00:00:00
[Music]
00:00:13
thank you
00:00:15
hi everyone developers Builders and
00:00:17
innovators from all around the world
00:00:20
welcome to Devcon our second virtual
00:00:23
developer conference I'm thrilled to be
00:00:26
here today as our goalie as CEO and
00:00:29
delighted delighted even to witness this
00:00:32
Gathering of the best and brightest in
00:00:34
our tech industry
00:00:36
Defcon marks a pivotal moment in our
00:00:39
journey
00:00:40
where we celebrate the power of
00:00:42
collaboration Innovation and the
00:00:44
boundless possibilities of Technology
00:00:47
we're here today to Deep dive into the
00:00:51
tech to challenge existing boundaries
00:00:53
and to Forge New Paths that will shape
00:00:56
the future of the search and Discovery
00:00:58
industry
00:01:00
in the past year we've seen amazing
00:01:02
leaps forward in the application of
00:01:05
artificial intelligence
00:01:07
and you know we needn't tell anybody on
00:01:09
this call with the Advent of chat GPT at
00:01:12
the end of last year the amount of new
00:01:15
emerging applications that people are
00:01:17
creating with generative AI is just
00:01:20
amazing as we try to figure out how we
00:01:22
harness artificial intelligence in
00:01:24
pretty much everything that we do these
00:01:26
days and we here at algolia we didn't
00:01:28
sit idly by either after the acquisition
00:01:31
of search i o we worked hard behind the
00:01:34
scenes to build and launch the next
00:01:37
generation of AI powered search and
00:01:39
discovery
00:01:40
on May 2nd I'm very proud to say that we
00:01:43
unveiled algolian neural search
00:01:47
today our ever-evolving search platform
00:01:50
uniquely handles both keyword and Vector
00:01:54
search in one product one API and it
00:01:58
uses AI in query understanding in
00:02:01
retrieval and in ranking so in essence
00:02:05
we are truly end-to-end AI
00:02:08
and we apply it to every single query I
00:02:12
think that's what a really big takeaway
00:02:14
Point
00:02:15
um that we'd love you to sort of walk
00:02:17
away from this with too
00:02:19
because nobody else in the industry does
00:02:20
it and nobody else does it have five
00:02:22
nines of availability which is when you
00:02:25
think about it 78 seconds of downtime a
00:02:28
quarter yes let me just repeat that 78
00:02:31
seconds of downtime a quarter
00:02:34
so you'll see the experience and the
00:02:37
full power of algolian neural search
00:02:39
here at devcom today
00:02:42
so in today's rapidly evolving online
00:02:46
world that we all live in these days
00:02:48
it's it's developers like you and us
00:02:51
that are really the true architects of
00:02:54
change
00:02:55
you're the builders who transform
00:02:58
abstract Concepts into tangible
00:03:00
solutions that improve lives create
00:03:03
efficiencies and connect us in profound
00:03:06
ways
00:03:07
your code your Creations have the power
00:03:11
to revolutionize Industries disrupt the
00:03:14
status quo and Spark transformative
00:03:17
experiences for people across the globe
00:03:20
you are the builders that make people's
00:03:24
lives better
00:03:26
at this conference we're United we're
00:03:29
United by a common passion
00:03:31
the Relentless pursuit of innovation
00:03:35
today you'll have the opportunity to
00:03:37
expand your horizons to learn from our
00:03:40
subject matter experts experts and to
00:03:43
engage with like-minded individuals in
00:03:45
our developer community
00:03:47
and to discover the latest advancements
00:03:49
from algolia
00:03:51
that are shaping our Collective future
00:03:55
we've curated a rich program of
00:03:58
presentations and Technical workshops
00:04:00
and interactive sessions that we hope
00:04:02
will inspire you and challenge you as
00:04:05
well as equip you with the knowledge and
00:04:07
tools necessary to tackle the complex
00:04:09
problems of search and discovery
00:04:12
from artificial intelligence and query
00:04:14
categorization to image search Vector
00:04:17
hashing from conversational AI
00:04:19
conversational Commerce to seamless
00:04:22
connectivity with major platforms like
00:04:25
major e-commerce platforms for example
00:04:27
amongst others and our agenda is
00:04:30
brimming with topics that reflect the
00:04:32
cutting-edge trends and emerging
00:04:34
technologies that will Define the next
00:04:36
chapter of our industry
00:04:38
significantly we are harnessing the
00:04:40
power of AI and making it available to
00:04:43
you part of that democratization
00:04:46
and you won't have to train the models
00:04:48
that's kind of what I mean by the
00:04:49
democratization of it I know that's you
00:04:51
know a little bit
00:04:53
like magic
00:04:55
um you know and that's because we train
00:04:57
on multiple llms however this conference
00:05:00
is not just about learning or about the
00:05:03
technology it's also about fostering a
00:05:05
community a community of
00:05:07
forward-thinking individuals who share a
00:05:10
common purpose it's about building
00:05:12
lasting connections
00:05:14
sparking collaborations
00:05:16
empowering each of you to scale New
00:05:19
Heights
00:05:20
throughout the conference I encourage
00:05:22
you to engage with your peers ask us
00:05:25
questions share your insights and Embark
00:05:28
upon a collective journey of growth and
00:05:32
learning
00:05:34
as we embark on this virtual experience
00:05:36
here for the next few hours together
00:05:38
including a local Meetup in our Paris
00:05:41
office where I am at the moment
00:05:43
I want you to remember that the true
00:05:45
magic lies in your hands your passion
00:05:48
your ideas your commitment to pushing
00:05:51
the boundaries of what is possible are
00:05:53
the driving forces behind our Collective
00:05:57
success
00:05:58
so let's embrace the power of this
00:06:01
moment let's celebrate the you know in
00:06:03
the spirit of collaboration Innovation
