Yeah.
This one, this one's the. This one's always.
Windows 12 Choose, there's no problem.

(36:46):
Yeah, yeah. I can't.
I can't help that so, but this one I can.
That's perfect, perfect. Excellent product.
Thank you. All right, so I got the PDF, let
me upload that and that's done. So let's save it.
So it's it's chosen a default location to upload that it's in
S33 account data works, gets some S3 space allocated to it

(37:08):
automatically. That's perfectly, perfectly
fine. I don't need to change that.
So we'll just move forward. And when we move forward, you
will see an additional choice here.
I'm not going to change anything.
I'm just going to go next, but I'll just talk about what this
is. This allows us to plug in custom
data flows for ingestion, like, you know, chunking and

(37:29):
vectorization. And then where do you want to
store it? So you could pick your own
embedding model, for example. There are a number of different
options for strategies and various parameters and how you
control that. Ultimately this is.
But in the default case, like I said earlier, it is going to go
into Mongo, which is just fine. So no changes here, we just use
this go forward to. Pick the table.

(37:49):
But people have the choice to choose their own depending on
their particular application andwhat embedding models suit that
application more right? So.
Right. Yeah.
So if you have a embedding modelthat is particularly tuned for a
domain like, you know, financialservices, you could use that.
We have a general purpose document, so a general purpose
embedding models, it's just fine.
Perfect. This is where you configure your

(38:11):
retrieval. The conventional drag sequential
pipeline that I talked about in the first slide.
Of course, this is a little bit more complex because there are
additional best practices that have been put into this and
there are more things you could do with it.
I'm not going to change it, but in the default you get a hybrid
retriever, which basically meansthat it will hit both a lexical

(38:34):
index as well as a semantic index.
The semantic index is Mongo withcosine similarity with .55 as
the threshold, and these are allgood defaults, so we just let
this finish. So that's.
I'm a big fan of defaults, but Ilike the.
I like the fact that you can change them there if you wanted
to. So yeah, that's perfect.
Yeah, yeah, make the default easy, but make the hard

(38:57):
possible, right? Yes, exactly.
Exactly. Yeah.
So, so this is, this is, this has gotten started.
It's going to pull in the data, it's going to do its thing.
It's going to provision a RAG application when it's done.
So that's great. While it's gonna take a minute
or so. So first, these things.
Do let's just. Do the next step which is
connecting to creating a tool toconnect with mongo, right?
So if I go here to configuration, I get the option

(39:21):
to connect to a number of different systems.
There's a pretty long list here.I have a option to connect to
Mongo and I have already configured it in the.
If you go to the GitHub the GitHub report you'll see that
Dataworks is hosted A publicly accessible read only mongo DB
instance. Perfect or get.

(39:42):
That to make the whole demo easier.
Yeah, yeah, OK yeah, perfect. Thank you.
So you can just use that, but there are also instructions on
how to bring your own Mongo DB if you would prefer to do that.
So you could go either way. I'm just using the one that
Dataworks is hosted. I'm just going to kind of so
it's fine, we'll configure it. OK, All right, so now I have a
connection to a Mongo database. I'm just going to copy the ID

(40:05):
here. Now the next thing I need to do
is basically I want the agent tobe able to send an order ID and
then get an order object back with all the details in it from
the operational DB. So I'm going to create a tool
that does exactly that. So this is this is the agent
hub. This is this is where we
launched the agent builder and also the tools repository.

(40:26):
And Shane, I want to. I'm going to take this
opportunity while I'm here with you on your show to announce
that we are launching Agent Hub,the Agent Builder in preview and
we would love for the audience to give it a try and give us
feedback. Thank you.
We've done something fantastic here and we're really excited

(40:49):
about and we hope that everybodywill find it interesting.
It is in preview, so you might have some issues.
Reach out to us. We have a stack community.
We are actively building a new version or the next version of
it that's going to come out in the next couple of weeks.
But you can already do a lot with what is there, and I'll
show you exactly what is possible.
Perfect. Listen, that's I'm thrilled you
announced it here and that's great.

