WEBVTT 00:00.720 --> 00:04.350 -: The second principle to try is to specify the format. 00:04.350 --> 00:06.930 Define what rules you want the AI to follow, 00:06.930 --> 00:10.080 and set a required structure for the response. 00:10.080 --> 00:12.120 All right, so we're gonna look at this 00:12.120 --> 00:13.560 through a text models first, 00:13.560 --> 00:15.780 and then we're gonna go onto images. 00:15.780 --> 00:17.850 Looking for a list of 10 product names 00:17.850 --> 00:20.670 for a pair of shoes that can fit any foot size. 00:20.670 --> 00:22.500 And this works fine. 00:22.500 --> 00:24.420 If you wanted to engineer the response 00:24.420 --> 00:26.370 to make it more reliable at scale 00:26.370 --> 00:28.770 if you're running this hundreds of thousands of times. 00:28.770 --> 00:30.720 Like if you're making a product name generator 00:30.720 --> 00:31.920 that people can use. 00:31.920 --> 00:35.100 But then you want to put a little bit more work into it. 00:35.100 --> 00:38.220 We've applied all the five principles to this prompt here. 00:38.220 --> 00:39.510 And specifically we're gonna look at 00:39.510 --> 00:41.580 how we specified format. 00:41.580 --> 00:44.820 And so this is a product template if you wanna grab it. 00:44.820 --> 00:47.880 But specifically in the format where we've said 00:47.880 --> 00:50.850 what we want is this section here, 00:50.850 --> 00:53.130 we said comma separated list. 00:53.130 --> 00:57.030 And all of the examples also provide a comma separated list. 00:57.030 --> 01:00.390 And we've also then told it we need three product names. 01:00.390 --> 01:02.640 We don't want a big list of 10 things. 01:02.640 --> 01:05.970 Now the reason why this is important is that the response 01:05.970 --> 01:08.880 to a prompt isn't the end of the task typically. 01:08.880 --> 01:11.550 It's usually just the input of the next step in the chain, 01:11.550 --> 01:12.930 which could also be another prompt, 01:12.930 --> 01:15.843 or it could be a human created task as well. 01:17.190 --> 01:18.960 And what impact on the results does it have? 01:18.960 --> 01:20.910 Depending on how you want the format, 01:20.910 --> 01:22.890 you could have it as a numbered list. 01:22.890 --> 01:25.770 This is if we change the prompt to have a numbered list, 01:25.770 --> 01:27.330 it comes out like this. 01:27.330 --> 01:28.770 But then if you're a developer 01:28.770 --> 01:31.530 and you're working on a product, you typically want 01:31.530 --> 01:34.170 the data to be structured in a way that you can pass it 01:34.170 --> 01:37.380 and use it in the next step, like displayed in the UI. 01:37.380 --> 01:40.590 So JSON is a very common mode, 01:40.590 --> 01:44.520 and in fact is actually a mode you can use in the OpenAI API 01:44.520 --> 01:47.100 and force it to return JSON. 01:47.100 --> 01:48.420 And then there's also YAML, 01:48.420 --> 01:50.280 which is more human readable, 01:50.280 --> 01:53.520 a little bit easier to pass in some cases. 01:53.520 --> 01:56.940 Whatever the data format you need, you can specify that. 01:56.940 --> 01:58.620 And one of the main things you want to do 01:58.620 --> 02:00.900 with prompt engineering is test different ways 02:00.900 --> 02:02.370 to make sure it always comes back 02:02.370 --> 02:05.310 with the required response format. 02:05.310 --> 02:06.930 So let's look at this in image models, 02:06.930 --> 02:08.370 'cause it's a little bit different. 02:08.370 --> 02:09.630 We're gonna look at stable diffusion. 02:09.630 --> 02:12.540 So this is prompting stable diffusion XL. 02:12.540 --> 02:15.120 And we can see, you know, we have a shoe 02:15.120 --> 02:17.670 that fits any foot size and that's fine. 02:17.670 --> 02:20.670 But you can see one of the responses it came back 02:20.670 --> 02:22.350 with was like a diagram. 02:22.350 --> 02:23.760 We actually want product photography. 