00:06:04
and continuous learning let's unlock
00:06:07
those new possibilities and embrace the
00:06:09
challenges that lie ahead and that the
00:06:11
technology opens the door to and
00:06:14
together we truly can
00:06:15
create and shape a brighter future
00:06:19
um more than I think at any other point
00:06:22
in our history
00:06:24
in the search sphere so thank you for
00:06:27
being part of this virtual developer
00:06:29
conference I'm excited to see the
00:06:31
incredible ideas that will emerge and
00:06:33
I'm confident that our Collective
00:06:35
efforts will Propel us towards a class
00:06:38
of endless
00:06:40
innovation
00:06:41
alrighty welcome to Defcon and let it
00:06:44
begin
00:06:46
indeed AI is everywhere thank you
00:06:49
Bernadette hello everyone my name is
00:06:52
Bharat and I'm the chief product officer
00:06:54
at algolia and your host today welcome
00:06:57
to algolia developer conference 2023 we
00:07:01
are thrilled that you have joined us and
00:07:02
over the next two days of action-packed
00:07:05
updates learning sessions and community
00:07:07
building we hope that you're going to
00:07:09
have a really enjoyable time
00:07:11
we have a lot of exciting things to talk
00:07:13
about and even more exciting things that
00:07:15
I want to announce today at algolia our
00:07:18
vision mission and purpose is powering
00:07:20
Discovery we wake up every morning and
00:07:23
obsess over how we can help our
00:07:25
customers put their data into motion so
00:07:28
it is discoverable across all use cases
00:07:30
and segments this is a never-ending
00:07:33
problem and journey and that's what
00:07:35
makes it both challenging and rewarding
00:07:37
for our customers and for us I want to
00:07:40
start with a quick look by numbers of
00:07:42
how algolia has helped our customers
00:07:44
power Discovery we process over 1.75
00:07:48
trillion searches per year and this
00:07:51
number is only growing dramatically
00:07:54
that translates to over 35 billion
00:07:57
searches per week
00:07:59
at our scale that means one in six
00:08:02
internet users are touched by algolia
00:08:05
through one of our customers
00:08:08
we have more than 5 million developers
00:08:10
who have used algolia in some way shape
00:08:13
or form
00:08:14
and we have more than 17
00:08:17
000 actual customers who are in
00:08:19
production
00:08:21
we return results in milliseconds which
00:08:24
is speed at scale that is unmatched
00:08:26
anywhere
00:08:28
in 2022 we had a 119 customer and scale
00:08:33
related releases to make the search and
00:08:36
Discovery experience even more magical
00:08:40
operationally we maintain five nines of
00:08:43
availability so you don't experience
00:08:45
downtime we even have customers who use
00:08:47
the algolia index as a simple key value
00:08:50
store because of its reliability
00:08:53
last Black Friday we hit over 120 000
00:08:57
queries per second and not a single one
00:08:59
of our Engineers was paged because of
00:09:01
scaling challenges or incidents we
00:09:05
understand what scale means and we
00:09:07
understand what speed means at algolia
00:09:10
we recently celebrated our 10-year
00:09:12
anniversary and we are extremely
00:09:14
thankful for every single customer who
00:09:17
has trusted us with their search and
00:09:18
Discovery needs thank you very much
00:09:21
and we're just getting started I
00:09:23
couldn't be more proud to have shared
00:09:25
some of those numbers with all of you
00:09:28
now before I get into the actual product
00:09:30
announcements section I want to start by
00:09:33
giving a shout out to our community both
00:09:35
our official algolia ambassadors from
00:09:38
all around the globe and our developer
00:09:40
Community who have built great search
00:09:42
experiences
00:09:44
we wouldn't be here without any of you
00:09:46
so thank you for your support all these
00:09:48
years
00:09:49
please don't miss Jacob's fantastic talk
00:09:52
later today about how nuts.js built
00:09:54
blazing fast search with algolia
00:09:58
I also want to talk about two of our
00:10:00
developer customers the first is resend
00:10:03
which is a new API for developers to
00:10:05
send emails we power all of their
00:10:08
documentation the second is built at
00:10:10
light speed which is a browsable catalog
00:10:13
of thousands of themes and uis to skin
00:10:16
your jam stack websites
00:10:19
okay let's get to the main section
00:10:23
so I want to start by you know talking
00:10:25
about the ways we've made developer
00:10:26
lives easier over the last year then
00:10:28
we'll talk about all of our new
00:10:30
Innovations and then we'll wrap it up
00:10:33
so we really care about making your
00:10:36
lives as developers easier we know what
00:10:39
it means to build applications in all
00:10:41
its moving pieces and we want to make
00:10:43
our part search and Discovery as
00:10:46
straightforward as possible
00:10:49
the first step in making it easier for
00:10:52
all of you was to bump up our free plan
00:10:55
now called the build plan
00:10:57
it gives you a million records for free
00:10:59
compared to the 10 000 before it has no
00:11:03
time limit and our goal is to make a
00:11:05
free plan so interesting that there is
00:11:07
no side project or startup idea that you
00:11:11
couldn't build with it
00:11:14
last year we released the algolia
00:11:16
command line interface and the response
00:11:18
to it has been phenomenal more than 700
00:11:21
applications have been using it to
00:11:23
configure their indices create snapshots
00:11:25
and load backups directly from the
00:11:28
terminal or in CI Scripts
00:11:31
we've been using it extensively
00:11:33
internally as well and we all like it so
00:11:36
much that we've been working on a web
00:11:38
command line interface version
00:11:40
speaking of which we'll be unveiling
00:11:42
this web command line interface version
00:11:44
in the next iteration of our