(41:11):
I hope hopefully you get some users to try it out and get some
valuable feedback as well too. That's brilliant.
That would be perfect. All right, so I don't have
anything right now here, but on the left is where I would see my
agents, any agents I've created,and on the right is where I
would see any tools that I created.
Let's start with creating 1. So like I was mentioning

(41:31):
earlier, tools are the mechanismby which we plug in external
data sources or make external data sources available to an
agent. And that could be, it could be
databases, it could be REST API web services, it could be rag
applications, or it could just be utilities and all kinds of
things that you can plug in. We are shipping with a bunch of

(41:54):
pre canned integrations and there's a bunch more that are
going to come out in just like aweek or so.
So, but in this case, yeah. So in this case, we are
connecting to Mongo. So let's let's start with that.
And the way you deal with a toolis it's like, Shane, if I were
to explain to you how you, how you should use something, right?

(42:15):
So I would give it a name and then I would, I would describe
it to you. If I give you a toolbox with a
bunch of tools you aren't familiar with, I would tell you,
I would point out a tool and say, hey, this is this is this
tool and it does this, right? So that's exactly the thing that
we need to do with an agent. We need to give it, give the
tool a name, a name that kind ofindicates what it does because
it interprets that. So in this case, this is a good
name because we are getting an order object from the Leafy.

(42:38):
For the Leafy portal, we need togive it a description so the
agent understands what this tooldoes, and then we need to kind
of tell it what inputs it takes.In our case, it takes an order
ID, it's of a type string, and we give it a description.
We say that it is the order ID of the user wants to get that.
Sure. Enough.
Right. And what this does is that if we

(43:00):
if we throw a bunch of tools at the the agent, it can basically
understand how to chain them together, how to use one's
output into the next input in order to create a solve a
particular problem. That's that's the goal here.
OK. OK.
So and others would, would, would this be fair to say kind
of others would call these kind of reasoning engine type logic

(43:22):
etcetera as well too, right. It's understanding which is the
right tool for the for the job with the details those extra
details parameters that you're giving it.
Exactly. And yeah, that's the reasoning
part where it understands this and then puts a plan together to
execute the to solve the problem.
That's exactly. That's exactly right.
That's the reasoning part. So every, every tool is gonna
have a similar concept where youdescribe it this way and we've

(43:43):
not got the input, we need to give the output.
Now the output in this case is alittle a little bit more
elaborate because we are going to plug in the output from the
operational database as it is tothe agent.
No changes, no transformations, no simplification.
We are just going to point it tothe orders database and we are
going to tell it that this is what the order object looks

(44:05):
like. And then the agent just going to
figure it out from there. So, so I'm basically describing
it in it's, it's like any object.
You basically say this is the name of the object.
It has these fields, fields havethese types.
And then again, there's a description.
If you have more complex fields,for instance, this is a list of
status history. You know, as the order went

(44:25):
through delivery states, new items got added here every time
the status changed. So this has a status and
timestamp and so on and so forth.
And this describes the output. So now given this, our agent
knows that this is what it's called, this is what's meant
for, this is what I should pass it, and this is what I will get
back from it. OK, right, very clear.
And the only thing that's left at this point is actually tying

(44:50):
it to a physical data source. So we connect it to the Mongo
connection that we had made earlier.
This is the same ID under add copy.
And then we got to specify what database you want to connect it
to. This is the name of the
database, what collection we want to connect it to.
This is the name of the collection orders.
And then because the thing that we're doing here is we are
saying given an order Idi want an order object back, we need a

(45:13):
query predicate for Mongo. This is the the language in
which Mongo understands a query predicate.
So we plug that in. The only interesting thing here
is that this is a placeholder for what the agent will
eventually pass in as the order ID.
So if we get an order ID 123 when we actually make the query,
we will replace this with 123. Perfect.
Very clear. OK, OK all.

(45:35):
Right. So let's let's say this and we
actually we've done creating ourfirst tool for Mongo.
We need to create one more whichis the one for the application.
But so let's go and check on what happened there.
We gave it enough time so it is done.
Job It has it. Is done, so let's take it for a
spin. We'll ask it a couple of
questions and see whether it's working fine.

