GitLab #EveryoneCanContribute cafe, 14 Oct 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: 4. #everyonecancontribute cafe: Jina.ai

Description

Blog: https://everyonecancontribute.com/post/2020-10-14-cafe-4-jina-ai/

- Repository: https://github.com/jina-ai/jina
- Examples: https://github.com/jina-ai/examples
- Documentation: https://docs.jina.ai/index.html
- Contribute to Jina.ai: https://github.com/jina-ai/jina/blob/master/CONTRIBUTING.md

A

So now we should be live and good evening good afternoon good morning, everyone um to our everyone can contribute cafe our fourth iteration and today I'm super excited um to welcome um developers founders, everyone from from. Is it china, ai, china, china? I don't know.

B

A

Pronounce china, um so I would say it's best uh alex if you, if you want to start the introduction round and then hand it over just to share a little bit about yourself.

B

Okay, more than happy to I'm alex cg, I'm the open source evangelist here at gina, ai, we're a neural search company based out of beijing and berlin, though we have people all over the world and we're always looking for more uh yeah, so we'll be talking a lot more about gina moving forward. So I won't hog my time here and I'll. We can go on to my colleague rituja, who will also be speaking.

C

C

This is rutuja.

B

Do you want to give a quick self intro.

C

Yeah, so I work with gina ai as an ai engineer and I'll, be uh giving a brief introduction to the technical aspects of gina and we'll be doing a deep dive into the internals and the tech stack.

B

Okay: next up, uh we've got patient.

D

Hello, everyone, my name is paishan, but you can also call me pay, I'm the developer relations specialist at gina, working together with alex and I'm base. We both are basing billing and uh yeah nice to meet you and I'm very looking forward to the talk later.

B

Yep lovely sunny horrible rainy bowling.

E

E

And we are trying very hard to make neural search easy for you, so do try out gina. uh If, if you find it of use to you, I think you will love it.

B

Nicely said anyone else.

A

Yeah, I would hand.

B

A

F

Okay, so I'm nicholas I'm currently working as senior devops engineer at sverity.

F

I'm today light jon snow. So I don't nothing and I'm happy to learn new things also about ai yeah. So I think the next one will be michael, the other michael from austria.

G

Hi, my name is mike: I'm from austria. I work at a company called zkw.

G

We build auto headlamps, for example, for the automotive industry, and I personally be a product manager and software engineer for high performance computing software like uh raytracer engines and 3d cd software, and I'm also really interested in this ai based search topic, because I can need such thing.

A

So and last but not least, I'm the the other other michael, I'm also originally from austria, but I moved to nuremberg, germany, eight years ago I've been with git lab uh since month now and I'm the kind of crazy guy or hosting all all the the coffee chats we have together yeah and I would say, um I think, alex you will be kicking us off.

B

Yep I'll start kicking okay, I'll share my screen and we'll get going get the show on the road. Okay, can everyone's whoops see everything.

C

B

Okay, excellent so yeah uh we've given a quick intro to everyone here, here's a chance to look at all our lovely faces once again, so yep gina gina is an open source search framework. We're powered by deep learning, we're cloud native and we've been going since early this year, we're pretty new startup, but we're back by a full-time venture back team. So I think they're about 20, maybe 30 of us now uh yeah. So, let's dive into these things a little bit more.

B

So by search framework, I don't just mean a search engine that you install and just plug and play, though we are pretty simple to use we're a framework where you can build your. You can use to build your own search engine to search any of your data that you want cloud native, meaning we support all the lovely buzzwords like containerization microservices, all of those and deep learning.

B

Well, we use all of the top deep learning frameworks, things like tensorflow, pytorch, uh mindsport, all the ones you know and love and have maybe played with. Maybe not um the main part of gina is gina core. So it's a state-of-the-art design pattern for such workflows use the latest ai yada yada works out of the box. The key points are we're universal, so we run on a lot of platforms, uh mac and linux right now.

B

Windows is in the works uh not running on microwaves yet, but you can run us on a raspberry pi, we're really easy to use. um I mean I knew barely anything about machine learning before I joined and I've been building my own examples with gina to search like all the lines from star trek. The next generation, because that's who I am cloud native, as I said, and with that cloud nativity comes scaling, we're extensible. We have a hub where you can a bit like transformers hub.

