WEBVTT 00:00.030 --> 00:00.863 -: Okay, so great. 00:00.863 --> 00:03.480 So you've got billing enabled, you have an API key, 00:03.480 --> 00:06.600 and you're ready to get started using OpenAI's API. 00:06.600 --> 00:08.880 What I want you to do is make sure you've downloaded 00:08.880 --> 00:10.800 and cloned to the GitHub repository. 00:10.800 --> 00:11.633 Then I want you to go 00:11.633 --> 00:15.270 to the openai_features_and_functionality folder. 00:15.270 --> 00:18.300 Then click on core_features_walkthrough. 00:18.300 --> 00:20.250 We're gonna go through and just walk through a couple 00:20.250 --> 00:21.360 of the core features. 00:21.360 --> 00:22.860 I'm not gonna write any code, 00:22.860 --> 00:24.420 but what we're gonna do is just go through 00:24.420 --> 00:26.100 each of the core services 00:26.100 --> 00:28.440 and look at what's happening roughly, 00:28.440 --> 00:29.820 and then in the future lessons, 00:29.820 --> 00:32.160 we'll dive into each individual service 00:32.160 --> 00:33.480 in a bit more detail. 00:33.480 --> 00:35.370 So installing a bunch of packages, 00:35.370 --> 00:39.270 we're then importing the os module and OpenAI module. 00:39.270 --> 00:42.390 Then we are creating a global variable for the model. 00:42.390 --> 00:46.500 In our case, we're gonna be using gpt-4.1-mini 00:46.500 --> 00:49.050 for most of the core features. 00:49.050 --> 00:51.180 Now, you will need to put your API key in here, 00:51.180 --> 00:53.610 so I'll leave that up to you to do that. 00:53.610 --> 00:55.200 The first thing we're gonna have a look at is, 00:55.200 --> 00:59.010 how do you generate text directly from OpenAI's API 00:59.010 --> 01:01.200 using the responses API? 01:01.200 --> 01:02.640 You'll see we have a prompt. 01:02.640 --> 01:06.270 So, "Write a one-sentence bedtime story about a unicorn." 01:06.270 --> 01:09.060 We then create our client.responses.create, 01:09.060 --> 01:11.880 we put in the model that we want to use and the input. 01:11.880 --> 01:15.420 And what you'll get back from that is this .output_text. 01:15.420 --> 01:18.840 You can also as well control the model's parameters 01:18.840 --> 01:21.180 by adding things like a temperature. 01:21.180 --> 01:24.450 So the lower will give you more deterministic outputs 01:24.450 --> 01:26.910 or very similar outputs again and again 01:26.910 --> 01:30.090 and a higher temperature will give you more random outputs. 01:30.090 --> 01:31.920 And you can go and run this cell 01:31.920 --> 01:34.080 and you'll see you get two responses here. 01:34.080 --> 01:36.540 The prompt for each one and the response. 01:36.540 --> 01:40.200 And all we're doing is using that response.output_text 01:40.200 --> 01:42.030 and that will give you some text 01:42.030 --> 01:45.000 specifically from the responses API. 01:45.000 --> 01:46.980 If we're having a look at the second example, 01:46.980 --> 01:49.350 we're using something called structured outputs. 01:49.350 --> 01:54.300 So you'll see we have the same client.responses.create, 01:54.300 --> 01:57.900 we have our model, the input is we have a system message, 01:57.900 --> 02:01.140 which remember is an instructional developer message 02:01.140 --> 02:05.580 that you can put into the start of a conversation 02:05.580 --> 02:08.130 and we're telling it that it's a generator AI, 02:08.130 --> 02:10.500 convert the user's input into a UI, 02:10.500 --> 02:13.080 and we're saying make a user profile form. 02:13.080 --> 02:16.230 Now, we've got here is something called JSON Schema. 02:16.230 --> 02:19.170 JSON Schema is a way for us to tell the model 02:19.170 --> 02:21.600 what we want the structure of that data to be 02:21.600 --> 02:24.480 when it comes back from the model's response. 02:24.480 --> 02:26.790 You'll see we have this type of JSON Schema, 02:26.790 --> 02:30.270 the name, the description, and we have the schema itself, 02:30.270 --> 02:31.770 which is the JSON schema 02:31.770 --> 02:34.080 from this bit all the way down to here. 02:34.080 --> 02:36.930 And we also have this strict is True. 02:36.930 --> 02:39.030 Now, if we go and run this, 02:39.030 --> 02:41.760 then what we get back comes from a string. 