(45:56):
So this is the interface within data works to try and use your
AG applications. There is an API could integrate
this into like a search widget or whatever you want, wherever
you want it within it makes sense.
But let's ask a simple question so we don't have to read through
a lot of text. All right, so that's start and
suite. It says yes, you can cancel an

(46:17):
order and then it points us to the relevant section in the
document from where it. The reference is there, the
reference tracker if you need itanymore.
Exactly perfect. And let's try another one that's
also that also should be a shortanswer.
All right, so that's great. So it looks like it's working
fine. The one thing that I wanted to

(46:37):
show here is that we have this probe that allows you to get
visibility into what's actually happening when a conventional
rack system executes. Yeah, right.
Nice. That's great visibility into it.
Perfect. Yes.
So every step of what it did, and there are a few more steps
than my slide because you know, we do a few more optimizations

(46:58):
on it. But at a high level, this is the
vectorization of the question. This is the the query to the
lexical search. This is the query to Mongo.
So you see the results of the Mongo, the results from the
Mongo database for the vector search.
And then finally the context is populated into a into a prompt

(47:19):
for the LLM and we get a response from the LLMS and this
is the response from the LLM. Very nice.
The level of I would imagine with your clients when you're
showcasing this such and it gives a level of of of
confidence and comfort in what'sbeing built in the back end.
If if they're going to, you know, leverage agents as they
should do to do, you know, thesetasks and certain amount of

(47:42):
reasoning inside of these tasks.They want to be able to have, as
you call it there, the probe to see under the hood exactly
that's what's going on, right? Exactly, right.
So you're absolutely right. So the same visibility that we
are providing to RAG applications is will be
available in agents so that you can see exactly what the agents
doing, what is the reasoning or tools are provided to it.

(48:04):
So what was the plan it created?What what was the, what were the
steps of execution went through,what did the tool return all of
that. So you know you know how to
tweak. It's full transparency, right?
So that kind of, you know, I think one of the concerns, I
mean, I do some of these shows and particularly we've done a
couple of agentic ones is like, you know, am I letting this
system off to do its own thing? Where am I safe?

(48:25):
Guards for the guard rails? How do I know?
You know? As I said, the reasoning behind
this. But this is superbly transparent
here, Sachin, so that's great. Yes.
And you made an excellent point again, Shane, which was that the
other big reason to do this is for audit purposes, especially
in regulated industries. So this is saved for every

(48:46):
interaction, every question that's asked.
So users can always go back to historical questions and see
what happened, what happened there.
And you can also use this for analytics.
You can see, you know how long something is taking, how large
some context is, and then make some choices in terms of
optimizing costs for by using the right LLM.
And there's a bunch of things that you can do with it.

(49:06):
OK. Is there in this rag portion
here? Is there is there an element of
memory inside and data works forthis?
If you're asking the the same sort of questions, does it build
up that over time in terms of quicker responses or is it that
not the case? There are there are abilities to
plug in caching, so you could save on the cost of the prompts

(49:29):
and all that provided your data itself is not updating.
Good. All right, so we built the Rag
application, right? So Shane, the next step we had
was to actually expose it as a tool so our agent can use it.
So let's go back here and createtool.
It's an AI app, so we picked that.
Let's let's give it a name. I think I already have a decent

(49:50):
one here, which is get leafy portal policy details.
That's that's basically. What we're doing makes sense.
And that's right, a description to it.
The description is basically going to say that get answers to
questions around store policies like shipping, returns,
cancellations, and that should give it a general idea about
what it can do. So we just have one application.
So let's pick the one that we just created and we use this.

(50:13):
So this this interface is already populated because we
know we're going to ask questionand then you're going to get
response. So that's already in there.
Let's just give it a couple of examples.
So the agent kind of knows what,what this tool can do.
And those are the questions we asked earlier.
So we can use that and, and, andthat's it.
Otherwise it's the same, same concept.
It, it gets a description and gets a tool name.

(50:35):
So let's save it. And now we have two of our
tools. That's, that's what we wanted to
do. Now comes the interesting and
exciting part of creating the agent and making this all come
together. So we, we click create agent to
launch our agent builder. And the first thing it asks is
what do you want the agent to do?
Right. And you know, I actually know
exactly what I want my agent to do because my product manager

(50:57):
told me what he wants it to do. You had the brief perfect.
I'm just going to go here and I'm going to paste the brief
right in here. You're going to put all of that
brief in there, Sachin, just like that.
I'm. Just going to dump it.
Yeah, absolutely. I'm just going to put it in
there and I'm going to create. And what it's going to do is
it's going to understand what's provided and it's going to, it

(51:20):
understands data works agents. So it's going to assign parts of
the brief to the right parts of the agent.
It is going to understand scenarios.
It will break it down into scenarios.
It will also hopefully suggest some tools to us so that we can
use the scenario. And so it kick starts your agent
development in a big way. In fact, it might, it might take