B

So with just a few lines of code, you can suck in models from somewhere else and use those and we're state of the art like using tensorflow pi torch uh onyx fast annoy all of the common machine learning libraries.

B

The other thing to bear in mind is that we are still very new. We haven't had version 1.0 yet so, if you do use us be aware that we're dynamic we're building fast things, may change from release to release and hey. If, if it breaks, you get to keep both pieces, so what's not to like.

B

So when I talk about a search framework, the common question would be oh well. What does it search right? Well, hey? What doesn't it search? We do text, we do images, we do audio, we do video and that's not all. We do cross modal and multimodal all for the low low price of well, nothing at all, because we're open source so cross modal and multimodal are words that not many people really know. I didn't know before I came on board.

B

So when I'm talking about cross modal, I mean something like google images you plug in text and images come out and we've got an example of this I'll show you now.

B

um This is pretty new but yeah, as you can see here, typing in someone wearing a hat. Damn someone wearing a hat comes up.

B

Or we've got a kid playing and we've got kids playing, so you get the.

B

Picture on top of that, we offer multimodal, which I think is even more exciting, so cross modal. You put in one query in one form of media like text you get out another in another form of media like graphics, but with multimodal.

B

You can have image to text and text to text. For example, if you're searching red nike sneakers imported gina can look at all your pictures in your product, catalog of sneakers and with its deep learning models, it can find the ones that are red and that are nike.

B

But just by looking at a picture, you can't tell whether these are imported or not so it'll, also search through the text attached to those images. So if you want to build the next amazon or something, this would be a good use case.

B

A few cool features: uh it's really quick to get going. We we use cookie cutter, so that'll help you create all the boilerplate code. You need just by running a command from your terminal. It'll, give you interactive choices. You just hit one or two or three or four to choose text or graphics or whatever. You want to search it's cloud native, as I said, universal, all kinds of data, first class support for ai models, uh fast annoy, onyx, all the lovely buzzword ones and plug and play so with gina hub.

B

You can extend gina with python scripts or docker images especially suited for your search domain.

B

Now I've done enough talking, you want proof, you've seen one demo, we got more, so one of the ones I like is lyric search. So this lets you search for lyrics in real time it has a user friendly, friendly front and then best of all it has beyonce.

B

So I've got another video demo to show.

B

B

So all of these demos are up on our repository. Most of them will run in docker and if you just want to run them in a virtual environment or even on bare metal, they're, pretty easy to get going with as well.

B

So another example uh because who doesn't love pokemon right, so this uses a public data set of all the pokemon sprites from um the 16-bit pokemon games, anything from game boy up in color, but not 3d, fancy stuff from any next-gen consoles, so learn decks of pokemon, sprites and it'll do a similarity search based on a user query, and with this we use our ginabox frontend, so I'll switch over to that now.

B

So here's our front end so with ginabox. You can search images, video, audio or text uh this time around. I'm running a docker image on my local machine, which is yeah our pokemon search example docker image.

B

So with this you can drag an image of a pokemon like this sort of jellyfish manta ray fella here, and we can search and oh similar things come up.

B

But what, if you want to search something, not a pokemon and find out what the closest pokemon is like this cute little axolotl? So you can try this cute pink animal goes in different cute, pink animal comes out, so it's doing something all right. uh We can even do it with photos.

B

So your favorite star and mine beyonce. What does she look like wow? I mean she's fierce as a singer and fierce as a pokemon too.

B

So yeah, that's that example. Out of the way, uh we've got a whole bunch of other examples on the website. On our examples, repo um patient will be posting links to that in the youtube comments.

B

um Our most basic example, I would say, is hello world. So this uses the really common fashion, mnist data set, which is a huge data set of, I think, 24 by 24 monochrome images of shirts tops bags socks. Things like that, so it'll find a random it'll, dig out a random subset and then find closest things in the whole data set that match. So there's no real user query here, it's just randomly generated and to save time, uh where are we here's one I made earlier so yeah on the left?

B

We've got the randomly selected images from the data set and on the right. We've got a whole bunch of images that gina thinks is similar.

B

You could use this on the original mnist dataset as well, which is handwri, which is samples of writing numbers for handwriting, well, not recognition, but similarity matching.

B

So we've shown you what gina can do. We've shown you some examples, and now how do you actually do it? So gina has a whole family of different components that all interact together. So you can get a high level view a mid-level or a low level view where low level you'd be dealing directly with the models but on the highest level uh you're dealing with the flow, so the flow deals with a specific task you want to achieve and in search, that's generally indexing or querying, and each flow consists of different pods.

