Algora | Algora TV

Michele Riva, founder of Orama - Open source interviews #35

Algora

Watching now

00:00 Hey everyone, welcome to the Open Source Founder Podcast.

00:02 Joining me today is Michele Riva, co-founder and CPO of Orama,

00:07 an open source full-text and vector search engine.

00:11 Michele, thank you so much for joining us.

00:13 I'm super excited to start with some information about your background

00:16 and how Orama got started.

00:19 Yeah, thank you for having me, first of all.

00:22 I'm a big fan of your project, so it's a real pleasure to be here talking with you now.

00:27 Appreciate it.

00:28 I've been working for around a decade right now as a software engineer professionally.

00:34 And in the last past few years, I've been involved into open source and public speaking and knowledge sharing.

00:44 to open source and public speaking and knowledge sharing.

00:47 So I've been very active in the community for creating open source projects, contributing to others.

00:51 But most importantly, I was working at Nearform,

00:55 which is a very important company for me.

00:57 And just for the audience to know,

00:59 Nearform is one of the few companies

01:02 that have full-time people working on Node.js.

01:05 So it's truly committed to open source.

01:08 And I was part of the developer experience team,

01:10 which is the team that works on Node.js specifically.

01:12 I wasn't working on Node.js myself, but the rest of the team was.

01:16 I was more concentrated on TC39 for certain things.

01:21 So I was basically creating some polyfills

01:25 for newer versions of JavaScript.

01:27 And I was more interested in building new solutions

01:31 rather than working on existing ones.

01:33 I know that's very typical for engineers,

01:36 but that's how it is.

01:38 And I've been lucky enough to work in this company

01:40 for almost a couple of years.

01:43 And for a big conference which is called

01:46 we are developers in berlin in 2022 if i remember correctly i wanted to create um something cool so

01:54 i created a full-text search engine in javascript and it was you know very sketchy it didn't work

02:01 very well but it was very fast um the reason why it was fast is, first of all,

02:06 JavaScript is fast. A lot of people say, no, it's not performant, but that's false. It's pretty

02:12 performant if you know how to write it. And I was lucky enough to be in a team where people deeply

02:18 cared about performance. So they could mentor me and teach me a lot of stuff on JavaScript

02:23 performances, Node.js performances.

02:27 So it became very fast. It was in the microseconds area for most searches on millions of records in JavaScript again.

02:33 And after I gave the talk in Berlin about this full-text search engine,

02:38 which was more a talk on algorithms and data structures in JavaScript,

02:42 someone put it on Hacker News.

02:44 It went viral.

02:45 And eventually, the boss of the company,

02:48 the founder of Nearform came to me and said,

02:53 okay, look, we are a consultancy firm.

02:55 We can bring this as a product.

02:58 So I'm going to help you and I will connect you with the people

03:00 to create a company out of this.

03:02 So thanks to Kian, the founder of Nearform,

03:06 I have been able to skin off my own company.

03:10 And yeah, that's how everything started, basically. It was just a little toy and eventually became a proper company.

03:15 And it's still in the process of becoming a proper company after almost one year now.

03:23 I love it.

03:24 Well, thanks so much.

03:25 Thanks so much for the background.

03:26 And, you know, in one of your conference talks, I saw that you said there's no slow programming

03:31 languages, just bad algorithms and data structure design, right?

03:36 Yes, that's the point.

03:37 And so I particularly like that quote.

03:41 And I saw you mentioned how basically, you know, you're a fan of Elasticsearch, but of course, you know,

03:46 it's hard to deploy.

03:48 The memory consumption is hard.

03:49 And then who wants to work with Java?

03:52 That's another key point there.

03:56 And then, you know, there's Algolia, but of course,

03:58 like at scale for an enterprise, it's way too expensive.

04:00 It's kind of like a black box.

04:02 You can have a little tinker with it and extend it.

04:04 So now moving to Oramaama which is solving these problems and of course achieves like performance

04:09 that is like incredible I would love to hear what the journey looked like in 2023.

04:15 Yeah so first of all just for you to know that's a fun one I'm a huge Elasticsearch fan I want this

04:21 to be very clear I sent my resume like twice to Elasticsearch. They

04:26 never replied. So I took this very personally

04:28 and I created

04:30 a company on an Apple

04:31 engine. I'm kidding, of course.

04:34 But it was

04:36 fun to mention. I've learned

04:37 a lot, of course, by

04:39 working with Algolia firsthand.

04:42 Sorry, Algolia and

04:43 Elasticsearch firsthand.

04:45 So, of course, I do still appreciate those search engines,

04:49 even though they are competitors now.

04:51 But I still love them because they're great.

04:54 So with that being said, we started the company in 2023,

04:58 and we hired the very first engineers to work on something very different. So Elasticsearch and Mailysearch and also Minisearch, which is a very nice project in

05:12 the JavaScript ecosystem.