(51:42):
us a very long distance here. And then you have access to
doing whatever you want to do after that.
So for instance, so here, here'sa brief I gave and here's the
response from the agent builder saying that it's created an
agent. And let's just quickly take a
look at what's happened here, right?
So the agent in Dataworks is actually a declarative

(52:03):
specification. So it's basically defined by
this Jason file and you don't actually need to do any
programming or like or connect to any library or anything that
you basically instruct the framework what you want and that
prevents the agent for you creates the right prompts, the
right reasoning, all of that. And the way we've put it
together is that, you know, I was talking about this in an

(52:25):
earlier slide, you compose a, anagent with a bunch of scenarios
and, and the scenarios with tools, right?
So, So what we have here is a visual representational same
Jason. And if you remember the brief,
we have a persona for the agent that's been extracted from the
brief that it says you're a helpful and professional

(52:47):
assistant representing BFE portal e-commerce.
So this came to brief. And then the other thing that
came from the brief was, you know, if a user asked the
question that the agent doesn't understand, then it should
basically say this. And that's what it's extracted
and populated that. And it also identified that

(53:07):
there were two scenarios that the manager had in his brief.
One was to ask FAQs and the other was like to operate on
orders, right? So let's, let's expand this,
let's expand this and see what we got.
It's a scenario to handle FAQs, greater store policies.
We may have to expand that a little bit that might not that
maybe that's fine. We'll see.
We'll try, we'll test it out andwe'll see if we need to change

(53:28):
that. It picked up examples from the
brief. We could ask it to create
additional examples if you wanted it to do that.
And here it says that hey, I think this is a good tool to add
to this scenario. OK.
So it's suggesting the tool for that Perfect so.
Now we had only tools, 2 tools, but it did exactly.
But as you said earlier, like ifsomebody's using data works

(53:50):
across the organization, there'sgoing to be plenty of pre
populated tools there. So exactly, yes, in this
instance we only had two. It's going to pick from the 2.
But as you say, you know, we're not reinventing the wheel here
in terms of tools. So if the organizations need
these tools and, you know, they're, they're applying them
to customer service, then they're applying them to
shipping, then they're applying them to, you know, maybe HR

(54:13):
right or something. Yeah, I get it.
Exactly. Yeah.
Exactly. And so and even between the two
tools that we had, it did identify that the the Rag 1 is
the right one to be added here. So so let's add that now.
So that's that's fine. And then this one says to manage
operation specific order. And again, I think that's a
little brief, maybe we can expand that, but let's let's go

(54:34):
and see what's going on. And here it says the Mongo get
portal or a Leafy portal order is a good tool.
That's that's true. I like that recommendation.
It's not sufficient though. So I'm going to go pick the
other one as well because I wantit to be in the context of the
policy documents. So we add that as well.
And you know, at this point we are actually done with what the
PM wanted for the most part. We might have to tweak it a

(54:56):
little bit because I think some of these descriptions may be a
little short, but let's give it a shot.
Let's see what it does. So let's let's use that and say,
let's ask a question that actually it it doesn't know
about. We haven't asked this before.
We didn't ask the dagger. We didn't put it in the examples
of the tool. It isn't in these examples
either. So it's, it's just a completely

(55:17):
brand new question. And let's see if that works.
It doesn't work. We might have to change the
description so at this point the.
Upper live demo. Exactly.
So at this point, it worked justfine.
Yeah. So the agent understood the
intent, mapped it to the scenario and then executed this
tool with that question and thenprovide us an answer back.

(55:39):
So that worked just fine. That's perfect.
Yeah. And let's try and see if we can
get it to do something for the second scenario, right?
So I know that this order exists, so I'm just going to ask
this question. Sure.
And this one I suspect we might have to go to.
No, I didn't have to. This one worked fine as well.

(55:59):
So that's great. So we have a, this came from
Mongo, right? And it's it's great because it
got us the order. I find that, you know, LLMS are
remarkable pieces of tech. I mean, without me having to
tell it, it's actually figured out that the the final order
status is delivered, which is great.
Yes, I don't want to rely on that every single time.