B

um These are the things that actually the tasks that build up the flow and we'll go into those in a bit, so it acts as a pipeline telling each part what to do and what the next part in the chain is and for building your own flow. We have three ways to do it. um I'm a fan of the yaml root myself because I live in vim, but you can also do it directly in python or if you're, a more graphical person, we've got gina dashboard, so you can do it.

B

You can design it all via drag and drop. Don't have to touch an editor at all if you don't want to.

B

As I said, each flow is made up of pods, so here we've got gateway, encoder and indexer.

B

So if you're indexing, these are the steps you would use to achieve your specific tasks like segmenting. The input document encoding it into vectors uh in querying you'd, be ranking the top k so yeah. If you're searching for red nike sneakers, what is the top match?

B

What is the second top match and putting them in order instead of just going blah here? Are all the red, nike sneakers again, it's all defined in yaml and I like to think of it as the brains of gina, because this is the bit that uses the uh ai models. um So, as you can see down here, we've got transformer torch. Encoder uh we've got distilbert, which is a really common language model uh based on bert that was created by google and yeah.

B

Beyond the flow and the pods we've got, peas, we've got executors, we've got drivers and retuja will go into some of those in a bit and then to run it all it's as simple as python app.py index, or you can dockerize it, and it's really simple to run it. That way too, once you're up and running, you can monitor it with the dashboard and also in the dashboard.

B

You can create your own flows, it's a low code, monitoring and management environment, and then you can use the ginabox front-end to search it or you can build your own front-end, because we offer a full restful api and, after that, if you've built something amazing, you can upload it to a hub and share it with the world so time for me to wrap it up.

B

So to summarize, we do multimodal cross multi mode with cross modal text images, audio video scalable cloud native. All of that, if you want to learn more, some good places to start are gina where you can meet the different components that I talked about.

B

My first gina app, which is a really simple tutorial to just get you up and running and building something fast.

B

I talked about cookie cutter before it uses that so it's there's almost no hands-on coding required with it dead. Simple we've got our examples, repo with a whole bunch of stuff up there, and we've got uh gina doc's site docs.gina.ai. So yep, that's all from me, and I am now going to turn it over to retuja thanks. Everyone.

D

C

I hope my screen is visible.

F

C

Yeah so I'll be now diving deep into the core internals of gina, starting with executors, so in gina at the moment, we have these four important executions.

B

Do you want to hit the present button to make it full screen.

C

C

Yeah, so any document is initially crafted in the sense it's subjected to different sorts of pre-processing and transformations into uh different chunks, at which the document is broken down before it enters the encoding phase. So the encoder is when we are actually making a vector representation of each chunk and post this. It gets to the indexer phase, where we actually uh try to uh store this indexed document and then later on during query time.

C

This is used for retrieval and that's where we sort the results and like alex mentioned about the flow api, the entire context, management of all these pods and executors is handled by the flow api. So basically, we could imagine all these spots to be different micro services that are very well uh executed using this flow api. So as a user, I don't need to bother about where my pods are running and I only can orchestrate it very easily using simple yaml files.

C

So basically, each flow object would correspond to an actual task and the entire flow right from the pre-processing encoding. Storing, retrieval a query results all is handled by the flow.

C

So it's essential for orchestration of ports and we can define it very easily in python code and yaml file and it's backed by multi-threaded processes and also extend like also extensible to kubernetes and docker swarm.

C

So a typical index flow looks like, as alex mentioned in the previous slides. We have different stages, including chunk segmentation uh document indexing, encode and so on. So uh typically, any um flow in gina has an entry point through the gateway, and then we have different parts that are either sequential or parallel, and uh all this is possible to uh it's possible to implement this flow very easily, using gina dashboards or through the cml file. So it makes context management. Super simple in general, using flow as we mentioned, gina supports different modalities of search.

C

It could be either text image, audio or video in this case. uh What's common across all these modalities? Is the input function for indexing?

C

So we make sure that, irrespective of the mode, the input is always in the form of a byte buffer, and we allow different functionalities like index, search and train for index for feeding in the index data. So we have different crafters for different types of modes that are supported by gina and we allow uh all indexing and searching at different levels like if there's a text document that has several lines, we can call index lines. If there are different files, we can call index files directly and supply the input function.