00:11:46
documentation we'll be releasing a new
00:11:48
beta version of our documentation very
00:11:50
soon
00:11:51
same content but improved ux and DX
00:11:54
thanks to interactive code Snippets and
00:11:57
a dark mode to name a few
00:11:59
to know more about all the developer
00:12:01
experience considerations that came into
00:12:04
the new version of the dock don't miss
00:12:06
Khalid and Lloyd's talk later
00:12:11
on this new iteration of the docs we
00:12:13
decided to move away from our custom
00:12:15
search implementation instead we'll now
00:12:18
follow a eat your own dog food approach
00:12:20
and use doc search for our own docs the
00:12:23
size of our documentation is massive and
00:12:26
as we deeply care about the ease of use
00:12:29
the developer experience and the user
00:12:31
experience we have used our own
00:12:34
documentation as a playground in the
00:12:37
past to try out new search patterns and
00:12:39
user experience today all those
00:12:42
learnings have been included into doc
00:12:44
search V3 so there is no longer any
00:12:46
reason to continue with a custom
00:12:48
implementation now any Improvement we
00:12:51
make for ourselves will also benefit you
00:12:56
as developers ourselves we understand
00:12:58
the pain of writing plumbing and
00:13:02
orchestration and Scaffolding code we
00:13:05
want you to focus on writing the code
00:13:07
that counts for your application and let
00:13:09
us handle the boring parts so we
00:13:12
introduced our no code platform that we
00:13:15
call connectors you can easily plug your
00:13:18
algolia index on the receiving end of
00:13:20
external Json apis and have it
00:13:23
automatically update when new content is
00:13:26
added edited or removed we take care of
00:13:29
all the plumbing of synchronization for
00:13:31
you so it is easier than ever to add
00:13:33
algolia search on top of your existing
00:13:36
API
00:13:37
since our last developer conference we
00:13:40
have added direct support for Commerce
00:13:42
tools apis and we have big things coming
00:13:45
up the release of our bigquery connector
00:13:47
and our big Commerce connectors
00:13:51
as I said earlier we want you to write
00:13:54
the code that counts we use the same
00:13:56
logic when updating our instant search
00:13:59
front-end Library we removed all the
00:14:01
boilerplate required to pass our
00:14:04
insights events and distilled it down
00:14:06
into a simple incises true Boolean flag
00:14:09
this is disabled by default because
00:14:12
we're privacy conscious but you as a
00:14:14
developer can now enable event sending
00:14:16
with a flip of a Boolean don't miss out
00:14:19
her own stock where he'll dive deeper
00:14:21
into all the complex developer questions
00:14:24
that led to this very seemingly simple
00:14:27
change you'll be surprised by how many
00:14:29
considerations we had to take into
00:14:30
account
00:14:33
we really wanted to make this actioning
00:14:35
events easy for you because events are
00:14:38
critical in this new AI world inside
00:14:41
events are signal in all the noise they
00:14:44
tell our underlying models what actions
00:14:47
are performed by your users and they are
00:14:50
used to train models on your specific
00:14:52
data and user Behavior this in turn
00:14:55
produces better informed search results
00:14:57
which creates a virtuous cycle of more
00:15:00
qualitative events being sent none of
00:15:03
that is possible without the initial
00:15:05
sending of events because
00:15:07
events are the electricity that powers
00:15:10
AI events that are specific to your data
00:15:13
and your users is what powers the
00:15:16
amazing set of innovations that we are
00:15:19
building
00:15:21
so that was a very quick reel of how
00:15:23
we're making your lives as developers
00:15:25
and Builders easier I hope that you all
00:15:28
appreciate that and are able to take
00:15:30
advantage of all of these tools I now
00:15:33
want to talk about all the exciting
00:15:35
stuff that I promised at the start of
00:15:37
the keynote that we want to talk about
00:15:39
and announce today so let's get into it
00:15:41
[Music]
00:15:42
I want to start with Doc search
00:15:45
when we built doc search in 2016 our
00:15:49
conviction was that every open source
00:15:51
documentation deserved a great search
00:15:53
experience as developers we spend a lot
00:15:57
of time in open source documentation
00:15:58
looking for relevant information as a
00:16:01
company we knew that we wouldn't be
00:16:03
where we were if it were not for open
00:16:05
source building doc search and giving it
00:16:07
away for free was our way to give back
00:16:10
to the community
00:16:12
reception from the developer Community
00:16:13
has been huge doc search now helps you
00:16:16
find relevant information in more than
00:16:18
7500 projects including some of the most
00:16:21
popular on the internet like react
00:16:23
laravel or nux our whole doc search
00:16:27
infrastructure handles 1.6 billion
00:16:30
search requests per year for you for
00:16:33
free
00:16:35
we want to extend the power of Doc
00:16:37
search to Beyond just open source
00:16:39
documentation we want every piece of
00:16:42
technical knowledge to be powered with
00:16:44
the doc search this means your technical
00:16:46
blog your wikis and even documentation
00:16:49
for your closed Source software or your
00:16:52
SAS company
00:16:53
as long as it helps share written
00:16:55
technical knowledge we want to help your
00:16:57
users find relevant information for free
00:17:01
so I'm very excited to announce that
00:17:04
today we are inviting all of you to join
00:17:07
the program and apply to have doc search
00:17:09
Power your technical content if we
00:17:11
rejected your application in the past
00:17:13
because your project was not open source
00:17:15
now is your time to apply again we'll
00:17:18
crawl your website and send you all the