(56:20):
So, and I also find this a little hard to read.
So let's let's try and give it an instruction to maybe format
this output a little bit more legibly, OK, and be a little bit
more consistent about it. So.
Fine tuning the output really now a little bit.
Exactly. And and we're just going to
instruct it. We don't need to do anything
more than say that if the user requests the full order details,

(56:43):
then you must use this format. But for everything else you can
just respond briefly. OK, right, so nice.
So let's ask the same question again and we'll get a completely
differently formatted answer. Actually, it follows this
format. And in this case we've actually
told it that for status, you must show the most recent status
from status history. So there is no, no, we aren't

(57:05):
relying on the, the LLM doing the right thing.
We're instructing it that this is the right thing to do.
So that's, that's one way of doing that.
Now, now I want to talk about, you know, we talked about the
agent builder, right? And this is the agent builder.
I mean, you can just give it a brief and it'll kick start a
agent for you. If you want to add new

(57:26):
scenarios, again, you can just type it in English and it'll
create a new scenario and do this the same thing, you know,
propose tasks and all of that. But you can also interact with
it with this conversational agent.
So for example, if I wanted to say let's do this.
So for the for the queue scenario and the what's the
other one called Order Order of scenario of scenario update the

(57:51):
instructions to be. When a users question cannot be
answered, respond with with and then please append the scenario
selection failure message. So we are just going to instruct
it to do something just just plain text.
You're just talking to the agentbuilder and you're saying, hey,
this is what I want, right? And hopefully what it'll do is
it'll update the scenarios for us through a conversational

(58:12):
capital. OK, OK.
It's incredibly impressive, verypowerful, and you know, to be
able to interact at such a, you know, as you said, you took the
brief and you just cut and pasted the full brief into the
beginning and then it's gone offon this.
But I love the safeguards that you have, the ability to go

(58:32):
expand out each of the sections,see what it's present to do, and
go back in and modify those. It's a incredibly comprehensive
Sachin. It's very, very clear to follow.
Thank you. I I really appreciate that
feedback. And there's one one key thing on
this saying that I didn't stressbecause I wanted to talk about
it separately. The thing about the agent

(58:53):
builder and the reason why we are so excited about this is
because this this interface on the left you have this more
traditional UI based interface and or a form based interface.
On the right you have this conversational thing that you
can interact with. All of this is powered by an
agent that is written in the exact same database agent

(59:13):
framework, the same framework that we are giving out to our
customers with no additional capabilities.
So anybody could build a sophisticated UI with
capabilities like this using ouragentic framework.
And right from the, the time we started down this thing, you
know, Sachin Smothra, who's the,the CEO of Dataworks and I, you
know, we, we wanted to create a conversational UI that

(59:38):
demonstrates the capabilities. It also pushes the boundaries of
what our own framework is able to do.
And we wanted our own stuff in production before any customer
actually had to do it. So we rely on it ourselves, and
we have. You're, you're using, you know
you're, you're using the tool tobuild the tool, right?
So. Exactly, exactly.
And, and not just that, I mean, we have a Jason that describes

(01:00:01):
our agent builder quite like this one, except that it's a lot
more, it's larger and more complex, but it's still exactly
the same concept. And the fact that we can build
this on our agent framework, which means that, you know, we
understand much better now what's what's easy, what's hard,
what's possible, and. Yeah, so, so this is our our
take on, you know, building agents using natural language.

(01:00:26):
We took inspiration from the therevolution in coding agents, for
example, Although in that case, you know, like Cursor and you
know, Aider and a bunch of thoseguys or copilot for that matter.
But in those cases you do end upgoing back to code and working
on that. That's something we wanted to
avoid here. So we went down a declarative
path. So you basically build out the

(01:00:46):
Jason and then you have you havea conventional UI and a
conversational UI that allows you to do that.
It's, it's incredibly powerful. I can only imagine the, the
things that people get to build in here too.
The one concern I have is we've all met the product manager,
Sachin. So he goes, here's my brief,
turn around 30 minutes later andgive him backup product.