C

So next I would move on to the concept of documents in gina and how exactly is any document represented in gina? So there are two important concepts here. One is that of the chunk and the other is that of the matches.

C

So any document, uh if we were to look at this red block here, this is the document that we want to say a query in this case. So what we do is we are breaking it down at a granularity level, one where we are breaking it down into two chunks.

C

So basically, that's like dividing the document into different segments. So at each chunk level we are looking at the matching uh chunks that are available in the indexed documents. So whenever we are trying to see anything that is matching a chunk.

C

That's when we are looking at the adjacent chunk adjacency in the sense similarity, so uh we have these concepts of granularity and adjacency where in granularity, we are going down breaking the document further breaking the chunks even further and at the adjacency level we are basically looking at uh how similar is this document with the other uh chunks.

C

So to imagine this as a tree structure, this is how a a sample tree traversal looks like in gina. So this is like, given granularity, 0 and adjacency 0, we are moving down at the chunk level. Then we move down at match level and so on. So the trick here is we move on from 0 0 to 1, 0 and then 1 1. We cannot move directly from 0 0 to 1 1, so that's an important concept here, while traversing so uh to illustrate this in a better manner.

C

uh This is an example where we have a car, so at the chunk level, what we'll be doing is we'll be uh looking at having a zoomed in look inside the car uh looking at smaller parts of this image. So this is where we are going down at the chunk level. So this is at level one, and this is further down at level two. Where we only see the headlight, then what we are doing is at this depth of level.

C

uh At this depth of chunk level, we are going to find the matching images so uh for this zoom in car. We get another match here, which is quite similar. Similarly, if we were to go down to the headlight level, then we find a similar headlight of the silver car corresponding to this green car, and uh this is like different levels at which we are trying to match.

C

So uh in this case, we need the parent id for traversing all the way back to the main parent document to which the smaller chunks belong to, and we have a vector indexer that has the id information in this case.

C

So next I'll be talking about few design patterns that are commonly followed in gina, starting with the concept of compound indexer. So we have two types of indexes here: one is the vector indexer, so the rationale behind having vector indexer is the concept of similarity.

C

We want to embed documents in a similar space and compute the similarity and make sure that, um while retrieving or while querying and ranking the documents, similar documents were indexed closer to each other in the same embedding space so uh further down, it's easy to retrieve the corresponding text for the document. If we, if we were to have a key value index one, so these two indexes vector and key value indexes, are often used, sequentially and parallely for different purposes.

C

So this is a common design pattern in g now. uh Next, I would talk about the text document segmentation concept. So in g now we introduce the concept of chunks or a recursive document structure where we say that one document is in turn composed of multiple other documents at different levels of depth, making it a recursive document structure, and each of these compositions is a chunk. Basically.

C

So we have a crafter that segments documents into smaller parts.

C

Also, we have indexers available at different depth levels. So uh it's important to note that when we are searching that search is actually performed at the chunk level, where we follow the compound index or pattern, and if we are to find out the actual document that corresponds to the particular chunk. That's where the document indexer comes into picture as a final step.

C

So uh another important concept is what happens in index time and what happens at query time so at index time, the two indexers, which is the document indexer and the chunk indexer they work in parallel. So what happens is chunk? Indexer is responsible for receiving messages from the encoder and doc. Indexer is the indexer that actually fetches the documents from the gateway, so gateway is like the entry point for the flow.

C

So in at query time document and chunk indexes work sequentially in the sense that a document gets messages from chunky indexer with a ranker that we have implemented called the chunk to dock ranker, and then the ranker ranks these chunks by relevance, and then it gets the parent ids corresponding to these chunks and then to track which document this id belongs to. We use the doc indexer.

C

So uh at query time we have sequential indexers working.

C

So to summarize, uh gina supports a lot of different types of modes of search, and it makes sure that um different piece together are embedded in a pod, and this pod acts like a micro service. It could be for different types of executors, like a ranker, a crafter, an indexer, an encoder and so on, and the flow api gives a very seamless experience in the context management for all these pods and uh for refining our results to make our search even more uh better.

C

We have recently introduced the concept of recursive document structure, a tree traversal, while indexing and in future we also have plans of supporting kubernetes and managing this entire thing. That is now available as a docker image hosted on docker hub also will be ported to kubernetes so yeah. That's it from my side.