05:14 It's a JavaScript full-text search engine.

05:16 Think of Fuse, Lunar, et cetera.

05:18 They have something in common.

05:21 They only run on the browser, only run on the server, but

05:27 they can't run at the edge. So that was something cool we wanted to make happen with Orama. So

05:32 the first thing we did was to take the open source Orama, which is the one you can get on

05:37 github.com slash Orama search slash Orama, and put it at the edge. So we are now basically,

05:44 the entire year we spent in building a system,

05:47 a distributed system that basically deploys your index,

05:50 even larger ones, to 300 global locations.

05:54 And the advantages of running something at the edge

05:57 is that you get virtually infinite scaling,

06:00 meaning if I am doing Black Friday, for example,

06:03 and I have an e-commerce and I get a spike

06:05 of requests, I don't have a server to manage. So it's truly serverless in a sense. So you get

06:11 billions of requests per second, no problem. CDNs were meant also to propagate DDoS's

06:17 tracks. So we are cool with that. It's also very performant. We have big limitations in terms of CPU time, for example.

06:27 So we can perform most searches in the microseconds area and we can aggregate the results to the browser and you get the data back from the server,

06:47 has basically very limited latency because there is a node that is always close to you.

06:50 Like we deploy in 100 countries, like 300 cities worldwide.

06:55 So, you know, wherever you are in the world,

06:57 even mainland China, for example,

06:59 which is not super accessible with AWS, et cetera,

07:02 like not every provider has a server that we got it.

07:06 And even if you connect from there,

07:08 you have a server that is very close to you.

07:11 So the time that the request takes to get back to you,

07:15 it's very limited.

07:17 So it's very fast.

07:18 It's very scalable, but also very cheap

07:21 because the price, it's basically the same as fetching an image at the edge,

07:28 you know, on a CDN, which again,

07:30 it's meant for serving billions of females.

07:32 It's a very nice model, pretty difficult to program with.

07:36 I gotta be honest.

07:38 But yeah, this is what makes edge applications

07:41 a bit more appealing to me than, you know,

07:44 standard servers, for example,

07:45 server-side applications, at least, or browser-side applications only, like the ones I mentioned that

07:50 are fantastic products, but they could potentially run there, but to make them run on the CDN,

07:56 I experienced this firsthand. It's going to be very difficult.

08:00 Totally, and we'll see, you know, later on, people sticking around the demo,

08:04 actually, that performance in microseconds.

08:06 And I'm excited to get there later on.

08:09 How's the open source motion been for you and your team in terms of the community as well as managing the project?

08:15 And I know you're familiar from the past, but I'd love to hear what the experience with Arama has been like so far.

08:21 And whether there's any advice for other maintainers or open source founders. Yeah, so first of all, you know, I've been very happy with the fact that

08:31 we made entire releases without the core team working on those releases. Like, community-first

08:37 releases. People like wanting, I don't know, Vietnamese language support. I don't speak

08:41 Vietnamese and it's really difficult for me to, you know, build and test proper stemmers, removing proper stop words, et cetera. But we got people getting

08:50 there and saying, yeah, no problem. I can do that. So they just do that for us. Sometimes we give

08:56 them a bounty. Why not? But, you know, other times they just show up and they want to implement

09:03 something that is missing and they do that. And again again i'm very happy to be able to make entire releases without having theorama

09:10 team working on those releases because that's that's just fantastic um that's pretty unexpected

09:16 i wasn't expecting that like two years ago when i first started the project so that's cool and

09:23 again it's not a super accessible project gotta be honest it's very

09:28 technical and you know it's algorithms and data structures again plus a lot of optimizations

09:33 because if you want to run on every single javascript runtime nowadays you got to support

09:38 ban you got to support cloudflare workers dino browsers. So you really need to know JavaScript.

09:46 And there's no abstraction there

09:48 that can help you with that yet.

09:50 We should probably invent it.

09:52 But, you know,

09:54 it really teaches you

09:56 to think in JavaScript, which is very nice

09:58 in my opinion.

10:01 Absolutely.

10:02 And I think in the process of

10:03 using Orama, there's also like a test suite that actually people can test out their algorithms and data structures and learn in the process of building with Orama and become better engineers and better optimizing their systems.

10:15 So I really like that. And again, maybe we'll get to that later on.

10:18 You come from a functional programming background, working with Haskell, for example.

10:23 So how have you experienced these changes?

10:26 So that's very funny.

10:29 Yes.

10:31 I've been working a bit with Haskell

10:33 and a bit of Elixir.

10:38 So I got to say,

10:40 working at the CDN, for example,

10:42 is a functional model, in my opinion.

10:44 So Orama is an immutable database.

10:46 We treat it as an immutable data structure.

10:49 So every time you want to update Orama at the edge,

10:52 for example, you have to redeploy the index.