00:17:20
relevant JavaScript Snippets under 48
00:17:22
hours
00:17:25
next I want to talk about neural search
00:17:27
the world's first end-to-end AI powered
00:17:31
hybrid search engine which we made
00:17:33
available a month ago
00:17:35
the Quantum Leap in generative Ai and
00:17:38
the consumerization of chat GPT at the
00:17:40
end of 2022 changed everything in terms
00:17:43
of consumer expectations companies are
00:17:46
now in an existential race to become AI
00:17:49
powered and meet the new expectations of
00:17:51
their customers to be truly understood
00:17:55
we are uniquely addressing that Demand
00:17:58
by bringing together the best aspects of
00:18:00
our world's fastest and most accurate
00:18:02
keyword search and our newest Vector
00:18:05
search engine
00:18:06
keyword search is still important it is
00:18:08
silly to match with vectors when you
00:18:10
know exactly what you're looking for and
00:18:13
the matches available at the same time
00:18:16
it is important to return matching
00:18:18
Concepts we do this hybrid evaluation in
00:18:21
real time on every single query and
00:18:24
return the most relevant results merged
00:18:27
between our keyword results and our
00:18:29
Vector results and this is unmatched in
00:18:32
speed and accuracy
00:18:34
I'd like to deconstruct search into
00:18:37
three very simple steps every search
00:18:40
query goes through three stages from the
00:18:42
moment you start typing
00:18:44
at the bottom to the moment the results
00:18:47
are presented to you at the top
00:18:50
so the first is query understanding the
00:18:53
first time a search engine sees a query
00:18:55
I.E when you type something into a
00:18:58
search bar
00:18:59
it is parsing it using a variety of
00:19:02
techniques and AI techniques like
00:19:04
natural language understanding or entity
00:19:06
extraction
00:19:08
and at this point the search engine has
00:19:11
not even started to retrieve any results
00:19:13
at this point the engine is preparing
00:19:17
the query to be sent in for evaluation
00:19:22
now once the engine is fed with the
00:19:24
prepared query it retrieves and returns
00:19:27
a set of results that it determines to
00:19:30
have the highest recall and precision
00:19:32
basically a relevant set of results that
00:19:35
are ordered in the most desirable way
00:19:39
after the results are retrieved
00:19:42
additional ordering can be done through
00:19:45
a re-ranking step where signals like AI
00:19:47
personalization or some other algorithm
00:19:49
like a learn to rank algorithm can
00:19:52
change the order for highest
00:19:54
desirability
00:19:56
now up till now the retrieval segment
00:19:58
the middle segment never really had any
00:20:01
tangible AI techniques applied to it as
00:20:04
it was extremely hard to do so
00:20:06
so for decades everyone in innovated
00:20:09
around the edges of understanding and
00:20:12
re-ranking and applied AI techniques
00:20:14
there I'll go there to the same too
00:20:16
until today
00:20:19
for the first time neural search is able
00:20:22
to apply the full force of AI by
00:20:25
applying vectorization and our own
00:20:27
neural hashing to any inverted index
00:20:31
some cynics have dismissed neural
00:20:33
hashing as locality sensitive hashing
00:20:35
which has a lot of drawbacks when it
00:20:37
comes to Performance and accuracy we
00:20:40
actually agree with those drawbacks
00:20:42
I'm also here to tell you that's not
00:20:44
what we do
00:20:45
recall the scale and speed numbers we
00:20:48
operate at Five Nights of reliability
00:20:50
millisecond response times 120 000 query
00:20:54
per second and unmatched accuracy we
00:20:57
maintain that performance and we have
00:20:59
actually improved the accuracy ahead of
00:21:02
that
00:21:04
as a result today we are the only search
00:21:07
company who applies AI to every single
00:21:11
step of the search query till the
00:21:14
results and to prove it here are some
00:21:18
real numbers from real customers in
00:21:20
production workloads
00:21:23
all of our customers that went through
00:21:25
the private beta saw impressive results
00:21:28
all of them had a massive jump in
00:21:31
conversion ranging from five percent to
00:21:33
a whopping 22 this was due to the
00:21:37
Improvement on click-through rates you
00:21:39
know as high as 11 and the huge impact
00:21:42
neural search has on finding results
00:21:44
even if no keyword matches come up this
00:21:48
drastically reduced the no results Page
00:21:51
by up to 70 percent these are numbers
00:21:54
that are not
00:21:55
typically heard of people are usually
00:21:58
operating in the low single digits if
00:22:02
that when they talk about massive
00:22:04
improvements here we're talking about
00:22:06
high double digit numbers and so for the
00:22:09
first time you know we are seeing the
00:22:11
power of neural search in production
00:22:14
that global fashion brand that you see
00:22:16
there is actually a true global fashion
00:22:19
brand but we can't actually reveal the
00:22:21
name because it's under NDA but I can
00:22:23
assure you that it is a globally known
00:22:26
and well-beloved fashion brand
00:22:29
so I'm very excited about neural search
00:22:32
and now let us actually see neural
00:22:35
search in action I'm going to hand you
00:22:36
over to Dustin quotes our principal
00:22:38
product manager in our AI search
00:22:40
organization who is going to do a demo
00:22:43
and talk about the next set of
00:22:44
innovations that we're announcing today
00:22:46
over to you Dustin
00:22:49
hey everyone neural search combines
00:22:52
algolia's full text keyword search
00:22:54
engine with Vector understanding into a
00:22:56
single API this provides the best
00:22:58
relevance on both headquaries as well as
00:23:00
the long tail all informed by user
00:23:02
Behavior
00:23:03
we can see this in action
00:23:05
let's start off by searching for
00:23:07
