1 00:00:03,120 --> 00:00:08,250 So in this basic app we are going to summarize a long article. 2 00:00:08,400 --> 00:00:16,379 As you know, one of the main problems of the foundation LMS is that they are limited by what we call 3 00:00:16,379 --> 00:00:17,820 the context window. 4 00:00:17,820 --> 00:00:27,270 And this means that, for example, the free version of ChatGPT cannot handle a text of more than 4097 5 00:00:27,270 --> 00:00:27,990 tokens. 6 00:00:28,320 --> 00:00:34,230 So what if we want to summarize a document longer than that limit? 7 00:00:35,190 --> 00:00:43,800 We are going to use long chain to help the foundation Elm summarize a document that exceeds the content, 8 00:00:43,800 --> 00:00:45,210 the context window limit. 9 00:00:45,210 --> 00:00:49,080 And we are going to do that in four simple steps. 10 00:00:49,140 --> 00:00:51,450 First we are going to load the document. 11 00:00:51,570 --> 00:00:54,840 Then we are going to check its token count. 12 00:00:54,840 --> 00:00:58,290 Then we are going to split it into smaller parts. 13 00:00:58,290 --> 00:01:04,769 And finally we are going to use a predefined line change chain to send the parts, the chunks of the 14 00:01:04,769 --> 00:01:09,300 article to ChatGPT and get a summary of the whole document. 15 00:01:09,420 --> 00:01:13,050 So here we have the code we are going to use for that. 16 00:01:13,050 --> 00:01:21,060 In the initial part of the Jupyter notebook, you will see what we normally have in every notebook. 17 00:01:21,060 --> 00:01:26,460 This is just an invocation a call of our dot env file. 18 00:01:26,460 --> 00:01:29,730 So we have the dot m file in the same folder. 19 00:01:29,730 --> 00:01:37,890 We have this Jupyter notebook, and we are calling this file in order to load the OpenAI API key. 20 00:01:38,600 --> 00:01:49,250 Once we have that A, we are a loading the OpenAI component in order to create an instance of our realm, 21 00:01:49,250 --> 00:01:52,280 and we are going to load the text file. 22 00:01:52,370 --> 00:01:57,470 So we have a text file in a folder called data. 23 00:01:57,650 --> 00:02:08,960 And we are going to load this file and save it in a variable called article which is a variable of type 24 00:02:08,960 --> 00:02:09,680 string. 25 00:02:10,009 --> 00:02:15,620 If we print the first characters of this article, we will see the title and the first paragraph. 26 00:02:16,220 --> 00:02:20,780 Once we have that ready, we can, uh. 27 00:02:21,930 --> 00:02:26,640 Take the number of tokens that this article has with the function. 28 00:02:26,640 --> 00:02:30,570 This is a default function in launching a. 29 00:02:30,570 --> 00:02:39,840 To check the actual number of tokens in this article, and we see that this is more than 6000 tokens, 30 00:02:39,840 --> 00:02:45,600 which as you know, is over the context window limit we have. 31 00:02:46,020 --> 00:02:48,870 So how can we solve this problem? 32 00:02:48,870 --> 00:02:53,910 We are going to solve this problem using a text splitter and a chain. 33 00:02:54,240 --> 00:03:00,630 The text splitter we are going to use from Lang chain is called recursive character text splitter. 34 00:03:00,630 --> 00:03:05,310 And the first thing we do is we configure this text splitter. 35 00:03:05,310 --> 00:03:11,820 So we are going to tell that we want to separate, you know the text by paragraphs. 36 00:03:11,820 --> 00:03:15,870 The the size of our chunk is going to be over 5000. 37 00:03:15,870 --> 00:03:21,450 And the size of the overlap is going to be over 300. 38 00:03:21,450 --> 00:03:28,080 Once we have this configuration we can create our article chunks variable. 39 00:03:28,080 --> 00:03:35,700 Applying to this text splitter, we have configured the function at the default function create documents. 40 00:03:35,700 --> 00:03:41,850 So this is going to uh divide the article in different chunks. 41 00:03:41,850 --> 00:03:50,370 And as you can see, if we print the length of the article chunks, we see that we have eight chunks 42 00:03:50,370 --> 00:03:51,360 of text. 43 00:03:51,990 --> 00:03:58,860 And uh, right now what we are going to use a chain called load summarize chain. 44 00:03:58,860 --> 00:04:09,300 This is a predefined chain in in long chain that is going to help us to create a chain that includes 45 00:04:09,300 --> 00:04:18,269 our LM and uh, the chain tab, which is going to be the summarize uh, summarizing uh, change type. 46 00:04:18,269 --> 00:04:25,680 So once we have the chain configured we apply that chain, uh, to the article chunks. 47 00:04:25,680 --> 00:04:29,430 We have uh, build before. 48 00:04:29,430 --> 00:04:39,210 And once we have that, if we run this, uh, by ChatGPT, we are going to produce the summary of the 49 00:04:39,210 --> 00:04:40,260 whole article. 50 00:04:40,260 --> 00:04:47,070 So basically what we have done with this chain and the execution of this chain, uh, applying the article 51 00:04:47,070 --> 00:04:55,680 chunks, we have sent the eight chunks of the article to the to ChatGPT, and ChatGPT has been receiving 52 00:04:55,680 --> 00:05:02,460 these eight, uh, chunks and has been creating a summary for all of them. 53 00:05:02,460 --> 00:05:08,130 So what what it does is it does a summary of the first one, a summary of the second one, etc., for 54 00:05:08,130 --> 00:05:08,820 the eight chunks. 55 00:05:08,820 --> 00:05:12,120 And then it does a summary of the eight summaries. 56 00:05:12,570 --> 00:05:21,720 So as you see in a very simple application, we have solved one of the main problems that the foundation 57 00:05:21,720 --> 00:05:23,040 LMS has.