10:55 So in a sense,

10:57 I was mainly inspired by functional programming

11:00 in working that way.

11:02 There is an inversion of paradigm

11:04 when it comes to search in general.

11:07 It's not just full text, but also vector hybrid search and newer kind of searches that

11:14 ChargePT taught us that we need to implement.

11:17 And making everything running at the edge and making it, you know, in a functional

11:22 architecture actually made it possible.

11:25 So that's something that I kept from my functional background.

11:30 But there is a good quote I can give you from the, it was the V8 documentation, I think,

11:37 that said, like, if you want your JavaScript to be as fast as C, just write it as you write

11:43 C. So it's not functional.

11:46 I'm sorry.

11:47 Functional programming JavaScript, it's very fun.

11:50 JavaScript itself started as a functional language back then in 1995.

11:55 But nowadays, if you want to have cool performances, you

12:01 have to make two compromises.

12:02 First of all, write imperative code.

12:06 Second, if you really care about performance,

12:08 it's going to be pretty ugly.

12:11 But again, that's a compromise.

12:12 And sometimes you can't compromise with speed.

12:14 For example, if you want to execute your code

12:17 on Cloudflare workers, you only got 50 milliseconds

12:19 for executing the code.

12:20 So you see you can't compromise with performances there.

12:24 And sometimes functional programming makes it a bit harder

12:27 on JavaScript specifically to run that fast.

12:30 Then you have other programming languages such as OCaml, Haskell, etc.

12:34 that are faster, but for different reasons.

12:38 But that's not the case for JavaScript, sadly.

12:41 And Yatoram, of course, has the crazy performance.

12:44 And again, we'll get to that.

12:46 I'm curious how you have experienced the go-to-market motion for Orama which can be done

12:51 you know via outbound, can be done through marketing and the inbound interest and of course

12:56 you mentioned that a lot of people you know prior to funding the company had their eyes on you

13:00 through the Hacker News post that someone did on your behalf. It's amazing when that happens.

13:06 And so how did you experience last year on that front?

13:10 Was it mainly just building, building, building and letting the work speak for itself?

13:14 Or did you also have, you know, a quote unquote enterprise, like outbound sales motion to

13:20 get the feedback from big organizations who have such big scale needs with their search?

13:28 Yeah, so we definitely have some design partners that are helping us at scale.

13:32 So if you think of large enterprises, for example,

13:35 we have some large enterprises that are helping us in shaping a search engine

13:41 or as we call it nowadays, an answer engine, as they wish it to be.

13:46 And those companies are mainly looking for switching off, you know, a different engine

13:52 into a Rama. So this is really helping us because we can see the problems that we need to solve.

13:59 And we also acknowledge the scale at where we need to bring our product to. So that's

14:06 pretty cool. We spent the past

14:07 year basically building and it

14:10 was pretty hard because again

14:11 there

14:13 are no many applications running

14:16 at the edge and there's not a lot

14:18 of literature

14:19 on how to run

14:21 applications that are so

14:23 complex at the edge.

14:25 Because again, if you think of Cloudflare workers, for example,

14:28 your worker size must be like one megabyte,

14:33 but an index could be one gigabyte.

14:36 How do you solve that?

14:38 So we had to solve that problem, you know,

14:40 and it took forever, literally.

14:43 Like that's hard programming.

14:46 That's some hardcore programming going on.

14:49 And it wasn't easy.

14:50 We had some help from many people getting there.

14:53 And we have a team now that finally has been able to prove that this model is working.

14:59 But it took time.

15:01 So we are now ready to onboard some first customers and get back to those

15:06 enterprises that are helping us and and start finally to onboard their clients um so i can't

15:14 say a lot more about our strategy at this point we will probably announce something in in a few

15:19 weeks or months um but i won't be the person announcing this. I think our CEO and our marketing persons

15:25 are the ones that are best fit for this. I'm just a technical guy. I'll stick to the technical

15:31 parts.

15:33 A technical guy who goes around the world, of course, like presenting and educating people.

15:38 And that's super unique. And I really admire that, actually. And congratulations and keep

15:43 it up. I was also curious to ask

15:46 what might be coming up next. Is there a milestone you could share on the product front?

15:51 If you see, for example, the edits that we made to the

16:00 engine, until version 1.2.4 we were just a full-text search engine. Until version 1.2.4,

16:07 we were just a full-text search engine.

16:08 We had all the goods,

16:09 like face setting,

16:12 grouping, sorting.

16:13 We had it all.

16:15 Whatever you could do on other search engines,

16:16 you could do it with Orama.

16:17 Then we implemented vector search,

16:20 so you could provide a vector

16:22 and retrieve results

16:24 based on their semantic meaning.

16:26 And now with Rama 2.0, we have geo search.

16:29 We have hybrid search.

16:32 So we basically perform both vector and full text search, aggregate the results and understand if you're searching for something rather than something else.