something that would reasonably be a
00:23:09
popular query and so one that we could
00:23:11
Target and optimize for
00:23:13
maybe we want an organizer
00:23:16
these results are looking great for
00:23:18
keyword search
00:23:20
let's now go even deeper we instead want
00:23:25
say an organizer for toys
00:23:28
we have a few results but might there be
00:23:30
even more
00:23:33
now when we search a neural search we
00:23:35
are matching based on concept as well as
00:23:37
on keyword
00:23:38
and so now we're returning all the
00:23:41
results that are relevant without any
00:23:44
synonyms added at all
00:23:48
well we're very happy with neural search
00:23:49
we haven't stopped working to make it
00:23:51
even better
00:23:52
the first thing we want to show you is
00:23:54
called adaptive correction adaptive
00:23:56
correction allows you to influence the
00:23:58
language model that powers neural search
00:23:59
using your expertise of your business
00:24:03
for example if we were to search for
00:24:06
something like living room there are
00:24:08
some results here that really are ideal
00:24:11
for some when someone searches living
00:24:12
room
00:24:13
the way that we might influence this in
00:24:15
the past is we would create rules and we
00:24:17
would pin results up at the top but it
00:24:19
really only influenced that single query
00:24:22
and for those records that we choose
00:24:25
with adaptive correction however you can
00:24:27
tell the model this is a good result and
00:24:31
it will understand that and it will feed
00:24:33
it back into its understanding so that
00:24:35
the results change overall not just for
00:24:37
that single record
00:24:39
let's see what it looks like for this
00:24:41
query
00:24:43
now when I search for a living room I'm
00:24:45
probably looking for something like
00:24:48
this sofa right here or maybe this sofa
00:24:52
and
00:24:54
let's say this sofa as well
00:24:57
and let's go ahead and validate that
00:25:00
adaptive correction
00:25:02
and what we'll see is that now we're
00:25:05
seeing a lot more sofas appear in those
00:25:07
first results the language model has
00:25:09
learned that for us when we search for a
00:25:12
living room and of course we're the
00:25:14
experts here
00:25:15
that sofas are what should be showing
00:25:17
higher up again we're not just pinning
00:25:19
the results we're actually influencing
00:25:22
the model that works underneath
00:25:24
this is incredibly powerful because you
00:25:27
now get to marry the intelligence of
00:25:30
those language models with your
00:25:32
expertise of your business
00:25:36
adaptive correction isn't the only
00:25:37
intelligent ranking Improvement we're
00:25:39
working on we're also happy to show you
00:25:41
Fusion ranking
00:25:43
Fusion ranking is the first of what we
00:25:45
call the ranking stack it is a central
00:25:47
hub for all the signals that matter to
00:25:50
come together and make the perfect
00:25:51
search ranking
00:25:53
there are two important pieces of fusion
00:25:55
ranking first is that with recent
00:25:58
advancements in Ai and data there is
00:26:00
increasingly more information that you
00:26:02
have that should be shaping your search
00:26:04
relevance
00:26:05
you might start off with straightforward
00:26:07
keyword and Vector relevance mixed such
00:26:09
as what we have in neural search
00:26:12
then you want to mix in your business
00:26:14
metrics as well as personalization and
00:26:16
popularity via dynamic re-ranking
00:26:20
all of these signals provide Vital
00:26:22
Information on search relevance
00:26:24
and it's just the start
00:26:26
Fusion ranking is the ranking stack and
00:26:29
that means in the future it will expand
00:26:31
to take in even more rear signals and
00:26:33
apply it to search results to improve
00:26:35
ranking
00:26:36
how we are building Fusion ranking makes
00:26:39
it a true ranking platform we are
00:26:41
introducing a framework that accepts any
00:26:43
new signals and folds them into the
00:26:45
ranking in a seamless manner
00:26:48
you may start off with Vector keyword
00:26:50
and popularity signals and then expand
00:26:52
to include learning to rank or even
00:26:54
bring your own re-ranking model
00:26:57
the second important piece of fusion
00:26:59
ranking is that we know that it is not
00:27:01
easy to know which signal is the most
00:27:03
impactful
00:27:04
there's just too much data for us to
00:27:06
control ourselves but this is a perfect
00:27:08
task for AI
00:27:10
inside Fusion ranking the AI learns each
00:27:15
signals influence on clicks and
00:27:18
conversions and overall search success
00:27:22
it then optimizes those different
00:27:24
signals and gives more weights to those
00:27:27
that have the biggest impact
00:27:30
and it learns over time so as user
00:27:32
Behavior changes the influence of
00:27:34
different signals might change
00:27:37
through AI driven Fusion ranking you can
00:27:39
be sure that you're getting the ideal
00:27:41
search relevance and it will grow with
00:27:43
you
00:27:44
and the final thing we want to announce
00:27:46
today is our new multi-language support
00:27:48
starting today we support 50 languages
00:27:51
for neural search out of the box
00:27:53
and for those of you who have indexes
00:27:55
that bridge multiple languages like many
00:27:57
of our Canadian friends we are excited
00:27:59
to say that neurosurge supports multiple
00:28:01
languages at the same time
00:28:03
multi-language neural searches available
00:28:05
today while Fusion ranking and adaptive
00:28:08
correction will be available later this
00:28:10
year
00:28:10
to receive a notification when you can
00:28:12
start using them sign up for our mailing
00:28:14
list
00:28:16
thank you Dustin I cannot be more
00:28:18
excited for our new re-ranking framework
00:28:21
that allows for any learning to rank
00:28:23
model to be plugged in and uses AI to
00:28:26
