WEBVTT 00:00.240 --> 00:03.400 Let's define a variable called information. 00:03.400 --> 00:05.560 And let me put some placeholders here. 00:05.760 --> 00:09.440 And here we want to put here some text about Elon Musk. 00:10.880 --> 00:14.000 So let me go to Google and let me search Elon Musk. 00:14.520 --> 00:18.120 And let me take the information from Wikipedia. 00:18.400 --> 00:20.200 So let me just copy here. 00:20.760 --> 00:24.240 Let's copy here this text only the first part. 00:24.480 --> 00:25.680 And let me paste it here. 00:26.080 --> 00:29.280 So right here we have some information about Elon Musk. 00:29.280 --> 00:33.400 And this information here we want to eventually propagate to the LM. 00:33.840 --> 00:35.960 We want to do that through a prompt template. 00:36.160 --> 00:39.280 And for that let's write a template first. 00:39.560 --> 00:44.200 So it's going to be this template over here given the information. 00:44.200 --> 00:50.360 And right now we have curly brackets information about a person I want you to create a short summary 00:50.360 --> 00:52.920 into interesting facts about them. 00:53.320 --> 00:59.920 So from this string here you can probably guess that this information placeholder here, it's not going 00:59.920 --> 01:00.400 to stay. 01:00.400 --> 01:04.770 It's going to be substituted by the information variable. 01:04.770 --> 01:12.290 We have defined a couple of lines earlier with the Elon Musk information, but it's going to be filled 01:12.290 --> 01:13.250 in runtime. 01:13.810 --> 01:16.490 And if we want we can do this programmatically. 01:16.490 --> 01:21.650 And we're going to do this with the prompt template object of LinkedIn. 01:21.770 --> 01:24.970 And we initialize an object of prompt templates. 01:24.970 --> 01:28.410 And we initialize it with the template. 01:28.410 --> 01:31.930 So this is the summary template that we wrote before. 01:32.170 --> 01:39.970 And then we give it input variables which is going to be a list containing the keys of what we're going 01:39.970 --> 01:43.530 to be filling and plugging in in runtime. 01:43.730 --> 01:50.810 And this should match the placeholders between the curly brackets that we have in our template here. 01:51.290 --> 01:55.050 Now this is very much reminding us of an F string. 01:55.250 --> 01:57.650 And you might be thinking, why do we need all of this? 01:57.690 --> 02:00.290 And why don't we just use f strings instead? 02:00.810 --> 02:04.490 And this is because the prompt template is going to enforce that. 02:04.490 --> 02:12.330 We supply exactly the variables that we expect, and if we forget an input or misspell it, we get clean 02:12.330 --> 02:15.730 air instead sending us the broken prompt. 02:16.090 --> 02:23.010 It's also going to give us reusability and more clarity, because those prompts can become reusable 02:23.010 --> 02:27.010 in another link chain chain if we want to. 02:27.210 --> 02:29.730 And it's a first class citizen in link chain. 02:29.730 --> 02:33.370 So this means it's going to get logged, it's going to get traced. 02:33.370 --> 02:35.490 And we'll be showing this very very soon. 02:35.730 --> 02:40.610 And overall it's going to make our life debugging much more easier. 02:40.890 --> 02:46.570 And it's going to also be safer against prompt injection because it can enforce strict formatting and 02:46.570 --> 02:47.330 structure. 02:47.530 --> 02:51.850 And this is very powerful, especially when we're going to be using output parsers. 02:51.850 --> 02:55.850 And I know I'm talking a lot of things right now and a lot of buzzwords. 02:55.850 --> 02:59.050 And don't worry, you'll understand everything by the end of this course. 02:59.250 --> 03:05.980 But just to conclude, if strings tend to encourage just jamming the text here, which is fine, but 03:05.980 --> 03:10.180 if you want our code to be more reliable, we really want to use those prompt templates. 03:10.900 --> 03:15.060 All right, let's now create a OpenAI chat model. 03:15.060 --> 03:17.420 And I'm going to call it LLM. 03:18.340 --> 03:24.500 And this is the way we're going to interact with the OpenAI model GPT five. 03:24.660 --> 03:32.180 And this is what eventually is going to make the API calls to OpenAI to send the text and the prompts 03:32.180 --> 03:34.860 and to get back the response from Dlrm. 03:35.060 --> 03:37.700 And who's going to pay for all those API calls? 03:37.940 --> 03:44.980 And I remind you that in the previous video, we created an API key in OpenAI, which is connected to 03:45.020 --> 03:48.100 our account with our payment information configured. 