16:41 And I can give an example of why this is important.

16:44 So that's a very silly example.

16:47 And I'm so sorry because I couldn't find

16:48 another greater example than that.

16:51 So if I search for a movie to watch

16:54 with my six years old daughter, right?

16:58 I'm clearly looking for something.

17:01 And I don't know how to express what I'm looking for

17:03 other than this must be a family

17:05 movie, you know. But if I search, for example, for a movie about a doctor that is possessed,

17:13 like I'm looking for the exorcist, you know, the movie. And that's totally different. Like these

17:19 results are totally different. I can give you results for The Exorcist if you're looking for a family movie. So this is

17:25 where Orama makes the distinction. If you're searching by meaning or you're searching by

17:31 keyword, we get this and we give you the right result every time. So Orama 2.0 has done that.

17:39 There will be Orama 2.1 and 2.2 and etc. and then Orama 3.0, I can't spoiler where we're going but we have a very nice

17:49 plan to get there. We will certainly enhance our AI capabilities. Right now for example with Orama

17:57 3.0 we also generate our own embedding models that are faster, cheaper, and closer to you than OpenAI.

18:06 So you can generate embeddings automatically with Rama, but we will see that in just a second.

18:12 And so we are moving a bit more in the RAG space where you have this way of searching that, again,

18:23 it's more about the meaning, not just the keywords.

18:26 And we want to provide you the best experience

18:28 either on the browser or the server or the cloud

18:31 or the edge.

18:33 You choose, we provide that to you.

18:35 Love it.

18:36 And I'm super bullish, I have to say.

18:38 Is there any mistake that, you know,

18:42 might have happened,

18:43 something you could urge other founders to avoid?

18:47 I know it hasn't been too long of a time

18:49 and it's going really well from the outside.

18:52 I'm just wondering if there's anything

18:54 we could share with people here.

18:57 I think it's not a proper mistake,

19:00 but sometimes I believe you should really define

19:04 your design partners.

19:07 And because eventually these are the users you're going to get.

19:11 And you should definitely always listen to your users.

19:16 So whenever you have a doubt, for example, are my users going to like this?

19:21 Do they want this in the first place?

19:24 The only real people you should ask is the users at this point.

19:28 So one mistake I can see is probably I would have loved

19:32 to learn that a bit earlier.

19:36 Even when that was an open source project

19:40 and nothing more. That would have simplified things so much.

19:44 Now we are in a situation

19:45 where eventually we

19:47 respect that rule, let's say.

19:50 It's a startup rule,

19:51 if you will.

19:53 But I wish I learned that

19:55 earlier. User

19:57 is the king and eventually

19:59 you're going to serve users.

20:01 Just build stuff for the users,

20:04 not for yourself or, you know.

20:09 Thanks for saying this.

20:10 And to follow up,

20:11 is there any specific process or heuristic

20:16 you use for prioritizing?

20:18 Because, you know, your design partners

20:20 and your users,

20:21 they might try to pull you in a different direction.

20:23 They might have all these different requests

20:25 and there's only so much bandwidth.

20:27 There's only so much time every day.

20:28 So how do you go about prioritizing?

20:31 And I know this is super challenging for everyone.

20:33 Just curious to hear your thoughts.

20:34 Yeah, so thankfully enough, the founding team,

20:38 so it's me and a CEO and a head of design.

20:42 So we are three people with three different departments,

20:47 you know, and we have three different perspectives.

20:50 So there is the business perspective,

20:52 there is the product perspective,

20:53 and there is the technical perspective.

20:55 That's awesome.

20:57 You know, in Y Combinator, there is the hacker,

21:00 the hustler, and the designer on the team.

21:03 Like, that's the perfect three kind of forming team.

21:07 And that's us.

21:08 Like, I really love this.

21:10 And I see why this is such important

21:13 because those three perspectives really helps you understand

21:16 what you need to prioritize, for example.

21:19 And most importantly,

21:21 to understand the point of view of different users.

21:23 Like I can definitely understand the point of view of a user that is a developer, but I can't really understand fully a marketing person or a business person or anyone else that is not deeply technical and needs a solution that is not super technical.

21:41 So having, again, multiple points of views, it's really helping.

21:46 And then I got to say, we have a part-time project manager that really, really, really

21:51 helps.

21:53 So having a third person, you know, third party that comes and say, okay, we really

22:00 need to prioritize and we prioritize with some strategy, with some structure.

22:06 really need to prioritize and we prioritize with some strategy with some structure at a certain point of a lifespan of a startup it's becoming you know pretty handy so i'm really happy about

22:12 that that's great that's great thank you thank you so much for uh for sharing this do you have

22:16 uh any any conferences that you're planning to attend the next month throughout this year

22:21 anything planned not yet I got a couple,

22:25 but they're not announced yet.

22:27 So I can't really tell.

