How-to Structure a Q&A ML App

00:00:00.640 | Okay so I had an idea. I thought maybe what we could do is actually put together a few videos

00:00:08.400 | and what we'll do in the next few videos is build out a full Q&A project. And this is something

00:00:15.520 | that I've actually been wanting to do for quite a while now and I think it will actually be really

00:00:21.360 | cool. I was thinking through it this morning and I thought okay today we're going to start it.

00:00:26.320 | So what you can see on the screen right now is the github repo for this project which is

00:00:32.880 | completely empty almost. There's a readme, a gitignore and this one architecture drawing that I

00:00:39.360 | set up literally you can see four minutes ago. So I'm going to show you that and I'm going to

00:00:46.480 | explain what we're actually going to go through and I think we're going to cover a lot of different

00:00:53.040 | things. So I think that is one of the reasons why I think this is such a cool project.

00:00:58.320 | So at the moment this is like the basic architecture of what we would need to build

00:01:06.800 | a Q&A model and the end goal is to have a front end which looks kind of like this.

00:01:20.320 | So we'll have like a search bar here and we'll have some visual up here. It's going to be a

00:01:27.200 | little bit better than the sigma and I'm going to show you what I already have and maybe we

00:01:31.600 | can use that or maybe we'll do something different I don't know. And what we're going to be able to

00:01:35.120 | do is ask a question here and we're going to be able to answer the question based on

00:01:45.760 | stoic philosophy books. So I haven't really read any of these, I've read like little bits

00:01:51.600 | but they're pretty interesting and I think quite unique. So as far as I know there's

00:01:59.040 | definitely not anything like this out there at the moment where you ask a question and you get

00:02:05.280 | the answer back from some ancient stoic philosophy book. And there's only really two books that I've

00:02:12.400 | thought of so far which is Meditations by Marcus Aurelius and Letters from a Stoic by Seneca.

00:02:24.800 | And the good thing with both of these is that we can find both of them online for free so we can

00:02:34.960 | use Python requests to get these. So I'll just kind of put a little list here of what I think

00:02:41.360 | we're going to need. So the first one is actually extracting and downloading this data so we'll be

00:02:49.280 | using requests for that. And then once we have actually got that data we need to pre-process it

00:02:57.840 | and when we're pre-processing it I think that will just be a case of using Regex

00:03:05.760 | more than anything else but I'm not sure yet so let's see. So after we pre-process it and

00:03:13.120 | that's when we get into this stuff over here. So this whole sort of stack that you can see

00:03:19.040 | without the API. So this is a typical, it's called a reader or retriever

00:03:31.600 | reader and what we do is we use this up here, this is our database, it's a document store,

00:03:39.360 | Elasticsearch document store and what we're going to do is feed all of these.

00:03:46.080 | Sorry it's getting a little bit messy with the color so let me change it.

00:03:58.400 | So what we're going to do is take these and we're going to feed them

00:04:01.840 | into our document store and once we have that what we want to do is build this retriever reader

00:04:10.880 | stack and it will allow us to query the retriever down here and what the retriever will do

00:04:21.680 | is send that query to Elasticsearch here which is what you can see happening there.

00:04:28.080 | And returning from that we'll get so many different contexts. So all of the text from

00:04:36.640 | meditations and letters from Stoic, we'll split them up by maybe paragraph and saw them in here

00:04:43.360 | and what these contexts will be are the most relevant paragraphs.

00:04:51.360 | And once we've done that, this retriever will pass on the context to our reader model down here

00:04:58.880 | and what the reader model will do is say it's given a long sentence like this or paragraph,

00:05:07.360 | it will say okay the actual answer that you want is actually only these three words here and it

00:05:13.840 | will return those three words and what we want to do is return those three words in our answer

00:05:20.880 | back to our API. But alongside the answer we're also going to include the full context here

00:05:28.720 | as well. So we get a few things back and I think that that's like going to be the more

00:05:38.240 | machine learning side of it but obviously we need to support all the machine learning side

00:05:42.560 | and I mean the very first part of that that you can obviously see here is the

00:05:49.280 | API. So let me so let me write down so we have the the ML part

00:05:54.080 | and for that we're going to be using something called haystack.

00:05:58.720 | And once we get out of that part we move on,

00:06:07.440 | I'm just going to use a different color here, move on to our API.

00:06:15.280 | The API we'll just we'll use probably fast API to set that up.

00:06:20.000 | Then once we set that up we go on to our front end part

00:06:26.400 | and the front end I don't I'm not a front end developer I just mainly use Python but I do know

00:06:35.600 | Angular a little bit. So what we're going to do is build all of this part using Angular.

00:06:45.200 | So this will be our Angular front end and that's essentially everything we'll be covering but

00:06:58.320 | there's you know there's quite a lot in here. In particular as well what I've missed

00:07:04.240 | is alongside you know we have our reader model down here but what I want to try and do is rather

00:07:12.400 | than just taking the reader model from Hugging Face Transformers as we normally would, I want to

00:07:21.440 | actually try training it and for that we need to use something called MLM which is mass language

00:07:30.800 | modeling. So we would need to train a BERT model using MLM or fine-tune the BERT model I should say

00:07:39.760 | on the data from our books up here and then we'd also want to train it so that it performs Q&A

00:07:49.040 | and for that we need to use the squad data set, probably squad anyway.

00:07:58.000 | So you know there's quite a lot that I think we would have to do to build this

00:08:04.000 | and I think it'd be pretty interesting. So that is what we're going to be covering in

00:08:12.160 | sort of the next few videos and the one other final little thing. Okay so over here we have

00:08:21.840 | the Marcus Aurelius Sit-Man and I thought maybe something like this would be cool. I don't know

00:08:28.080 | this is something I drew ages ago and this is Marcus Aurelius and I think something like this,

00:08:36.480 | maybe this or something like it would be pretty cool to just have in the middle of the web page

00:08:40.640 | and underneath we have a search bar and keep it pretty simple. So I think that's everything really

00:08:48.800 | for the plan and I think the first thing we're going to do in the next video is actually

00:08:55.440 | sell requests and download that data and maybe pre-process it as well or they might be two videos.

00:09:03.040 | So that's everything for this video. I hope that you're as excited about this as I am because I'm

00:09:11.040 | really looking forward to actually building all of this. I think it'll be super cool and

00:09:15.280 | I mean ideally at the end it's one it's going to look cool and two we're going to learn like a huge

00:09:23.520 | amount of stuff if you even put all this together there's so many different things that you need to

00:09:27.440 | know in order to make everything work. So it should be really cool and I'm looking forward

00:09:34.160 | to getting started with it. So I will see you in the next video where we'll actually download the

00:09:41.040 | the data and pre-process it.

00:09:42.900 | [BLANK_AUDIO]

How-to Structure a Q&A ML App

Chapters