(01:01:09):
He's just going to give you moreand more briefs.
But you know, this is the way the world has changed, right?
So no more. Here's the functional
specifications and we'll see youin eight weeks.
And can you conduct 4 sprints, you know?
Exactly, exactly. And, and it's, it's amazing you
bring this up, Shane, because wehave this customer who, who, who

(01:01:30):
had this, who has this problem into tracking processing of a
customer's order within the enterprise, a large enterprise.
And the processing goes through very many stages.
And I think they've probably built it up organically,
acquisitions, whatever. So it's a little hard for them
to kind of bring it all together.
So they had a traditional project they put together, you
know, get the teams on boarded, get, get all the, everything set

(01:01:52):
up. That takes time.
Then you put a solution togetherthat takes time.
And here's Nikhil, who's the CTOof Databerks walking into a demo
and building these things in the, in the demo and showing a
prototype for in like half an hour.
So it's, it's, it's transformative.
And like you said, I'm really excited about where people will
take this. It's incredibly impressive.
And look, as I said, I've done, this is probably my third show

(01:02:15):
with Dataworks. I see what your team builds so,
but this has been the most impressive demo today because I
just love how it all. It builds on everything that
I've seen Dataworks do before, but now it's such a compelling
solution. I think it's amazing and it
seems so slick and effortless. Sachin, were there any kind any

(01:02:35):
unexpected challenges or roadblocks putting this together
that you you hadn't expected andhow did you overcome them if
there? Were no, there were.
There were a bunch of things. Because so much of this is new,
we started fiddling around with agents, experimenting with them
sometime last year. Our first implementation was
React, and React was great, but we struggled with certain

(01:02:55):
aspects of React. We just felt a little
uncontrolled, I don't know maybeif it was our implementation or
or reacts gotten improved since then, have no idea.
But we kind of moved away from React and picked another
framework to implement called LMcompiler that's served as well.
But that was one of our first was to kind of choose a
foundational agent implementation that would that

(01:03:17):
would be good. The other thing was coming up
with, you know, coming up with what would be easy to use to
build agents, because that was our promise.
We did that for RAG. We want to do that for agents.
And we were so early with it. It's it's a little hard to
imagine whether it'll all work out.
So that was that was challenging.

(01:03:37):
Originally they thought this would be a proof of concept, but
the proof of concept kind of convinced us that we could
actually put it in the product. The third, I think challenge is,
is just the user experience, right?
This is, this is new even for our, our, our UX person.
She's, she's also coming to terms with, you know, what are
the interaction models? How, how do you present this to

(01:04:00):
not how do you, how do you present this and how do you
present this to the development team that needs to build it?
How do you capture all the interactions?
So that's, that's been interesting there.
There's so many new areas here that we are all learning
together. I think it's just something a
fun journey. Yeah, well, look it, it
certainly looks superb. I know we're over time.

(01:04:20):
I appreciate everybody still joining us etcetera as well too.
Throughout I put up some of thisthat you had shared with me,
Sachin. So if anybody wants to go back,
if they're looking at the recording, you can dig down
through the comments and you will find a link to the blog
posts, the video, the GitHub repo for this, and also a link
to sign up for for data works and try this yourselves.

(01:04:42):
So and and that's sign up for free and, and get in there and
start playing around Sachin, right?
Yeah, absolutely. Yeah.
This is the this is the read me that has very detailed steps for
everything we saw today. So you can, you can do this.
You can sign up for free. And you know, like I said,
there's a hosted Mongo DB that you can use.
So yeah, absolutely. The only thing you would need to

(01:05:02):
bring is your LLM key. Yes, you saw that on the way in
as well too. Well, look, I mean, I think it
was stunning in terms of, yeah, you started the demo, there
wasn't anything in there. OK, You had, you had added the
LLMS and you had your connections etcetera set up as
well too. But other than that, you sent it
off, it created the embeddings, it built that agent, it then

(01:05:24):
went off and you added the otheragent and then you drop the
brief in and it just worked, which is incredible.
So I think, you know, sometimes we see, as I said at the intro,
we see a lot of really cool AI demos and but they're, I don't
mean to say that they're smoke and mirrors, but as you said at
the beginning, they're not very practical.
This was incredibly practical. This you could easily see how

(01:05:46):
you could apply something like this to most businesses, most
industries. So I appreciate how swift and
how easy you made it all look, Sachin, and and as you said,
you're everyone else can go and check it out as as well too.
To wrap up this, which has been a great conversation.
Where's next for data work? So where do you see this type of

(01:06:07):
technology going? Because we really are, you know,
the, the Gen. AI space 2 and a bit years maybe
there. So, um, you know, LMS and
foundation models, et cetera, etcetera.
But the agentic space is superbly new.
So what's next? Yeah.
Yeah. And I'll keep this brief because
we are we got gone over time. Um, so we talked about