03:48.300 --> 03:55.500 And the API key is stored in an environment variable, which is called OpenAI API key, which is loaded 03:55.500 --> 03:56.740 into the runtime. 03:56.780 --> 04:02.780 Now LinkedIn is going to search for this environment variable and is going to use this as credentials 04:02.780 --> 04:05.230 to talk with OpenAI API. 04:05.830 --> 04:07.750 And I give you an exercise. 04:07.750 --> 04:13.870 Maybe if you want, you can check out the source code of the chat OpenAI model and see where it happens 04:13.870 --> 04:14.270 there. 04:15.830 --> 04:19.830 And we initialize the chat model with temperature equals to zero. 04:20.230 --> 04:26.510 And temperature will control how random or creatives versus strict and deterministic. 04:26.550 --> 04:28.350 The model response is going to be. 04:28.750 --> 04:37.550 So low values between 0 and 0.3 is going to make the response deterministic, factual and probably repeatable. 04:37.910 --> 04:45.190 And it's good for summarization, for code, for instructions, for test and high values like above 04:45.230 --> 04:47.550 0.8 until one. 04:47.670 --> 04:50.350 It's going to get us very creative results. 04:50.350 --> 04:54.870 And it's good for poetry and fiction and out of the box ideas. 04:54.910 --> 05:00.630 And if you're interested about how temperature works, then feel free to check out the theoretical section 05:00.630 --> 05:02.190 where I explain this. 05:02.430 --> 05:02.830 All right. 05:02.830 --> 05:04.390 Let's go back to the code. 05:04.390 --> 05:09.840 And the second argument that we give is going to be model equals to GPT five. 05:10.400 --> 05:14.320 And now we're going to have a link chain chat. 05:14.320 --> 05:22.480 OpenAI object which eats under the hood, is going to be using the OpenAI SDK to make API calls to OpenAI 05:22.640 --> 05:23.800 with our API key. 05:24.120 --> 05:27.840 And now we can go and create our first chain. 05:28.400 --> 05:35.200 And we have the variable chain which is going to be summary prompt template the pipe operator. 05:35.200 --> 05:36.520 And then ln. 05:36.880 --> 05:39.400 So let me break down what's happening over here. 05:39.560 --> 05:44.880 And we're using something which is called the link chain expression language or LCL. 05:45.280 --> 05:54.760 And in this LCL syntax we create a chain by composing two components a prompt template and a large language 05:54.760 --> 05:55.280 model. 05:55.560 --> 06:00.000 Now the prompt template is going to come from the variable summary prompt template. 06:00.000 --> 06:02.800 And this is going to format input variables. 06:02.800 --> 06:10.520 And in this case information into a prompt string which will eventually be propagated to the LM, and 06:10.520 --> 06:19.680 the LM variable is a chat open AI object, which takes in an input a prompt string and generates a text 06:19.720 --> 06:20.400 response. 06:20.440 --> 06:27.920 Now, this pipe operator in an expression language is going to create a new runnable chain. 06:27.920 --> 06:35.560 So this is a new term runnable by connecting the output of the left component as an input to the right 06:35.560 --> 06:36.360 component. 06:36.640 --> 06:45.600 So summary prompt template pipe lm means that we first want to format the input using the prompt template, 06:45.800 --> 06:51.600 and then to pass the resulting prompt string into the LM to generate the response. 06:51.840 --> 06:56.440 So the resulting chain is called a runnable object. 06:56.440 --> 06:58.720 So this means we can invoke it. 06:58.760 --> 07:03.760 Use the invoke method with some input variables matching the prompt template. 07:03.760 --> 07:06.120 And here in this case it's going to be information. 07:06.640 --> 07:12.410 Now I know this is a lot to digest and to be honest, the LinkedIn expression language. 07:12.410 --> 07:15.650 I think this is the hardest thing to comprehend in LinkedIn. 07:15.650 --> 07:21.290 So if you don't understand what's happening here right from the first time, this is totally okay. 07:21.530 --> 07:25.970 We are going to dive deep into this and see the implementation itself, and we're going to have many 07:25.970 --> 07:27.250 examples during the course. 07:27.490 --> 07:32.410 And at this point of the course, I'm not expecting you to know every bit and byte of it. 07:32.450 --> 07:37.410 So we just need to know from a very high level that we read these from left to right. 