arbitrate the different signals
00:28:28
I also love that we're allowing for
00:28:30
reinforcement learning through human
00:28:32
feedback which we call adaptive
00:28:35
correction
00:28:36
and the fact that our llm models now
00:28:39
Support over 50 plus languages means
00:28:42
that you get out of the box support
00:28:44
anywhere in the world at the speed scale
00:28:46
and accuracy of algolia I'm happy to
00:28:49
announce that we have opened up our
00:28:51
waitlist to self-service with the goal
00:28:54
to go fully open later this year when
00:28:57
you log into your algolia dashboard or
00:28:59
if you're signing up for the first time
00:29:01
you will get the option to join our wait
00:29:03
list to get Early Access please take
00:29:05
advantage of this and we cannot wait to
00:29:08
see you use neural search
00:29:11
let us keep going because we have more
00:29:14
announcements
00:29:15
chat GPT open up the possibilities for
00:29:19
workloads that we never imagined before
00:29:21
it also brought back to life the world
00:29:23
of chat Bots and conversations chat Bots
00:29:26
are not new but their experience has
00:29:29
been subpar at best now with the power
00:29:32
of llms applied to your data you can
00:29:35
provide a relevant conversational AI
00:29:37
interface that is actually engaging and
00:29:40
helpful
00:29:41
but slapping a chat bot on your
00:29:44
application is not enough the ux is
00:29:46
critical and in this age of generative
00:29:49
AI the race is on for who will provide
00:29:52
the best integrated ux
00:29:55
so let's see it in action I'd like to
00:29:57
introduce ayush Iyer who our director of
00:30:00
user experience who's going to introduce
00:30:01
algolia's conversational AI framework
00:30:04
over to you ayush
00:30:08
thanks Bharat and hello everyone
00:30:11
my name is ayush and I'm so excited to
00:30:14
talk to you today
00:30:16
our team has been diving deep into
00:30:19
understanding how AI can impact search
00:30:21
and Discovery experiences
00:30:24
our first area of focus has been the
00:30:28
shopping experience
00:30:29
we're curious about how AI helps
00:30:32
what you are excited about and
00:30:35
importantly what your users really need
00:30:39
we watch it on a deep dive on many AI
00:30:42
power chatbots in the market to
00:30:44
understand how they are performing
00:30:48
and with all of this we've collected our
00:30:50
key insights into this highly scientific
00:30:53
slide
00:30:55
many AI power chatbots promise the world
00:30:58
oh you'd be able to communicate with
00:31:00
customers in part Breaking ways they say
00:31:03
and yet they have so little context on
00:31:06
what your users are doing
00:31:09
to these Bots it doesn't matter if your
00:31:12
user has spent 10 seconds or 10 minutes
00:31:15
browsing your website
00:31:17
they just ask radharan emotionally
00:31:20
what can I help you with
00:31:23
also have you noticed how chai Bots
00:31:26
require you to type to be useful
00:31:29
I mean it's great in many circumstances
00:31:31
but we do believe that the power of
00:31:33
generative AI should be in so much more
00:31:36
than typing a message to expect a
00:31:37
response
00:31:39
I mean why can't we converse with AI
00:31:41
just like we search and browse a website
00:31:45
it was very clear here that there had to
00:31:47
be a better way so we talked to our
00:31:49
customers to see if they felt the same
00:31:51
pains
00:31:52
and resoundingly they did
00:31:55
our customers want AI That's truly
00:31:58
conversational
00:32:00
but not just in a little shot bubble
00:32:02
they want AI generated responses but
00:32:05
only those that truly understand user
00:32:07
Behavior so that they can respond with
00:32:10
empathy and care
00:32:12
and our customers definitely know that
00:32:14
the shopping experience is vast and
00:32:16
complex
00:32:17
they want AI that can be applied across
00:32:20
the customer Journey
00:32:23
with these learnings beeping hard at
00:32:25
work to build something new and are
00:32:27
delighted to announce alcoholic's
00:32:30
generative AI e-commerce framework
00:32:32
giving you the power to add the magic
00:32:35
and the light of AI in your shopping
00:32:37
experiences
00:32:39
so what's in this framework
00:32:41
well algolia's generative AI e-commerce
00:32:45
framework comes with a new set of UI
00:32:47
libraries that rapidly integrate
00:32:49
generative AI into your shopping
00:32:51
experiences
00:32:52
it uses llms instructed and tuned by
00:32:56
your own algolia apps data and events
00:33:00
as a framework it's composable by Design
00:33:03
which means that you as a developer can
00:33:06
Unleash Your creativity and inject the
00:33:08
light and solve for challenges across
00:33:10
your shopping experience
00:33:13
and while it is a great creative sandbox
00:33:15
it'll also come with patterns and
00:33:17
examples that allow you to quick start
00:33:19
your generative AI experiences
00:33:23
so I'm sure at this point you're
00:33:24
wondering that sounds great but what can
00:33:27
I build with it
00:33:29
every e-commerce journey is different
00:33:31
and we believe that with this framework
00:33:33
you can have the power of generative AI
00:33:35
across many spaces across this journey
00:33:39
this framework comes with two key
00:33:41
Concepts that allow you to start an AI
00:33:43
session or receive an incoming message
00:33:46
and together they can be used in some
00:33:48
really powerful and creative ways
00:33:51
for example you can use this framework
00:33:54
to create a generative AI powered guide
00:33:56
that loads for Search keywords
00:33:59
this makes it easy for customers to
00:34:01
discover vast categories
00:34:03
so the next time your user searches for
00:34:05
home theaters or surfing gear you can
00:34:08
give them a guide that gets them on the
00:34:09
right track
00:34:11
you can also use this framework to
00:34:13
refine and improve your users searches