B

um I've just got one more thing to share um a final slide with some links. If anyone wants more information, so hang on.

B

So yeah, that's all about contact info. If anyone wants to get in touch or you can email me directly at alex.cg at.

F

F

Yeah nice presentation.

A

C

A

Mind blown at the moment. I need to unpack so many things in my head.

B

That's why we've got the q a okay.

G

One question um is it: you have images, you have text and you store your or your structure, your data in in documents, so you extract data out of the stuff, so you can search it faster or you can search it in the amount of time.

G

Can I give the system any kind? Can I write this document myself so that I can, for example, use 3d data to put it in to search it.

B

That's a good question. I think we'd need to use libraries specific that can encode 3d data. I mean, let's put it this way, if there's a library that supports that it can be integrated very easily into gina.

G

So I understand the question, but I mean: can I put when I have the extractor when I can read 3d data and extract some information on the? I have a bounding box, for example, and I use the length of the box or something like that can is. There is the api able that I can put the stuff into it. So then I can search it.

F

E

Alex can I take a hit at this one yeah, of course,.

B

E

uh So currently uh talking in terms.

E

Which sort of detects certain things inside these images like option protection or segment? So these are two different.

C

Pieces one is like the return.

E

Of the image itself.

E

So we currently do not have anything that you are asking for this kind of applications, but these are very easy to sort of write as small modules, and this can be added to the gina and the rest of the file automatically. You just need to make sure that the requirement of these classes are satisfied, inputs and outputs.

G

B

B

If anyone wants an example of uh documents and chunks um and feel free to correct me, if I'm wrong, everyone like pratik and ratuji are the experts, but you in terms of recursive document structure, you could think of the document you want to index as let's use a book as an example.

B

So in a book you have chapters, so each of those would be a sub-document and in a chapter you have paragraphs again, sub-documents of the chapter so sub-sub-documents and in a paragraph you have sentences and sentence. You have words, so that would so you could search from the word level and that would bubble up or you could search for a paragraph level or, however, you want to do it. um Is that about right, return, practice.

E

Yep, it's right.

E

It's also not necessary to bubble up. If you want information at the junk level, also, like you can match documents to document itself, you can match chunks to chance itself. You can do any of this combination. uh We just need to define that recursivity in our flows.

B

Yeah, you can choose the level of document you put in. You can choose the level of document you get out. Yep.

B

Do we have any q a's coming up from.

D

Youtube not that I'm aware of.

A

Maybe we can try something live. um I was wondering about the pokemon demon this. I cannot.

A

Maybe maybe do something with the the gifts uh or do you have any other demons around which are catchy and uh well. We could just try it.

A

Out, I think you're, muted.

F

B

Do you have any demos uh you can show us.

E

uh Right now you know it will be difficult for me to start something uh but yeah we do have other demos in the pipeline.

E

uh But cannot show it live right now.

B

Yeah, the problem with showing search demos is that ninety percent of the time- yes, you spend you doing the indexing.

A

That's especially the case when.

B

It comes to machine learning because so.

G

Machine learning is.

B

Powerful but it's not the fastest thing in the world.

G

Looking at the progress bar.

B

In fact, I will say it's faster in terms of the time humans have to put into it, so.

E

If you're writing a traditional.

B

Search engine you have to write all sorts of rules and processing pipeline to do stemming lemmetization, all of that jazz with gina we're using pre-trained models that were like created by google. So they already know what they're doing so. You just throw the data at it and it returns relevant data, but it needs a lot of time to index that data. So it's the computer that takes the time not the developer.

A

Okay, then, maybe a different question. um You mentioned the github repo and all the other urls before. If I want to start contributing to your project um which steps are needed, which, how would I start contributing to your project?

A

Maybe maybe you can walk us a little through, maybe like sharing a screen opening up whatever document you might have and talk a little bit about this or if you need resources from someone else anything you can think of.

B

Sure thing, okay, so I can share my screen.

B

Now so a good first thing is to always join uh gina. Ai community slack um is adding invite links to the zoom and the youtube channel. The youtube video, I should say uh so. You've got a whole bunch of channels here, but the most important one is general.

B

So this is where you can find us where you can talk to us, ask questions and yeah: that's the best way to get started in talking to us, and we really like people to use the general channel and not private channels. For this, because you know sharing is caring. Let's share that information out there and let the world see it, so everyone can benefit.

