WEBVTT 00:00.020 --> 00:03.800 Hey, in this video you're going to learn how to use Lange chain chat models. 00:03.830 --> 00:09.530 Chat models work by having a history of chat messages where you can talk to them over and over again. 00:09.530 --> 00:13.070 As you'll see, they're the most common type of interface when working with a large language model. 00:13.070 --> 00:16.790 Firstly, we're going to install a couple of packages inside of the notebook. 00:16.790 --> 00:18.920 So I'm going to install Lange chain Lange chain open. 00:18.920 --> 00:22.010 I will also install Lange Chain Community. 00:22.190 --> 00:25.520 And we'll also upgrade these packages and put that on a quiet. 00:25.760 --> 00:28.880 If you've already installed this you can skip this step. 00:28.880 --> 00:33.770 The next thing that you'll need to do is import the OS package, and you'll need to set an API key. 00:33.800 --> 00:39.260 I'm not going to do this step, because I've actually already got an OpenAI key set in my terminal window. 00:39.290 --> 00:45.260 Then what you'll need to do is then we're going to import the chat model directly from Lange chain OpenAI. 00:45.290 --> 00:50.150 And what's interesting about this is you can easily switch out chat models by installing a different 00:50.150 --> 00:52.550 package and then importing a different chat model. 00:52.580 --> 00:57.800 The next thing that you need to do once you have done this is you will then need to create a client. 00:57.800 --> 01:00.620 So we'll call it chat OpenAI. 01:00.650 --> 01:03.590 And then you can specifically pick whatever type of model you want. 01:03.620 --> 01:06.110 I'm going to pick GPT four oh mini. 01:06.140 --> 01:08.900 And you can also have different types of parameters. 01:08.900 --> 01:12.170 So for example you could put the temperature in here etc.. 01:12.200 --> 01:14.870 Let's have a look at how we can then actually utilize that. 01:14.870 --> 01:20.870 So I could do chat dot invoke and invoke is a function inside of lang chain, which allows us to operate 01:20.900 --> 01:22.490 on a model or a chain. 01:22.490 --> 01:27.290 So I could say what is the capital of Egypt? 01:27.890 --> 01:32.300 And then what you can see here is then we get this returned AI message. 01:39.170 --> 01:44.390 Also notice you can see there is a response metadata which can be quite useful for looking at token 01:44.390 --> 01:44.870 usage. 01:44.870 --> 01:47.120 So we'll look at the response metadata. 01:47.150 --> 01:52.010 You can see we have the token usage, the model name that we used, a system fingerprint and a Finnish 01:52.010 --> 01:52.430 region. 01:52.430 --> 01:53.960 And the log probabilities. 01:53.960 --> 01:58.430 You can access multiple different parts from the invoke and you could save those to your results. 01:58.430 --> 02:04.700 So I could do for example result chat, invoke dot content and then the result now becomes a string. 02:04.700 --> 02:05.930 So I can print the result. 02:05.930 --> 02:09.690 And you can see that the capital of Egypt is Cairo. 02:09.720 --> 02:10.530 Okay, great. 02:10.530 --> 02:12.300 So that's one way that we can do that. 02:12.300 --> 02:18.450 Another way that we'll want to invoke and call the chat model is by actually using messages. 02:18.450 --> 02:20.340 There are multiple different types of messages. 02:20.340 --> 02:24.210 For example system message, human message are the most common. 02:24.240 --> 02:29.700 The system message is quite important and that is most respected inside of the large language model. 02:29.700 --> 02:31.590 It's often where you'll put the instructions. 02:31.590 --> 02:34.710 The human message is the message from a user. 02:34.710 --> 02:38.250 You also do have the ability to add in AI messages as well. 02:38.250 --> 02:46.260 We're then going to have some text and we'll say what would be a good company name for a company that 02:46.260 --> 02:47.520 makes colourful socks? 02:47.520 --> 02:50.910 And then we're going to have a messages Python list. 02:50.910 --> 02:56.010 And inside that we're then going to put a system message which says you are a helpful assistant that 02:56.010 --> 02:59.640 generates company names, and then we're going to have a human message. 02:59.640 --> 03:03.030 And for that we're actually going to put in the text here. 03:03.030 --> 03:05.460 So inside of this content we're going to put that. 03:05.460 --> 03:09.930 And then what you'll see here is we're then following on with this pattern of chat dot invoke messages 03:10.110 --> 03:10.890 content. 