00:34:16
for example if your user searches for
00:34:19
something that might be a little weak
00:34:20
such as baking appliances
00:34:23
generator AI can ask helpful questions
00:34:25
via query refinements that can make the
00:34:28
search more specific by asking things
00:34:30
such as what's the size of the appliance
00:34:32
you're looking for or do you have a
00:34:33
specific color or what are you looking
00:34:35
to bake out of it
00:34:38
and finally I'm sure you're wondering
00:34:40
can it build chatbots
00:34:42
and it sure can
00:34:43
with this framework you can message your
00:34:46
customers in intuitive ways that truly
00:34:49
recognize their behavior and needs
00:34:52
for example are you noticing a theme of
00:34:55
customers browsing your app and making
00:34:56
searches but not finding what you need
00:34:59
well with this framework you can program
00:35:02
a chatbot that identifies these users
00:35:05
and sends them a helpful message
00:35:06
directly
00:35:08
and really that's just the tip of the
00:35:10
iceberg with so little time today it's
00:35:13
hard to talk about all the things we can
00:35:15
build with this framework but that's why
00:35:17
we have a treat for you
00:35:19
later today
00:35:20
will be giving you a demo of the
00:35:23
framework in action
00:35:24
you'll be among the first to see the
00:35:26
future of generative Ai and e-commerce
00:35:28
apps
00:35:29
we're so excited to show you what we've
00:35:31
been up to and looking ahead we can't
00:35:34
wait to give this framework to every
00:35:36
developer
00:35:37
let's make some AI Magic
00:35:39
thanks for listening and back to you
00:35:43
thank you ayush please make sure to drop
00:35:46
into our principal engineer Sarah
00:35:48
Diane's talk for a full demo and
00:35:51
instructions to sign up to get access to
00:35:54
our conversational AI framework
00:35:58
okay folks we are coming to the end but
00:36:01
before we wrap up I do have one last
00:36:04
announcement to make
00:36:08
there are emerging llm stacks for llm
00:36:12
Native applications that are farming
00:36:14
every day you can see on the screen
00:36:16
there are innumerable ones popping up
00:36:18
and more will keep coming up
00:36:20
however one thing is very very clear we
00:36:24
need a way to store vectors which are
00:36:26
the output of a larger language model
00:36:28
and the answer today to that are
00:36:31
Standalone Vector databases
00:36:35
existing databases like MySQL or
00:36:39
postgres have very poor support for
00:36:41
Vector representations which are
00:36:43
basically very large floating Point
00:36:45
numbers they're also not equipped to
00:36:49
Traverse the vectors like how you would
00:36:51
in a SQL database that has rows and
00:36:53
columns
00:36:55
Vector databases on the other hand are
00:36:57
optimized for storing High dimensional
00:37:00
vectors in a graph like structure that
00:37:03
makes it easy to retrieve them when
00:37:05
presented with a query
00:37:07
they then use AI algorithms like
00:37:09
approximate nearest neighbor to find
00:37:11
vectors that are close together in
00:37:13
concept and return those results when a
00:37:16
query is passed
00:37:19
but Vector databases don't scale well ad
00:37:23
production workloads they are
00:37:24
prohibitively expensive there is a joke
00:37:27
on Twitter that I read that Vector
00:37:29
databases are great till you get your
00:37:30
first bill
00:37:32
next they are terrible at handling crud
00:37:35
operations so if you have data or
00:37:37
content that is changing Vector
00:37:39
databases become very slow
00:37:42
Vector databases were built to do one
00:37:44
thing to store vectors which means they
00:37:47
will never become a primary data store
00:37:49
as they cannot handle typical
00:37:51
transactional database needs with
00:37:53
asset-like properties
00:37:56
all of that means you have yet another
00:37:58
data store you need to maintain and
00:38:00
manage in your stack this requires
00:38:03
duplication and synchronization which
00:38:06
means more operational overhead
00:38:09
here is a super simple representation of
00:38:11
how a vector database might fit into
00:38:13
your stack as you can see you now need
00:38:16
to introduce new services to manage the
00:38:18
vector database and operationalize it
00:38:20
which all leads to more complexity in
00:38:24
your application
00:38:25
[Music]
00:38:26
so what if your current database could
00:38:31
already do Vector search
00:38:34
today I am extremely excited to announce
00:38:38
algolia's neural inference service this
00:38:41
will enable any data store to have
00:38:44
Vector search natively inside it
00:38:48
it is built on algolia's neural hashing
00:38:50
technology and it will enable fast
00:38:53
Vector search in my sequel in postgres
00:38:56
and sqlite and many many more use the
00:38:59
data store the query language and the
00:39:01
logic you already use without having to
00:39:04
shift your data to a vector database
00:39:06
duplicate your data somewhere else and
00:39:08
manage it somewhere else
00:39:10
in half of the talk I'm going to pass
00:39:13
you over to Hamish or gilby our VP of AI
00:39:16
who is going to do a live demo of
00:39:18
vectorizing and querying a traditional
00:39:20
database over to you Hamish
00:39:23
thank you tame is here to talk to you
00:39:26
about doing AI search from within your
00:39:28
existing database so that's not
00:39:30
requiring a new database but actually
00:39:32
doing it from within whatever you're
00:39:33
using today so this example is using
00:39:36
sqlite but in the session later on we're
00:39:38
going to run you through how you can use
00:39:40
our goalies neural hashes to make any
00:39:43
database or data store Vector search
00:39:45
capable so jumping into
00:39:48
the sqlite database here just want to
00:39:52
show you we have one table in here
00:39:54
called products so you can see here
00:39:57