22:29 But also I'm moving a lot.

22:31 So it's going to be a bit more difficult this year

22:33 because I really want to concentrate on

22:35 working, working, working for this year.

22:38 Maybe in 2025,

22:40 I'm going to move a bit more,

22:41 but we'll see.

22:42 We'll see how it goes.

22:44 That's great.

22:45 Are there any other notes that you think would be helpful

22:48 for people to hear before we proceed to the demo

22:51 that I'm super excited about?

22:53 I think you should all put a star on GitHub.

22:56 That would really, really help if you could.

23:00 And I'm going to show you why in just a second with the demo.

23:04 Absolutely, absolutely.

23:05 And actually, I guess I just thought

23:07 I think it would be useful to ask you this

23:09 before we go to the demo.

23:11 If you were starting today as a developer, right?

23:14 Maybe you're in college, maybe you're going out of college

23:16 and you want to start your career.

23:19 Is there any rule of thumb you would follow

23:22 or is there something that you could recommend

23:24 to people who

23:25 are just getting started and would like to work in tech and be a professional software engineer

23:30 oh that's hard i know this tough question and kind of generic no but it's like it's it's a

23:37 cool question and and i think every developer has a different answer so i'm just giving you mine

23:44 there are multiple levels

23:46 of you know a career that you can achieve in software engineering i truly believe that if you

23:53 if you want to see a very long trajectory going on so seeing yourself progressing in your career, you really have to invest a lot of time early in your career.

24:07 So the first two, three, four, five years, you really have to work at night. And I'm sorry

24:14 in saying that. I do understand that everyone has a life. I do understand that work is not

24:19 everything. But this is a very hard job and there's no college,

24:25 there's no university that it's going to teach you how to do that job for real.

24:33 I've been working with graduated engineers for a long time.

24:37 They are super smart, incredibly talented people and educated people,

24:43 but I can't tell the difference you know with

24:47 self-taught developers that really spend their nights early on their careers trying to learn

24:53 something more about programming so with that being said i'm not saying that college and

24:58 university is useless because it's not like if i could go back i would do college and university

25:03 Like if I could go back, I would do college and university.

25:06 But that's not for learning how to program. That's for knowing the unknowns that I'm facing today.

25:10 Because my problem as a self-taught engineer,

25:13 I couldn't go to college because I didn't have money and I needed to work.

25:18 My problem is that I'm facing a lot of unknowns, unknowns.

25:22 Like I didn't know that I had to learn this.

25:26 I didn't know that I had to, you know,

25:28 go look into this other stuff.

25:30 And I have this feeling that people that studied

25:33 have a different approach.

25:36 They have a method to follow

25:37 whenever they face this kind of unknown.

25:40 And they probably have a broader sense of the unknowns

25:44 that they are going to face moving on a project,

25:47 moving on a career, et cetera.

25:49 And this is very valuable.

25:52 Again, sometimes it's not a problem for me.

25:55 Other times it is.

25:57 But if I could go back in time,

26:00 I would probably study a bit more

26:01 because sometimes it's really frustrating.

26:04 But this is not preventing you

26:06 from getting a very nice career in computer science

26:10 or in programming in general.

26:13 I love these insights.

26:15 Would you say, just a last follow-up question here,

26:17 would you say that the skill set

26:19 that engineers need to build nowadays

26:21 has been expanded to also include building in public, doing media, doing marketing.

26:27 And is that becoming at all a necessity, quote unquote,

26:31 or like a very nice to have?

26:35 That's the tip of the iceberg.

26:37 You know, you see developers doing marketing,

26:40 you see developers doing content and conferences

26:42 and being influencers on Twitter.

26:44 But again, this is the tip of the iceberg.

26:46 This is like the 0.00001% of developers worldwide trying to express themselves.

26:55 And they're doing this pretty well, if you ask me.

26:58 But the vast majority of developers are not interested in having a career that goes you know into content

27:07 creation or in building in public in creating startups or being you know ctos managers of other

27:14 developers which is fine like that's totally cool um i think it's not necessary for everyone

27:22 to get on twitter and get a,000 likes on every post.

27:27 It's pretty useless if you don't care about that.

27:30 It's a pretty useless metric.

27:33 But I do also understand that sometimes having a good network, it's really helping you in your career.

27:53 So what I would recommend is, I said previously, you should really study during the night early in your career, but during the day, just make some connections.

28:05 So you only got 24 hours, make sure to also grow your networks and your connections, because this is what it's to be life-changing for you rather than your technical skills like your technical skills are going to be just fine you

28:09 might not be the strongest in algorithms but if you are a pretty good one in relationships

28:15 this is helping you a lot more absolutely absolutely and working in open source contributing

28:21 to open source can definitely help there especially for younger people who don't have that professional setting

28:25 experience working with teams.

28:27 Well, with open source, you can contribute to production software

28:30 and work with other people and get mentored in the process.

