WEBVTT 00:00.320 --> 00:05.690 And in this video, you're going to learn how you can automatically write blog posts using AI and ChatGPT. 00:05.930 --> 00:10.490 We're going to use Lang Chain, and we're going to also break this down into several steps to improve 00:10.490 --> 00:11.570 our outputs. 00:11.690 --> 00:16.460 Firstly, you're going to need to install a couple of different packages, mainly lang chain lang and 00:16.460 --> 00:18.740 OpenAI lang chain community. 00:18.740 --> 00:21.980 And then we're going to import a bunch of packages here. 00:21.980 --> 00:24.560 You can see that we're also using Pydantic. 00:24.560 --> 00:28.460 And if you're using the Lang chain version of that you can also use V1. 00:28.460 --> 00:32.900 Or you could add that to your imports at the top here with pedantic as well. 00:32.900 --> 00:36.830 Pedantic is now supported on version two for Lang Chain, which is great. 00:36.830 --> 00:41.150 And we're also going to now set up a chat model using GPT four. 00:41.270 --> 00:46.250 So notice how we've set up a pedantic model here that we're looking to extract. 00:46.250 --> 00:50.150 We have an outline sections key which is a list of strings. 00:50.150 --> 00:53.690 And in the description of that field we have the sections of the blog post outline. 00:53.690 --> 00:57.670 And if the point is a nested point, then add a number to the start of it. 00:57.820 --> 01:04.570 We then set up a system mock system message prompt template and tell the alarm that it's a helpful system 01:04.570 --> 01:07.030 that's going to be writing blog post outlines. 01:07.060 --> 01:10.990 The outline must be incredibly long, extensive, and detailed, and you are writing an outline on the 01:10.990 --> 01:13.210 topic of and put a topic there. 01:13.540 --> 01:18.370 Also, notice that we make a chat prompt template from the system message, and also we're using the 01:18.370 --> 01:22.780 model with structured output and passing in our Pydantic model. 01:22.780 --> 01:24.940 And this is getting piped the prompt. 01:24.940 --> 01:26.830 So the prompt gets piped into the model. 01:26.830 --> 01:31.450 And then the model is then using this with structured output to know that it needs to produce these 01:31.450 --> 01:33.700 sections, which we defined earlier. 01:34.150 --> 01:37.750 This is our blog post outline for our blog post generation system prompt. 01:37.750 --> 01:42.070 We're going to say that because we're not going to generate the entire blog post all in one go, we're 01:42.070 --> 01:44.530 going to say you are a helpful assistant to write blog posts. 01:44.530 --> 01:47.170 The blog post must be incredibly long and extensive and detailed. 01:47.200 --> 01:48.790 Here is the article topic. 01:48.820 --> 01:52.510 You can say, here are the last three sections of the article that have been generated. 01:52.510 --> 01:55.900 Here are the next three sections of the article that will be generated. 01:55.900 --> 02:02.590 You must render the article in a structured markdown format and also you must produce the content. 02:02.620 --> 02:05.320 Never include the section headings as these are added later. 02:05.320 --> 02:08.020 And then we just tell it to produce the current section content. 02:08.020 --> 02:10.030 And then we have the same kind of thing here. 02:10.030 --> 02:16.210 But notice that we're instead of using a Pydantic model, we're putting the prompt into the model and 02:16.210 --> 02:22.600 then into a structured output parser, which basically means that will be getting back a string. 02:22.600 --> 02:24.520 And you'll see why that's important in the moment. 02:24.520 --> 02:31.540 But to start with, we just do our blog post outline, uh, runnable, which will produce, as you can 02:31.540 --> 02:33.400 see here, this sections class. 02:33.400 --> 02:36.840 And then we can go into the dot outline sections. 02:36.840 --> 02:40.020 And you can see we've got all the various sections here which are a list of strings. 02:40.020 --> 02:45.570 And the ones that are in the numbers here are the the sort of nested ones that you can see. 02:45.570 --> 02:47.370 Now that's great. 02:47.370 --> 02:48.420 And we know that that works. 02:48.420 --> 02:52.440 But one thing we want to store is maybe like the last three sections of content. 02:52.440 --> 02:58.290 And this is important because we want to make sure that the LM isn't repeating itself from between sections, 02:58.290 --> 02:59.610 as we're doing a for loop here. 02:59.610 --> 03:05.280 So we're going to store the history and we're going to enumerate over the outline sections. 03:05.280 --> 03:07.170 And you can see that becomes the index. 03:07.170 --> 03:09.030 And we have the current section there. 03:09.030 --> 03:13.650 We're also going to go get the previous sections and the previous content for those sections. 03:13.650 --> 03:15.660 And then we're also going to get the next sections. 03:15.660 --> 03:21.270 Just so the LM is aware about what sections are going to be coming up next in the blog post, we're 03:21.270 --> 03:28.270 then going to invoke our blog post Generation Runnable Chain, and we're going to provide in here the 03:28.270 --> 03:34.150 topic and also the previous article sections and the next three article sections. 03:34.150 --> 03:37.750 We're then going to add that to the history with the current section and the section content. 03:37.750 --> 03:40.600 And then we're going to generate a message called generated section. 03:40.600 --> 03:44.470 And then at the end of that we're going to take everything that's in the history and just join it together 03:44.470 --> 03:45.340 with a new line. 03:45.340 --> 03:47.710 And then you'll start to see this is kind of how it would work. 03:47.710 --> 03:52.590 So if we're only generating three sections, then you know, this, this is how it would, would work. 03:52.590 --> 03:55.770 And then you can see this is the full blog post that it decided to generate. 03:55.770 --> 04:03.510 So the key points here are at the moment oh doesn't have for all users, doesn't provide the best generation 04:03.510 --> 04:05.640 in terms of an entire blog post in one go. 04:05.640 --> 04:10.830 So we've broken the problem down and decomposed it into two steps, the first step being a blog post 04:10.830 --> 04:16.640 outline, and the second step being that blog post generation on a section by section basis. 04:16.670 --> 04:22.310 The important thing that we've also done is we've included a couple of sections to help this chain have 04:22.310 --> 04:23.540 a bit more continuity. 04:23.540 --> 04:29.150 So this bit here, um, the previous article sections, hopefully that's going to really showcase the 04:29.150 --> 04:34.760 lamb, uh, what has happened in the last three sections so that the lamb has, you know, provides 04:34.760 --> 04:38.000 a better blog post and it's not going to be repeating itself. 04:38.000 --> 04:38.540 Cool. 04:38.540 --> 04:38.960 All right. 04:38.960 --> 04:39.770 I'll see you in the next one.