D

Additionally, additionally, there's also a support channel for specific, like queries like issues as well.

B

Support channel just here, we've also got uh the repo always has a bunch of issues. Gina ai, slash, gina,.

B

So yeah I'll bump up my font size, so everyone can see so some of these we've got some good first. Issues like anything with hacktoberfest is a good first issue, so lots of things for you to check out for you to help with we're, also always on the lookout for people who want to contribute examples.

B

So, as I said, we've got a whole bunch of things on our examples: wrapper gina 101, the first thing you need to learn and all of these I would suggest people start with my first gina app, which was written by yours truly.

B

So this lets you build your own search, app that searches all the lines from star trek. The next generation we've also got one that searches south park. If that's more, your cup of tea.

F

B

F

Was would be also.

G

B

So yeah you can run it from docker. You can query it straight in your web browser earlier. I showed you ginabox, which is just here, and this is all running on genus server. You just type in the end point of your running docker instance and it'll connect flawless well, seamlessly, I would say, and it just works out of the box, so you don't need to install any front end. You don't need to get npm going. You don't need to run a web server locally.

B

Don't need to run the front end in docker, no just go to gina dot, ai, slash, ginabox.js and boom. It's there ready. You can also install it on your own device. If you want it's all open source tear the code. Apart do what you want with it.

B

uh With querying you don't have to use ginabox, you can use anything that can connect via rest, so you could use curl or you could write your own thing. I'd put some json because, of course it does in terms of installation.

B

It's pretty simple. You just use cookie cutter.

B

um In this tutorial it is comprehensive. It goes through every little thing, so it is quite long, but that's because it's holding your hand, because this is the first project for a lot of people once you've done this project. You've already picked up a lot of info, so you can get up and running for the next ones.

B

So yeah and then, when you want to run it it's as simple as app.py index, which takes a while, because machine learning always does with text it's not so bad. It might take well depending if you use a lower max stocks. Value you'll get worse results because you'll index a lot less, but you can also get up and running a lot more, a lot quicker because it doesn't need time to index so much so it's a speed, quality trade off there and then to run it.

B

You just app.py search and then you can open your web browser to ginabox and boom, make it so and it'll come back with output and we go into a little bit here about flows and all of that stuff too.

B

uh Beyond that we've got uh gina 101.

B

Which introduces you to the gina family? I return, and I both spoke a little about this, so we've got documents and trunks, so chunks are just essentially sub documents, everything's defined in yaml, so dead, simple to use no actual code required, uh the executors do most of the work encoding, crafting, ranking and so on.

B

uh We've got a few other things like drivers, uh wrappers for the executors, the pods that help you get a lot of executed, running np's running in parallel to speed things up so replicas and shouting and all of that and then the top you've got the flow which is all about context, management, orchestration and all of that.

B

We've got the doc site at gina.ai with our beautiful animated gif, so you can install by a pip. You can run it by a docker. You can run it on a raspberry pi.

B

B

Any questions so far.

A

G

A

I think niklas will try it out right now.

F

Yeah, probably I tried to start it.

B

Another thing we do is: every month we have our engineering all hands. I think it's, the second tuesday of the month. uh Patience can.

D

B

Yeah, yes, so yeah! So that's when our engineers all get together, uh overzoom, it's a public event. Public webinar! Anyone can join in. Ask questions, learn more about gina there and you know find out what the engineers are working on right now and that's where you can hear from our ceo, hans xiao uh he's the brains behind gina.

B

Well, he's the guy who started gina. There are many brains behind genome there and yeah he's also the man behind the fashion mnist data set that I demoed earlier. He uh runs. He started the birth as service uh github. Repo he's done a lot in machine learning, uh we're always looking for new team members. So if you've got a background in uh yeah machine learning, artificial intelligence, uh evangelism and developer relations, product management reach out to us, because we're growing fast we're looking for new talent and yeah.

B

We've got teams in beijing and berlin, but we've also got a bunch of people in india in the us in spain, so yeah, and at times like this, who cares where you work anyway? Nobody can go to an office right.

G

Different times.

B

So yeah, oh um another thing we recently had a hackathon not long ago, a whole bunch of cool projects came out of that, including some real world solutions for using gina to detect fake news or using gina to how do I put it, create inspiration for ad creators to create new social ads on facebook or twitter and other ways to annoy people, um and we had several covid related projects that came out of that.