07:37.410 --> 07:43.490 And here we are plugging in the result of formatting the prompt template into the LM. 07:43.530 --> 07:45.050 So when do we run this. 07:45.210 --> 07:50.010 So we do this using the invoke method that the runnable interface has. 07:50.010 --> 07:57.690 So what's going to happen here is that when we run change dot invoke we actually going to invoke the 07:57.730 --> 08:03.490 invoke method of the summary prompt template which is a class of prompt template. 08:03.690 --> 08:10.140 And when we invoke it, we're going to give it the input of the information key to be holding the information 08:10.140 --> 08:10.700 value. 08:10.700 --> 08:14.660 And in this case it's going to be the Elon Musk information. 08:15.140 --> 08:23.700 So once we do that then we're going to create a prompt value, which is sort of like a fancy word of 08:23.700 --> 08:24.540 saying a string. 08:24.740 --> 08:28.700 And that string is going to be piped to the LM method. 08:28.820 --> 08:37.180 And this pipe is actually some very clever way to go and call the LM invoke method, which it has with 08:37.180 --> 08:40.660 the input to be the prompt value. 08:40.660 --> 08:43.540 So the final prompt that we want to send to the LM. 08:43.700 --> 08:49.340 So this is just a clever way to chain everything up together and hence the name of link chain. 08:49.500 --> 08:55.820 So eventually when we're going to run this code we are going to take this string that you see right 08:55.820 --> 08:56.660 over here. 08:56.660 --> 08:59.620 We're going to plug in the information of Elon Musk. 08:59.780 --> 09:03.740 And this string is what is going to be sent eventually to the LM. 09:03.780 --> 09:04.220 Right. 09:04.260 --> 09:07.620 So I know it sounded at the beginning very, very complicated. 09:07.620 --> 09:10.350 But actually what it's doing is quite simple. 09:10.350 --> 09:15.870 And in the next video we're also going to be tracing this and see the exact objects and debugging it. 09:15.870 --> 09:18.350 So don't worry, you will understand this. 09:18.390 --> 09:25.030 And I think this is the hardest concept to comprehend in length chain and the length chain expression 09:25.030 --> 09:25.590 language. 09:25.750 --> 09:29.390 And it is totally okay that you do not understand it by now. 09:29.390 --> 09:35.550 And just a fun fact, even implementing and understanding how agents work is going to be much easier 09:35.550 --> 09:39.950 than understanding this piping operator in the chain expression language. 09:39.950 --> 09:45.750 All right, so just to summarize the code here, all you need to understand is that we're taking the 09:46.110 --> 09:47.750 variable of information. 09:47.750 --> 09:51.150 We're plugging it in to this string over here. 09:51.150 --> 09:56.750 And we're dynamically going to put the value of the information about Elon Musk. 09:56.950 --> 10:01.870 And this is the string that we're going to be sending eventually to the LLM and get a response to it. 10:01.910 --> 10:02.230 Right. 10:02.270 --> 10:04.750 So this is everything you need to know from this video. 10:04.790 --> 10:06.870 You can forget about gene expression language. 10:06.870 --> 10:08.150 You can forget about everything. 10:08.150 --> 10:11.070 As long as you understand this, this is okay. 10:11.190 --> 10:12.430 And we can continue. 10:12.950 --> 10:13.510 All right. 10:13.510 --> 10:18.950 So let me go now and finally run this code and let's see what results do we get. 10:18.950 --> 10:23.910 And I'm going to fast forward it a bit because we're going to use GPT four. 10:24.110 --> 10:30.350 And GPT five is a bit slow LLM it's very capable but it's very slow. 10:30.910 --> 10:32.790 So let me go fast forward it. 10:33.630 --> 10:39.830 And at the end we're going to print response dot content, which is going to be the final string that 10:39.830 --> 10:41.070 the LLM returns. 10:41.230 --> 10:45.270 So we'll be debugging and we'll be tracing all of this in the next video. 10:45.390 --> 10:47.350 So let's just wait to see the output. 10:47.830 --> 10:50.190 All right so we can see now the short summary. 10:50.190 --> 10:54.430 And we can see now two interesting facts on Elon Musk. 10:55.110 --> 10:57.270 So this looks great. 10:57.510 --> 11:00.670 And now you can read it in your own time. 11:00.670 --> 11:03.750 And in the next video we're going to be debugging this. 11:03.790 --> 11:06.030 We're going to be examining the objects. 11:06.030 --> 11:09.430 And we are going to be tracing it with Lindsmith. 11:09.430 --> 11:12.350 And it should clarify everything that we discussed now.