WEBVTT 00:00.180 --> 00:01.013 -: In this video, 00:01.013 --> 00:03.840 we're gonna cover a concept called delimiters. 00:03.840 --> 00:07.620 Delimiters allow you to tell ChatGPT what part 00:07.620 --> 00:09.870 of the input you would like it to operate on. 00:09.870 --> 00:11.730 Reading the prompt from start to the finish, 00:11.730 --> 00:16.050 we have summarized the text delimited by triple quotes. 00:16.050 --> 00:18.960 So you can see we've put these triple quotes at the start 00:18.960 --> 00:21.210 and end of where we would like to put some text. 00:21.210 --> 00:23.610 And obviously, when we put some real text there, 00:23.610 --> 00:25.380 we're actually gonna get a real summary. 00:25.380 --> 00:29.130 The key to this is realizing we've used the word delimited, 00:29.130 --> 00:30.450 triple quotes. 00:30.450 --> 00:32.310 And then the language model is gonna know 00:32.310 --> 00:34.710 that whatever is inside of our prompt, 00:34.710 --> 00:36.720 which contains these triple quotes 00:36.720 --> 00:38.220 is where it's gonna summarize. 00:38.220 --> 00:40.440 We could also do different type of delimiters. 00:40.440 --> 00:42.660 So rather than using triple quotes, 00:42.660 --> 00:45.000 we could actually delimit all of the different bits 00:45.000 --> 00:47.070 of information such as this. 00:47.070 --> 00:49.140 So let's walk through this prompt. 00:49.140 --> 00:49.973 We have, 00:49.973 --> 00:51.277 "In the following scenario, 00:51.277 --> 00:54.037 "you'll be given an extract of a dissertation 00:54.037 --> 00:56.197 "and a proposed title for it. 00:56.197 --> 00:58.657 "The name of the thesis should not only inform the reader 00:58.657 --> 01:00.937 "about its subject matter in a succinct manner, 01:00.937 --> 01:02.707 "but also be engaging. 01:02.707 --> 01:05.497 "Should the given title fall short of these requirements, 01:05.497 --> 01:08.490 "kindly offer five different alternatives." 01:08.490 --> 01:09.323 And then we have, 01:09.323 --> 01:13.350 "Please input abstract and please input title." 01:13.350 --> 01:14.670 So what's great about this format 01:14.670 --> 01:16.800 is then we can put the abstract here 01:16.800 --> 01:18.180 and then we could put the title, 01:18.180 --> 01:20.670 which is "Chromosome Structures" in here. 01:20.670 --> 01:24.480 So we're able to use a multiple delimiter approach 01:24.480 --> 01:28.620 by having what we are delimit by either title here. 01:28.620 --> 01:29.820 And if we scroll up, 01:29.820 --> 01:32.610 then you'll see we've got the abstract as the top delimiter. 01:32.610 --> 01:33.443 As well as that, 01:33.443 --> 01:34.800 it's also important to note 01:34.800 --> 01:37.740 that we have attached these delimiters, right? 01:37.740 --> 01:39.690 So we should ideally have title 01:39.690 --> 01:41.640 and abstract capitalize 01:41.640 --> 01:43.350 so that when we're referring to abstract, 01:43.350 --> 01:45.660 we're referring to what's after the delimiter. 01:45.660 --> 01:47.926 And when we're referring to the capitalized title, 01:47.926 --> 01:50.250 ChatGPT knows what to refer to. 01:50.250 --> 01:51.870 Going back to our first example, 01:51.870 --> 01:54.510 let's have a look at a slightly more complicated prompt 01:54.510 --> 01:57.000 and walk through it step-by-step. 01:57.000 --> 01:58.867 So we have a prompt here which says, 01:58.867 --> 02:01.950 "Summarize the text delimited by triple quotes." 02:01.950 --> 02:04.260 And we've got, "My father is happy." 02:04.260 --> 02:06.360 Then compare the generated summary 02:06.360 --> 02:09.337 to the text delimited with square brackets, 02:09.337 --> 02:10.680 "My dad is happy." 02:10.680 --> 02:13.320 Provide a final answer on whether the summary 02:13.320 --> 02:16.380 and the square bracket text are similar in meaning. 02:16.380 --> 02:18.900 So let's give this a run and see what happens. 02:18.900 --> 02:20.820 It's firstly created a summary, 02:20.820 --> 02:22.950 as well as that we've got a comparison, 02:22.950 --> 02:24.690 and it's come up with a final answer 02:24.690 --> 02:25.950 that the generated summary 02:25.950 --> 02:28.470 and the text limited by square brackets 02:28.470 --> 02:30.060 are similar in meaning. 02:30.060 --> 02:32.220 What's great about using delimiters here 02:32.220 --> 02:35.137 is that we've specifically told ChatGPT, 02:35.137 --> 02:35.970 "First, 02:35.970 --> 02:39.067 "you need to summarize the text that's in triple quotes 02:39.067 --> 02:41.347 "and then compare that generated summary 02:41.347 --> 02:43.560 "to the square brackets text." 02:43.560 --> 02:45.390 By adding multiple delimiters, 02:45.390 --> 02:48.540 we can clearly separate out different parts of the task 02:48.540 --> 02:51.300 so that ChatGPT can work more effectively for you. 02:51.300 --> 02:52.133 Finally, 02:52.133 --> 02:54.450 it's worth noting that delimiters are universal 02:54.450 --> 02:56.130 and can be applied to any different type 02:56.130 --> 02:57.330 of large language model. 02:57.330 --> 02:59.370 For instance, here we have GPT-4, 02:59.370 --> 03:01.920 and we're using that same delimiters concept 03:01.920 --> 03:03.690 where we have some various articles, 03:03.690 --> 03:07.110 which are separated out with the article delimiter. 03:07.110 --> 03:10.470 And then we've given a system message on the left hand-side, 03:10.470 --> 03:13.440 telling that we will be providing a pair 03:13.440 --> 03:16.530 of articles delimited by XML tags. 03:16.530 --> 03:18.813 So you can use this method anywhere you like.