B

We are looking at running another one, probably sometime in the next three months, so yeah, that's another great way to get your feet wet.

B

But yeah the number one way would be join us on our slack. The link should be in the chat which is there or somewhere. I don't know where you people put your chat windows but it'll, be either on youtube or on our zoom chat. So you can find us there and that's the best place to reach out we're very friendly, we're eager to support you. We don't bite much so.

A

Maybe you bite me now um if I'm asking the question of the, what is the most anticipated feature in the next year's roadmap or.

C

A

What is coming and what? What are you personally, the most excited about.

B

Wow, where do I even begin? It's been such a rush, just focusing on uh getting the word out about what we already have. I haven't had time to catch my breath and see what's in the pipeline.

B

But yeah I'm really excited personally about seeing gina used for real world applications. We had another one built uh by arte tanona. He built uh a search system for european uh judicial rulings uh using gina that wasn't even at a hackathon. That was just something he did and I'm also looking forward to integrations with other frameworks. Other things at the hackathon someone built a hybrid chatbot, neural search system using gina and rasa, which is a machine learning powered uh chatbot platform, also based here in berlin.

B

Pratik any thoughts.

E

uh Regarding the upcoming plan.

B

That we can talk about, remember.

B

E

B

E

Careful we can give them a hint.

E

Everyone is excited about. This feature is that a lot of people come and ask us that in the current we can use the models that are available, but people want to train their own model and then use them. So this is something that we look forward to and it's a very complicated piece, uh because we have to do it in a distributed way in a cloud and also support lot of different models. So this is.

C

Something that we are.

E

In next year, definitely.

B

Yeah- and I should reiterate that right now- we're using pre-trained models which are pretty damn good anyway, but you know, there's always room for improvement for specific use cases. So yeah being able to train is a good thing.

C

Yeah, even I agree with what uh pratik said.

F

Cool cool, so I got also a last question or probably a question. So when is the juvenile release coming? Probably so, when we.

A

F

Use it on trivia meters.

B

Do we have a fixed date, for that? Can we talk about.

C

B

State for that.

C

Yeah yeah we'll be soon also using kubernetes. So currently we are planning to uh host it on aws infra and we are on the path of uh using terraform for uh infrastructure deployment post. This we'll be also uh connecting this and uploading it to kubernetes.

C

So all the pods that we spoke about will be available as kubernetes pods for micro services, and uh definitely it will help us scale out very well. So that's definitely our uh plan for the next quarter.

F

B

Yeah one of the cool things about the flow and the pods is that using a flow you can use pods on your local machine. You can use pods running in a container like in docker. You can use pods running remotely all in the same flow, no problem at all. It's the code.

F

Is dead, simple.

F

D

Just like alice keeps saying is, then damn simple, so I I hope.

B

You guys, like me, potty mouth.

D

D

Giving you like the wrong wrong expectation, yeah yeah,.

B

D

B

Like I said before, joining gina, I hadn't played around with machine learning much and after joining gina just playing around with this. It's I mean it. It's ridiculously simple: it's one of the easiest machine learning things I've used.

E

B

One of the next talks.

E

Maybe in one of the next talks you can bring a child a kid five-year-old and then maybe he can do the things and then prove it simple. Is it for certain use cases? It's really good like for some well-known cases like semantic search, uh it really runs runs out of the box. Otherwise you have to do tons of engineering like before.

E

Before working at gina. I used to make these things customly, and it would definitely take one or two months to get something running and to make something as robust. As you know, it would take months like it's not even possible to do it so yeah I mean certain use. Cases is dead.

B

All right we're coming up to the top of the hour.

E

A

Oh true, I was reading the source code while you've finished playing sometimes.

G

F

He has currently also set up a new model to see how many cheat sheets we can detect.

A

Well, I was actually wondering about the road member little and did it dive into the projects on github, and then I found out about something about cicd and then I was reading the source code, um but we can also do that asynchronously. So um I would say we we can wrap it up here, um thanks for the for the great introduction for everything, you've you've shown us- and I I have learned a lot today.

A

I just need to find the time to write a blog post, probably about it I'll just dive a little deeper into it. um We will share everything. We've talked about today, all the urls, all the the things um later on, in the youtube description and the blog post.

A

um So yeah thanks for attending, and with that I would just stop the live stream on youtube.