WEBVTT 00:00.020 --> 00:02.450 You have a look at what a chat prompt template is. 00:02.480 --> 00:08.000 It is a template that allows you to create a list of messages that will form your chat history. 00:08.000 --> 00:10.790 And you can have these curly braces in here. 00:10.790 --> 00:14.900 So you can see we've got curly braces name or curly braces user input. 00:14.930 --> 00:21.260 When you invoke a template, essentially what's going to happen is for each individual curly braces 00:21.260 --> 00:27.470 with the name of that, you're going to inject a variable which will then output a chat prompt value. 00:27.500 --> 00:28.640 Look at the messages. 00:28.640 --> 00:29.120 You'll see. 00:29.150 --> 00:35.960 These then have become formatted strings, allowing you to inject variables at runtime into your prompts. 00:35.990 --> 00:39.110 Let's have a look at an implementation of this inside of Lang chain. 00:39.140 --> 00:45.140 Now we're going to import a new import which is Lang chain called prompts dot chat. 00:45.140 --> 00:48.560 And we're going to import the chat prompt template class. 00:48.590 --> 00:52.850 Then what we're going to do is we're just going to start by doing chat prompt template. 00:52.850 --> 00:55.190 And then we're going to make a template here. 00:55.190 --> 00:58.940 And then what we're going to do is we're going to have a chat prompt template. 00:58.940 --> 01:04.490 And we're going to do from messages, which is a function which takes a list of tuples. 01:04.490 --> 01:08.870 And in each tuple on the first index of that tuple you have the message type. 01:08.870 --> 01:13.010 So you can see the first message is a system type message that we learned in the previous video. 01:13.040 --> 01:17.220 We've told it that it's a helpful assistant that's capable of generating company names. 01:17.220 --> 01:21.810 And we also have a human message which talks about a text. 01:21.840 --> 01:23.730 Now, a text is a dynamic variable. 01:23.730 --> 01:29.430 So what we can then do is just like we can invoke in the chat model, we can also invoke a chat prompt 01:29.430 --> 01:29.910 template. 01:29.910 --> 01:33.120 So I'm going to invoke and then we pass a Python dictionary. 01:33.150 --> 01:38.430 You just have to look in each of your chat prompt template variables and see what are the dynamic curly 01:38.430 --> 01:39.930 braces that you're going to need to replace. 01:39.930 --> 01:41.040 So I have text. 01:41.040 --> 01:45.300 So I'm going to put here and I'm going to say what would be a good company name for a company that makes 01:45.300 --> 01:46.140 colorful socks. 01:46.140 --> 01:49.020 And then what you'll see is we actually get a result. 01:49.560 --> 01:51.330 And we're just going to have a look at that. 01:51.330 --> 01:53.040 So I'm going to print the result. 01:53.760 --> 02:01.260 And then we have a result which if we have a look at the type you can see this is a chat prompt value. 02:01.290 --> 02:05.190 We can pass a chat prompt value directly into a chat model. 02:05.190 --> 02:06.750 Let's have a look at how we could do that. 02:06.750 --> 02:08.670 Firstly we're going to set up a model. 02:08.670 --> 02:11.100 So I'm going to do model is equal to chat OpenAI. 02:11.100 --> 02:12.900 And we're using GPT four mini. 02:12.900 --> 02:16.140 And then what we're going to have is an AI LLM result variable. 02:16.140 --> 02:17.880 And we're going to do model dot invoke. 02:17.880 --> 02:24.090 And we're going to take the result, which is the result of formatting our prompt by invoking it. 02:24.090 --> 02:27.090 Then we're going to invoke based on the template. 02:27.110 --> 02:29.690 And we're going to store that in an AI LLM result. 02:29.690 --> 02:35.090 And then just like we did in the previous video, do print AI LLM result dot contents, we're going 02:35.090 --> 02:38.360 to use that dot content key to see what the actual value is. 02:38.360 --> 02:40.940 So again you can see how this is kind of flowing. 02:40.940 --> 02:45.410 We first have now made a prompt template which has a list of messages. 02:45.410 --> 02:49.490 The first message being a system message, the second message being a human message. 02:49.490 --> 02:53.300 You could even add in AI messages as well, like you're seeing in the suggestion. 02:53.300 --> 02:59.420 And as well as that, when we have curly braces for each curly brace, we have to replace the text. 02:59.420 --> 03:04.970 If I, for example, uncomment this code and I try to invoke this, and I try and do this without, 03:05.000 --> 03:11.150 it throws an error and it says that it actually is missing a variable called text expected text received. 03:11.150 --> 03:12.380 But we didn't get any. 03:12.380 --> 03:17.690 So what you need to know is that every time you have curly braces, you will have to add those into 03:17.690 --> 03:20.990 the dictionary, which will then invoke the prompt template. 03:20.990 --> 03:23.630 So again we're then getting the prompt template. 03:23.630 --> 03:28.520 We're formatting it into a list of messages which we're then setting up a chat model. 03:28.520 --> 03:31.460 And then with the chat model we are invoking the result. 03:31.460 --> 03:37.370 And then once we have the result back from the LLM, we're then doing the dot content on that to get 03:37.370 --> 03:38.720 the actual output.