28:34 And just as a middle way in terms of kind of like creating also content

28:38 in some sense, I think a middle way would be to learn in public

28:42 or to have your own blog where you have the teal, today I learn kind of like document your journey do it for yourself as a journal and

28:50 maybe that's you know that can exercise that muscle a little bit uh should you find yourself

28:55 in a position later on where you'd like to express those talents even more uh definitely not a

29:00 requirement as you said uh just another avenue for people to explore their talents.

29:06 Absolutely.

29:07 Thanks for the insights here.

29:09 And unless there's anything else you would like to add,

29:13 maybe you can jump to the demo,

29:15 maybe start with showing the repository on GitHub

29:18 and then the demo for Amon.

29:21 Yeah, sure.

29:21 Let's start with the demo.

29:23 I will probably share my screen. Let me see. Right. Let's start with the demo. I would probably share my screen. Let me see.

29:26 Right. I will start with the repository. So in case anyone wants to add by putting a star,

29:33 this is the repo. Everything is open source. We got a bunch of great contributors and,

29:38 oops, sorry. And this is where you can find all the documentation and the highlighted features for Rama.

29:46 So whatever you need, just click and get to the documentation.

29:49 Everything should be documented.

29:51 And if it's not, I'm responsible for that.

29:52 So just ping me on Twitter or Slack.

29:55 And I will do my best to try to fix this.

29:59 All right.

30:00 But let's go back.

30:03 If you want to start with the demo, I want to be honest, I totally forgot about the demo.

30:09 So we're going to make something. No problem.

30:13 All right. So let's start with the Rama Cloud. I'll be signing up with GitHub.

30:19 All right. This is the dashboard. This is where you can create and deploy indexes. You can upload files.

30:25 You can connect to e-commerces.

30:26 And of course, we will be supporting multiple connection going on.

30:31 Or you can create different data sources based on Webhooks, REST APIs, or even DocuSaurus or other platforms that eventually we will support.

30:41 There is a nice analytics page, which is coming soon with a survey.

30:46 I would really appreciate it

30:47 if anyone could compile that survey.

30:50 And in developers tools,

30:52 this is where you can put your OpenAI API key

30:54 and we will see why you might need this later.

30:57 I've already inserted mine previously

31:00 because I didn't want to share it publicly.

31:02 I don't want to go bankrupt because of that.

31:05 So once you insert it, you can never see it again.

31:08 The only thing you can do, you see this is text.

31:11 You can just replace it with another one and that's it.

31:15 So let's start.

31:17 I'm going to show you the data set.

31:19 So this is going to be the data set.

31:21 It's going to be a thousand video games with title, description,

31:27 rating, which is a floating point number, and genres, which is an enumeration of genres. And

31:35 we are going to index this and deploy this at the edge. So let me show you how it works.

31:40 Let's create a new index. Let's call it argora-demo. Description is optional.

31:48 Let's select JSON file because we're gonna use JSON file and let's create the

31:52 index. So now we can drag and drop the file here and we will see a preview. This

31:58 is gonna help us in creating the schema. So if we create the schema, we should say title is of type string.

32:07 Description is of type string, again, but rating is of type number. Genres,

32:17 I have no idea how to pronounce this word, so I'm very sorry. It's of type enum, and it's going to be an array. You can also auto-detect the schema

32:26 if you wish, but for the purposes of the demo, I'm showing how to properly write it. What's

32:32 important to know is that you only should put data here that you mean to search through. So

32:38 if you don't want to search through title, for example, just remove it it only put the data you want to search through or filter

32:45 through so we want to search through the entire data here we will enable the automatic embedding

32:52 generation so we can select either openai or orama for the purpose of this demo i know a lot of people

32:59 is using openai already so i'll be just using OpenAI. We can select which properties to consider for

33:07 embedding generations. Right now we only consider string properties by default so

33:11 add all description. Yeah we'll select both. Save and deploy. So this is gonna take somewhere between

33:21 one minute and three minutes and the reason why's going to be so slow at the beginning

33:26 is because we have to calculate all the vectors

33:29 and we're going to insert all the vectors in our vector storage.

33:33 So we have to rely on OpenAI.

33:35 So for every single document, we have to call OpenAI,

33:38 get the response, put it in the storage, et cetera.

33:42 From the second time you're going to deploy them,

33:44 of course, there will be a cache.

33:46 So it's gonna be a lot faster

33:47 and you should expect such a small data set

33:50 to be online in 30 seconds

33:52 in 300 global locations worldwide.

33:55 You can always see the status by looking at the logs.

33:59 And as soon as this finishes,

34:01 you will see the endpoint and the API key

34:04 that you need to make requests.

34:06 So we'll do one thing. I'll go on Visual Studio Code. I already created a vid project, so we're trying to do that in like 10 minutes.

