1 00:00:03,130 --> 00:00:06,610 What are few shot prompts? 2 00:00:06,610 --> 00:00:12,880 And, uh, how do we use the few shot prompt templates in landscape? 3 00:00:13,510 --> 00:00:23,080 So if you remember, uh, until now, we have been using mostly what we call zero shot prompts. 4 00:00:23,680 --> 00:00:24,430 Yes. 5 00:00:24,430 --> 00:00:28,060 Regular instructions to the LM model. 6 00:00:29,590 --> 00:00:36,850 This usually works well with the LM model, but sometimes this is not enough. 7 00:00:36,850 --> 00:00:43,900 And this kind of prompting generates some level of mistakes or hallucinations. 8 00:00:43,900 --> 00:00:52,690 So in order to improve the quality of the output provided by the LM model, we can use a more sophisticated 9 00:00:52,690 --> 00:00:54,760 kinds of prompts. 10 00:00:55,840 --> 00:01:05,560 Uh, the most popular is the few shot prompt, which includes not only instructions but also examples. 11 00:01:05,560 --> 00:01:13,240 So we provide the LM model a few examples of what we mean with our instructions, so he can be more 12 00:01:13,240 --> 00:01:15,970 aware of what we try to do. 13 00:01:15,970 --> 00:01:24,190 For example here here you have on the right if we are, uh, trying, uh, the LM model to classify 14 00:01:24,190 --> 00:01:31,270 the sentiment of a book review or a product review, we can provide a few examples of a positive review, 15 00:01:31,270 --> 00:01:33,280 negative review or neutral review. 16 00:01:33,280 --> 00:01:38,380 So the LM understands exactly what we are looking for. 17 00:01:39,160 --> 00:01:45,460 There is a little bit more sophisticated kind of prompt which is chain of thought. 18 00:01:45,460 --> 00:01:54,010 And this kind of prompt not only includes instructions and examples, it also includes explanations 19 00:01:54,010 --> 00:01:58,030 about the reasoning logic behind each of the examples provided. 20 00:01:58,030 --> 00:02:05,980 So if this was a chain of thought prompt, we would include here an explanation explaining the LM model, 21 00:02:05,980 --> 00:02:12,520 why this review is positive okay and why this review is negative, neutral, etc.. 22 00:02:12,520 --> 00:02:17,470 So the most popular kind of prompt is the first one. 23 00:02:17,710 --> 00:02:20,590 The second one is also very popular. 24 00:02:20,590 --> 00:02:22,270 The third one is less popular. 25 00:02:22,270 --> 00:02:29,410 So in this exercise we are going to focus on uh how to do the first two. 26 00:02:29,410 --> 00:02:37,090 And we will give you, you know, with the long chain documentation to experiment and practice further. 27 00:02:37,090 --> 00:02:46,660 So the first thing we do as usual is to, uh, load the dotenv file where our credentials are in order 28 00:02:46,660 --> 00:02:54,190 to, uh, connect with the OpenAI, uh, API and being able to connect with ChatGPT, etc.. 29 00:02:54,370 --> 00:02:59,710 So zero shot prompting is what we have been doing until now. 30 00:02:59,710 --> 00:03:05,050 So it's nothing new for you in this case, uh, we are importing prompt template. 31 00:03:05,590 --> 00:03:09,340 We are creating the, the the prompt template. 32 00:03:09,340 --> 00:03:17,980 We are a entering the user input and we are getting their response okay. 33 00:03:17,980 --> 00:03:27,640 So this is a small application when we want or we want the LM to classify the sentiment of our review 34 00:03:27,640 --> 00:03:28,060 okay. 35 00:03:28,060 --> 00:03:32,050 So we enter a review I love this course. 36 00:03:32,050 --> 00:03:34,390 And we get the response. 37 00:03:34,390 --> 00:03:35,440 The LM is telling us. 38 00:03:35,440 --> 00:03:37,120 This is a positive review. 39 00:03:37,120 --> 00:03:37,690 Excellent. 40 00:03:38,700 --> 00:03:42,900 Second kind of prompting few shot prompting. 41 00:03:43,590 --> 00:03:53,310 We can do this in many different ways, as I show you before, we could do it inside the prompt definition 42 00:03:53,310 --> 00:04:02,610 so we can build a prompt or a prompt template that has the same thing that the previous one had, and 43 00:04:02,610 --> 00:04:05,040 also including the examples. 44 00:04:05,040 --> 00:04:08,640 Or we can have a different approach. 45 00:04:08,640 --> 00:04:13,710 This second approach is using the few shot prompt template module. 46 00:04:13,920 --> 00:04:22,620 And in this approach what we do is first we create a variable including all the examples. 47 00:04:22,830 --> 00:04:27,840 And then we follow the the the the the classic approach. 48 00:04:27,840 --> 00:04:36,660 We create a template a classic template a, and then we create a second prompt template using the few 49 00:04:36,660 --> 00:04:38,340 shot prompt template module. 50 00:04:38,340 --> 00:04:45,300 And here the main difference is to use this examples uh settings okay. 51 00:04:45,300 --> 00:04:52,410 So in this example settings we need to refer the variable that we have created with the examples. 52 00:04:52,410 --> 00:04:54,150 And that's it okay. 53 00:04:54,150 --> 00:05:02,100 So for you to understand that we have different kinds of prompting that you can experiment and try different 54 00:05:02,100 --> 00:05:06,090 ones in a in the land change documentation.