03:10.950 --> 03:14.940 And then what we're then going to do is have a look and see what that gives us. 03:14.970 --> 03:16.590 So we'll then print the result here. 03:18.150 --> 03:24.870 And you can see what we're doing is we're now adding in a multiple messages into our chat model. 03:24.900 --> 03:25.320 Right. 03:25.350 --> 03:26.640 So it's got a system message. 03:26.640 --> 03:30.810 We're then adding a human message and we're putting in the text directly inside of that. 03:30.810 --> 03:37.740 Then we're invoking the chat model with that history of messages and getting the content out into the 03:37.740 --> 03:40.740 result variable, and then printing that out. 03:41.190 --> 03:45.750 There is one other way of doing this, which is a more legacy way of doing this. 03:45.750 --> 03:48.150 So for example, we could have the original text. 03:48.180 --> 03:52.140 We could then also just say, let's have a bunch of messages. 03:52.140 --> 03:59.100 And in that we could say the system message is equal to the content of text. 03:59.370 --> 04:05.160 And then we could also then have the chat model directly operate on those messages. 04:05.160 --> 04:07.500 So you can see Lang chain is deprecating this. 04:07.500 --> 04:09.390 So I wouldn't recommend using this. 04:09.390 --> 04:11.460 But you might see this in the code somewhere. 04:11.490 --> 04:17.340 In general you should favor the invoke function, and we'll see that there are lots of different types 04:17.340 --> 04:19.380 of functions you can use in lang chain. 04:19.380 --> 04:24.930 Dot invoke will only give you a single result, but in future videos you'll learn how to generate multiple 04:24.930 --> 04:25.800 results. 04:26.010 --> 04:30.750 Also, one of the other benefits that I just want to highlight about Lang Chain is that it's essentially 04:30.750 --> 04:35.400 a piece of glue around the large language model, so you can easily switch the model out. 04:35.400 --> 04:43.710 So for example, if we were to go to if we were to go to the Lang Chain community, click on the integrations. 04:43.710 --> 04:48.090 What you'll see is there's a variety of providers that you can easily interact with. 04:48.090 --> 04:54.450 So for example, if we wanted to interact with pinecone, that would be Lang chain Dash pinecone if 04:54.450 --> 04:57.840 we wanted to interact with for example, a other language model. 04:57.840 --> 05:00.750 So anthropic that would be lang chain anthropic. 05:00.750 --> 05:06.420 And then we can then go and have a look at that on PyPI and see this is how the package would be imported 05:06.420 --> 05:08.250 for a similar type of model. 05:08.280 --> 05:14.220 So then you would basically pip install lang chain anthropic port from lang chain anthropic. 05:14.220 --> 05:16.300 And I know it's chat anthropic. 05:16.300 --> 05:21.820 And then we would then create our model and we could call this chat anthropic and pass in and pass in 05:21.820 --> 05:22.960 a specific version of Claude. 05:22.990 --> 05:27.820 Then we could have our messages which is equal to a human message. 05:27.820 --> 05:33.010 And then we could also do what we had before where we're taking a chat model, invoking that with some 05:33.010 --> 05:34.720 messages, getting the content out. 05:34.720 --> 05:39.100 And you can see from here we're then easily able to swap the model out, right. 05:39.130 --> 05:41.500 So we're getting our sock names just like we were before. 05:41.500 --> 05:46.540 But the important point is that actually we've already got the messages defined so we could get rid 05:46.540 --> 05:47.140 of this. 05:47.170 --> 05:52.810 And essentially all we've done is we've just changed one piece of code here, the word model for the 05:52.810 --> 05:55.120 variable earlier on which we had for chat. 05:55.120 --> 05:59.170 So you can easily see that all of the rest of the code stays the same. 05:59.200 --> 06:04.690 We're only changing out the chat model, and Lang chain has built a variety of providers that allows 06:04.690 --> 06:08.200 you to easily switch models in and out via chat models. 06:09.400 --> 06:14.710 In the next video, we'll learn how to easily set up prompt templates which allow you to inject dynamic 06:14.710 --> 06:19.180 variables directly into your prompts using Lang Chain Chat prompt templates.