34:18 We'll see if we can do that. Let's install a couple of dependencies so first of all ban because i really like ban

34:28 i want to be honest so we will be using ban and i will install orama slash orama and then

34:37 so we will install orama and orama cloud client let's see what are they about

34:48 client. Let's see what are they about. So Orama, of course, is going to be the Orama package itself,

34:51 but then we also have Orama Cloud client.

34:58 So this is going to be the client that we're going to use for searching on Orama Cloud.

35:05 The API are the same that we're using for Orama open open source but this time we are searching on the cloud.

35:11 In the meantime we can see that this is proceeding in inserting everything and that should be done now.

35:20 So let's go back for a moment in our terminal. All right we installed the dependencies. We can now go back here. I can't really remember how to start a project. So, okay, that was easy.

35:27 Ban, run, dev.

35:28 All right, this is what we got.

35:30 Let's go back.

35:31 This is a very fresh installation.

35:33 So we'll just go back here, remove everything.

35:37 We don't want CSS.

35:38 We want nothing.

35:40 My co-founder is the head of design, so he's going to kill me.

35:43 I'm very sorry.

35:45 I won't do anything very good.

35:49 All right, this is what we got.

35:51 And yes, we also have finished inserting everything.

35:54 So this took a couple of minutes, as you can see here.

35:57 Two minutes, exactly.

35:58 And we got the API endpoint and the API public key.

36:02 So let's go back here.

36:03 Let's create a new folder.

36:06 I'll zoom in a little bit. I call this lib where we're going to use all our libraries. Let's create oramma.ts

36:14 and now let's import client. So we can import a couple of things,

36:21 Oramacloud or Oramaproxy. I don't want to talk about oramaproxy right now,

36:26 but we will see why this is interesting a bit later. So import it and export it. Export const

36:34 oramacloud. It's equal to new oramaclient. Oops. So a couple of parameters, API key and endpoint.

36:45 So, the endpoint is going to be the API endpoint.

36:51 And the API key, that's public, so you can share this on the frontend, no problem.

36:58 API key, there it is. We can start the heartbeat with a frequency of, say, three seconds and a half.

37:11 So this is basically keeping Orama awake because it goes idle mode after you don't use it. So

37:18 you just want to keep it awake. We don't integrate this built-in because sometimes you don't want

37:24 this, but

37:25 there's no time maybe now to discuss this. Let's just leave it for granted that we want it.

37:30 So let's go back here and import our Orana instance.

37:36 Then let's import, and again, I'm sorry React developers, I'm gonna...

37:42 I'm not a fondant there, so I'm going to create something that is not super optimal,

37:47 but import useDate from React and maybe useEffect.

37:53 We're going to use that later on.

37:55 Let's create search result and set search results.

38:00 Okay, good. Great, Copilot, thank you.

38:02 Okay, good. Great, Copilot. Thank you.

38:10 We're also importing a type, which is nullable from Orama.

38:14 So nullable, it's basically telling you that a value could be null.

38:15 That's 3D.

38:21 We can also import results, which is the type of the search results. And you can determine the type going back to your index.

38:25 This is basically the type so you go back here and you say type

38:31 search result result it's going to be oops it's basically going to be this.

38:40 The num type is not a JavaScript sorry a TypeScript type we're going to use now. Just

38:45 treat it as a string. Rating is going to be a number.

38:50 And description is going to be a string again. So that's really it. So now we have, let's call it

38:57 type search results state, which is nullable,

39:07 and we can pass searchResult. Cool.

39:08 Now let's just create an input,

39:13 and we're gonna create a second state,

39:14 which is searchTerm,

39:17 setSearchTerm, which is a string.

39:20 And we now go here and say, all right, placeholder,

39:28 search, not a movie, something.

39:30 All right.

39:33 Then value is going to be search term.

39:35 And on change, we're going to, oh, my God, I love compiler.

39:36 Can I say that?

39:41 That's so, all the boring stuff, it's automated now.

39:42 I love this.

39:43 Cool.

39:48 So now if we go back here we will see hello cool

39:59 let's maybe remove the css from there all right again maybe i can zoom a little bit okay So go back here. Let's make some search. All right.

40:06 So useEffect.

40:12 We are going to spawn this effect every single time the search term changes.

40:19 And this is going to be the search term.

40:22 So like if I search, for example, for, I don't know, Pokemon, you know, this is going to be the search term. So like if I search, for example, for, I don't know, Pokemon, you know, this is going to

40:27 be the search term.

40:29 Then we have mode.

40:30 By default, this is full text, but I can also search for hybrid and vector.

40:34 These two are only enabled if you either provide embeddings on your own or you generate embeddings

40:41 with Rama, just like we saw previously.

40:43 So in that case, just limit to, let's say, five.

40:47 And you can paginate by using offset.

40:50 We're not going to do that right now.

40:52 Then, this is a promise.