you've got a products table you can come
00:40:00
in and see how many products are in here
00:40:02
there's about 21 000 products so no no
00:40:05
neural hashes or anything in here at the
00:40:07
moment but what we're going to do is
00:40:10
we're going to go and add them so
00:40:13
I have a script here
00:40:15
um this script basically is going to
00:40:17
take that the the database it's going to
00:40:20
look at the products table and it's
00:40:23
going to index it and it's going to do
00:40:25
that using the algolia inference API
00:40:28
which basically means it's going to send
00:40:30
the text for the fields and the API is
00:40:33
going to give back hashes which are then
00:40:35
going to be inserted into the database
00:40:37
so in this case it's using these fields
00:40:40
from the record as well and we'll go
00:40:43
more into that later but basically I hit
00:40:45
the button here and now the script is
00:40:47
running in the background so if we want
00:40:51
to see how it's progressing I can jump
00:40:53
in to the the log here and watch it
00:40:56
running so you can see it's adding about
00:40:57
100 a second or so and this will
00:41:00
actually speed up over time as it auto
00:41:02
scales as well but I'll quit that for
00:41:05
now and come back into the database and
00:41:08
so interesting thing here if I was to
00:41:10
look at the schema now
00:41:12
you'll see that along with the original
00:41:14
products table
00:41:16
we have two additional tables and these
00:41:19
have been added to create the neural
00:41:21
hash indexes which is what enables you
00:41:23
to run AI search in your database so
00:41:26
that's pretty cool and you can actually
00:41:28
see the progress
00:41:31
by checking how many things have been
00:41:34
indexed within the um the table so far
00:41:36
so if I run that again you'll see that
00:41:39
it's about three and a half thousand
00:41:41
have been added now of the uh 20 000
00:41:44
that are in this set so
00:41:46
um
00:41:47
I'm gonna add an extension here because
00:41:49
sqlite doesn't support the uh bit count
00:41:52
uh function natively so we add an
00:41:54
extension in to do that other databases
00:41:56
do so this is not something that you
00:41:59
need to do everywhere but for sqlite you
00:42:02
do
00:42:03
and then so now what we're going to do
00:42:05
is we're going to run a query this query
00:42:09
is coffee machine espresso with the milk
00:42:13
thing
00:42:14
here which is a bit of an odd query but
00:42:17
the purpose of doing this and boom we
00:42:20
have results purpose of doing this is to
00:42:22
actually show that you can use a a
00:42:25
textual style query that would not work
00:42:27
well with a keyword style search but
00:42:29
you're going to be able to get good
00:42:30
results from from the database using
00:42:33
that and other things to note here that
00:42:36
ran in five milliseconds
00:42:39
and this data set's too small to
00:42:42
actually see performance because each
00:42:43
query has an overhead and sqlite but
00:42:45
you'd probably find that that actually
00:42:46
took like a few microseconds to run it's
00:42:49
very fast uh in this case though looking
00:42:52
at the query we can take a look and see
00:42:54
what it's actually doing we are taking
00:42:57
the hash here and we're comparing it
00:43:00
against all of the hashes in the product
00:43:03
hashes table
00:43:05
we're going to pick the ones with a
00:43:06
score greater than 0.6 order by the
00:43:09
score limit by the top 10 and then we're
00:43:12
going to join it onto the products table
00:43:14
to return the results and
00:43:16
even though we haven't finished hashing
00:43:19
this table you can see here that these
00:43:21
results are indeed coffee machines and
00:43:25
there's a couple here that even have the
00:43:27
um the milk you can tell they have like
00:43:31
erytina milk device so we can run that
00:43:33
query again in your profile that results
00:43:35
will have updated and they're slightly
00:43:38
different now and that's because in the
00:43:39
background we've actually added more
00:43:42
hashes to each of the products and so
00:43:44
come in here and check we've now done 13
00:43:48
437 13 700 so you can see this in you
00:43:53
know a minute or so you've been able to
00:43:55
take the algola inference API and turn
00:43:59
your entire product catalog into AI
00:44:02
searchable within sqlite pretty cool
00:44:05
concept
00:44:06
really looking forward to showing more
00:44:08
of this in the session later but it's
00:44:10
back to you for now Brad thank you
00:44:14
thank you Hamish we are extremely
00:44:17
excited to announce algolia's neural
00:44:19
inference service and we're opening it
00:44:21
up to you for Early Access you're
00:44:24
hearing this for the first time at
00:44:25
algolia's developer conference please
00:44:28
sign up and we will be in touch with you
00:44:30
and how you can try it
00:44:33
you were using algolia for search in the
00:44:35
pre-ai world get ready for us to be a
00:44:39
part of the AI stack of the future
00:44:42
I want to thank you all for being here
00:44:44
you have known and loved algolia for
00:44:46
years you've prayed us for our speed
00:44:49
relevancy and great developer experience
00:44:52
we know we're a part of many developers
00:44:55
tool belts when they need to act search
00:44:57
to their applications we are also aware
00:45:00
that AI is finally here and we're making
00:45:02
it accessible to everyone a new era of
00:45:05
build possibilities is opening up I wish
00:45:09
you a very pleasant developer conference
00:45:12
would love for you to dig into more
00:45:15
topics that I only scratched upon and we
00:45:18
will see you in September at Dev bit
00:45:20
thank you everyone
00:45:22
[Music]
00:45:35
thank you
![Allah and the Cosmos - CREATION IN SIX DAYS [Part 1]](https://ruhnrawpgxrzdrgaohnf.supabase.co/storage/v1/object/public/images/video/Or7pwR60ilE/thumbnail.jpg)