(01:06:27):
conventional drag already and you know that powered, um,
information bots or search widgets and stuff for that
agentic rag. It's just been amazing.
I mean, in all my 25 years of experience, I haven't seen
people react to conversations about agentic Rag or demos or
presentations the way they reactto it.
The moment we talk about it, there are folks on the call that

(01:06:48):
have ideas whether it's, you know, automation or some
operational efficiency or just something transformation that
instantly clicks for them. So I think we are at the doorway
of a vast sea of opportunities with Integrat.
So we are trying to solidify theoffering.
We we have a number of ongoing Pocs and conversations around

(01:07:10):
agentic lag. So that gives us a lot of
insight into what people won't build and that'll go into the
product. Of course, there's a lot of work
already happening there. We are integrating it where it's
required, like, you know, Slag, Zendesk, Salesforce, wherever a
conversational agent could be required that is driven by this.
And then as, as, as it gets usedand deployed and confidence

(01:07:31):
builds the next step, and we're already having conversations
around this next step is going to be, you know, making actions
from it doing, you know, doing small things, maybe update
tickets, close tickets, but theneventually we start doing
actions on behalf of the user, you know, maybe updating
Salesforce. So then we get into actionable
rack and we have a couple of customers who are already making
these plans for, you know, how they can automate some of their

(01:07:55):
workflows by triggering them, triggering the agent from human
triggers, the agent with some inputs.
But that once they start to see enough confidence in the
reasoning and the planning and you know, it does the right
thing. And we have the supporting
tooling around that that's goingto move into triggered by an
event or triggered by, you know,schedule triggering whatever.
And that's going to be autonomous again.
So there is there's a whole journey here as the confidence

(01:08:17):
and experience with it increasesover time within the enterprise,
we'll see various aspects of it.And it's not going to be, it's
probably all going to live together, probably have some
autonomous agents and some action about diagonal of that.
And you know, it's, it's super exciting where we're going with
this. It's it's incredibly exciting
and it's it's also nearly overwhelming how quick things

(01:08:38):
are progressing as well too. You know, if I think back, you
know, when we started to see some of this and then, you know,
even back to when we started to see rag being built, it's like,
wow, we've moved on now we've got that.
That's it. You know, if you if you were to
do a yeah, rag demo still exists, but like, you know, the
agentic piece layered on top, the reasoning behind, you know,

(01:08:59):
how it knows where to go for answers, how it knows how to
choose. It's brilliant.
And and you know, then, if you take what data works have done,
which is as and kudos to your UIdesigner.
I think she she did an amazing job.
And then obviously the developers and implementing that
super easy to follow. So I very much look forward to

(01:09:19):
getting you back for one of these purple or green boxes when
we get to the actionable and theautonomous, right so.
That would be wonderful, yes. Get you back on the show.
But for today, this has been superb and really powerful.
Clear demo that I think everybody will agree it's been
great to get the comments in. Sorry we didn't get to go
through a lot of them, but I think it's great.

(01:09:40):
But as you, as I said, if you want to learn more, there's the
blog, there's a video and most importantly, there's the repo,
but also the link there to Dataworks where you can create
an accountant and see it for yourself.
So umm yeah, Sachin, chief architect of Dataworks, this has
been a superb stream. Thank you so much for your time.
Thank you. Very much Shane, it's a
privilege to be here and show all this to you and the

(01:10:02):
audience. Thank you so much for the
opportunity. Well, listen, it's a it's my
privilege to host you. Thank you so much for your time
and how clear and straightforward you made this
all good luck. And I know tons of work behind
the scenes to build products like this.
But the proof is very straightforward in that you
know, you, you set the scene andyou did the demo from scratch
and it just worked. And I, I think that's great.

(01:10:23):
So, umm, certainly we'll get youback as data works, move forward
into that journey further, have new products to show.
Delighted to have you or any of your colleagues back on the
show, Sachin, Thank. You really appreciate that.
Thank you. Excellent.
And look for me and all of the other folks here in MongoDB.
Thank you all for tuning into these live streams.
As I said at the intro, keep an eye on our YouTube channel and

(01:10:44):
LinkedIn for other streams coming your way that are exactly
like this. We do appreciate everybody
joining in and viewing and adding the comments.
So until next time from me, Shane McAllister, and from
amazing guest Sachin. Thank you so much for joining us
everybody. We will see you on another
podcast live, hopefully shortly.Take care.
Goodbye.