40:57 So we can just say set search results.

41:01 And we can catch any error with console error.

41:03 I know this is not great.

41:04 and we can catch any error with console error. I know this is not great.

41:12 Yes, I'm probably... I need to specify this because of TypeScript and...

41:19 All right, let's keep it as is for a moment. Now we got search results.

41:24 Let's see, if we have search results, let's just...

41:27 I'm sorry, I think this is going to be like this.

41:31 So results, basically.

41:34 I forgot to use the result.

41:36 As you can see, this is the API response from Orama.

41:41 So it's going to give you the count.

41:42 So if you search for Pokemon, for example, you have 1000 results. I'm going to limit this by 10 results per page, but I'm also giving you

41:51 the full amount of results inside your index. The hits, so the results themselves, how much time

41:58 has elapsed, and face sets in case you wanted the face sets and groups if you're grouping the request. So now we got

42:07 search results oops dot hits dot map like that as a key we can use result dot ed and then we can say

42:29 and then we can say result.title. So you see this is inheriting basically the search result type we just used.

42:35 We can probably also say div and make another div for result.document.description. Hopefully that should be enough. default document dot, I don't know, let's say description.

42:47 Hopefully that should be enough.

42:49 All right.

42:51 All right.

42:55 So by default, I'm going to, if you don't specify search string,

42:58 I'm going to return to you all the documents.

43:04 Of course, you can block this by saying just return. Right?

43:05 So you're not making any requests if there's no search,

43:09 which makes sense.

43:11 Let's maybe use H2 for this.

43:13 I know semantically it's not correct.

43:17 But I can go back here and say, Pokemon, this is working now.

43:20 And I can search for Guitar Hero maybe.

43:29 is working now and I can search for guitar hero maybe, but you will see there is something that is not going well.

43:30 Like if you search for guitar, you got guitar hero.

43:35 This is because by default, me give a threshold of 0 or 0.2.

43:50 By default, it's 1.

43:51 We're going to set 0.

43:52 That means, as you can see now, it works.

43:55 That means I only want results that has both of the search term in it.

44:02 of the search term in it.

44:06 So if I search for Guitar Hero, for example,

44:10 the search result, either the title or the description has to have both.

44:12 And I can have Guitar Hero and Hero Hero.

44:14 That's fine.

44:16 By default, we don't do that because you can customize this.

44:19 By default, you could say also like 0.5,

44:23 which means return all the results that has

44:28 guitar and hero plus 50% of the results that only has one of those. If you have few results that

44:35 could be handy, you see why. But this is not the case so let's just do that. So that's basically it and I can also say all right but I want them where in that case

44:48 rating is equal to or greater than three. So I can also say document, then result, document.rating.

45:06 And now we should only have search results where the rating is greater than or equal to three.

45:13 So let's try. Let's refresh.

45:17 Yes, that's correct.

45:20 You see, we basically don't have, if I search for A, for example,

45:24 we shouldn't have results that have more, sorry, less than three.

45:29 And I can also say, like, you see now, these are all three.

45:33 Basically, we can go back and say, all right, let's do with four.

45:38 And we go back.

45:39 That's now four.

45:40 And I haven't changed my search query.

45:41 and I haven't changed my search query.

45:45 We can do the same, of course, by saying only give me the, oh, sorry, I forgot to put here H3.

45:53 Result, oh, because I, all right.

45:55 This is not a string.

45:56 This is an array of strings.

45:58 So I can also say only give me the adventure.

46:01 So where genre is equal to adventure or includes adventure

46:06 and visual novel so you can go crazy with all the filters you want and this is how we basically

46:12 create an index on rama i believe that's fairly simple but i developed that with my team so

46:19 i'm looking for feedback here like if you want to join as a as an early adopter and try this out, I'd be very glad

46:27 to hear some feedback. And that's basically how we do it. And last thing we can share is mode. Maybe

46:33 do vector. So threshold, it's not available when you're using vector. Nice search, for example, for

46:43 using Vector, and I search for example for

46:56 game for children, you basically get data based not on the keywords but on the

47:09 meaning of description and title in that case. So that's how easy you can create you know your own vector search database that works on the client so on the browser at the edge etc and i truly love this product and i

47:17 believe this is going to be interesting for many people this is super powerful wow thank you so

47:21 much for watching us do this um are you Are there templates available or planning to make any templates available

47:28 so that people could do this, for example, with a ready-made template?

47:33 Yeah, I really wish we would get this.

47:36 Now, you have one template specifically, which is on Vercel Commerce.

47:43 So if you go on Vercel Commerce, for example,

47:47 you have Orama with the demo.

47:50 So this is a demo using Orama Cloud, for example,

47:53 or searching through the Vercel Commerce,

47:55 let's say, cap, for example. So you can probably use the fork

47:57 as a template for your e-commerce website

48:01 if you're using Vercel Commerce, for example,