02:23.760 --> 02:25.530 So how do we specify that? 02:25.530 --> 02:27.270 With a bit of prompt engineering, 02:27.270 --> 02:30.000 we have this prompt that is much more reliable 02:30.000 --> 02:33.060 and giving the same format back every time. 02:33.060 --> 02:36.870 And the part that where we specified format specifically, 02:36.870 --> 02:39.330 is if we go past the template here 02:39.330 --> 02:41.550 in section here of the prompt, 02:41.550 --> 02:43.710 we have photography, extremely detailed, 02:43.710 --> 02:46.740 studio lighting, 3.5 millimeter DSLR. 02:46.740 --> 02:49.890 So this is a way for us to keep it very consistent. 02:49.890 --> 02:52.470 If we append this on the end of every prompt, 02:52.470 --> 02:54.090 then we're gonna get a pretty consistent 02:54.090 --> 02:55.773 visual style every time. 02:56.820 --> 02:58.740 We've defined what rules we would like to follow, 02:58.740 --> 03:00.710 we've, you know, told it what sort of structure 03:00.710 --> 03:02.460 of the response comes back. 03:02.460 --> 03:05.640 And in image prompting you can get a little bit complicated 03:05.640 --> 03:09.510 because sometimes the format also blurs into the style. 03:09.510 --> 03:11.370 If you use product photography, 03:11.370 --> 03:14.790 then that's gonna necessitate certain types of style. 03:14.790 --> 03:18.600 And similarly, if you tell it you wanna style of a painting, 03:18.600 --> 03:20.640 then it's gonna clash with product photography 03:20.640 --> 03:22.380 and sometimes you get unreliable results. 03:22.380 --> 03:25.320 So really a case of testing both the first principle 03:25.320 --> 03:26.940 and the second principle together, 03:26.940 --> 03:29.133 and see which combination works well. 03:30.750 --> 03:32.880 Okay, let's look at the impact on results. 03:32.880 --> 03:35.310 Product photography is not the only thing we could do here. 03:35.310 --> 03:37.590 There's different types of product photography. 03:37.590 --> 03:39.870 So in this case, if we made it more of a model shoot, 03:39.870 --> 03:41.337 we wanted someone actually wearing it 03:41.337 --> 03:43.320 and we could achieve that. 03:43.320 --> 03:45.000 If we wanted street wear, 03:45.000 --> 03:47.130 now that's a different vibe entirely, 03:47.130 --> 03:50.160 and that you can see how it changes the style of the shoe 03:50.160 --> 03:52.230 as well alongside the format. 03:52.230 --> 03:54.480 So it takes some testing to get what 03:54.480 --> 03:56.370 you want in terms of your vision. 03:56.370 --> 03:58.830 And then the final one here is like Vogue Magazine. 03:58.830 --> 04:02.070 You can see that we have the classic problem with AI 04:02.070 --> 04:04.710 where she's got three legs. (laughs) 04:04.710 --> 04:06.750 Again, with AI, you need to test, 04:06.750 --> 04:10.170 you need to see how different changes to the prompt 04:10.170 --> 04:13.170 come back in terms of reliability of the results. 04:13.170 --> 04:15.630 And then also the quality of the image. 04:15.630 --> 04:17.220 I think it's a very high quality image, 04:17.220 --> 04:20.130 but it's obviously AI given that she has three legs. 04:20.130 --> 04:23.100 Maybe that's part of the alien vibe, I'm not sure. 04:23.100 --> 04:25.890 But whatever it is, getting the right format back 04:25.890 --> 04:28.110 will help you achieve the final task. 04:28.110 --> 04:32.760 Because if you're using these images for a magazine shoot, 04:32.760 --> 04:35.160 or if you're using them on your website, 04:35.160 --> 04:37.410 or if you're using on social media, you need to be able 04:37.410 --> 04:40.590 to specify what type of format you're expecting. 04:40.590 --> 04:41.730 Is it a model shoot? 04:41.730 --> 04:44.070 Do you actually want someone wearing these shoes? 04:44.070 --> 04:46.680 Or do you just want some product photography 04:46.680 --> 04:48.063 of the shoes themselves?