02:41.760 --> 02:43.710 So this output text is a string, 02:43.710 --> 02:45.360 and we're then loading that in 02:45.360 --> 02:47.970 directly from the json.loads package. 02:47.970 --> 02:50.460 And you'll now see that we have some structured data 02:50.460 --> 02:52.830 that's come out from OpenAI's API. 02:52.830 --> 02:54.390 This can be also done using 02:54.390 --> 02:57.300 a data validation package called pydantic. 02:57.300 --> 02:59.070 So we can import a base model, 02:59.070 --> 03:01.950 we can set up various models such as the step 03:01.950 --> 03:03.270 where we have an explanation 03:03.270 --> 03:06.660 and an output which have data types of string. 03:06.660 --> 03:09.480 We have some math reasoning which has a list of steps 03:09.480 --> 03:10.890 and a final answer. 03:10.890 --> 03:12.270 And then we can use, for example, 03:12.270 --> 03:14.280 the chat completions endpoint 03:14.280 --> 03:18.000 with the client.beta.chat.completions.parse. 03:18.000 --> 03:21.150 We pass in the model, the messages, and the model. 03:21.150 --> 03:24.063 And then we also put the response format as well. 03:25.380 --> 03:26.640 And what you'll see if you run this 03:26.640 --> 03:29.400 is you get a structured pydantic object. 03:29.400 --> 03:30.960 If I do math reasoning, 03:30.960 --> 03:35.490 what you can see is we can access this .parse 03:35.490 --> 03:37.710 and then we have a maths reasoning object 03:37.710 --> 03:39.480 where we can do the .steps 03:39.480 --> 03:41.610 and we can see here are all the different steps 03:41.610 --> 03:43.260 that the model decided to take 03:43.260 --> 03:46.530 when reasoning over that type of math problem. 03:46.530 --> 03:49.320 We can go in and have a look at the first one, for example. 03:49.320 --> 03:51.540 It's also possible to get ChatGPT 03:51.540 --> 03:53.490 to put an image in here. 03:53.490 --> 03:56.970 And then we will use the responses API with this input. 03:56.970 --> 03:59.370 And we're putting in a role of user 03:59.370 --> 04:01.177 and we're saying the content is, 04:01.177 --> 04:02.520 "What is in this image," 04:02.520 --> 04:05.730 and we have this type input_image and the image_url. 04:05.730 --> 04:07.740 And then when you run that block of code, 04:07.740 --> 04:11.460 what you'll see is ChatGPT will analyze in real time 04:11.460 --> 04:13.830 what is inside of that image. 04:13.830 --> 04:15.450 You also have text to speech. 04:15.450 --> 04:16.283 So we have this 04:16.283 --> 04:19.950 client.audio.speech.with_streaming_response.create. 04:19.950 --> 04:24.950 We have a different model for this called gpt-4o-mini-tts. 04:24.960 --> 04:27.330 We choose a voice that we're interested in 04:27.330 --> 04:29.100 and what text we would like 04:29.100 --> 04:31.770 to be converted into an MP3 file. 04:31.770 --> 04:34.110 And then what we do is have some instructions 04:34.110 --> 04:37.470 that help to standardize how that voice sounds. 04:37.470 --> 04:40.230 Then we use this response.stream_to_file 04:41.790 --> 04:44.790 and we provide a file path of speech.mp3. 04:44.790 --> 04:46.650 You'll then see if you look on the right, 04:46.650 --> 04:49.500 you then have a speech file which you can then play. 04:49.500 --> 04:52.110 AI: It's a wonderful day to build something people love. 04:52.110 --> 04:52.943 -: That's awesome. 04:52.943 --> 04:57.420 And we also have the ability when we have an MP3 04:57.420 --> 05:01.260 to convert that MP three directly back into transcription. 05:01.260 --> 05:03.540 So you'll see that we're just using the file 05:03.540 --> 05:05.010 that we just already made. 05:05.010 --> 05:06.360 And then we can take that file 05:06.360 --> 05:09.510 and you can see we can easily convert that back into text. 05:09.510 --> 05:12.420 The next concept is this idea of function calling. 05:12.420 --> 05:14.820 So what you can have is a function. 05:14.820 --> 05:16.800 You can set up a bunch of tools. 05:16.800 --> 05:19.350 So in this scenario we only have a Python list 05:19.350 --> 05:20.640 with one tool. 05:20.640 --> 05:24.300 We then create an input message, which is a role of user 05:24.300 --> 05:26.460 and has the content of, "What's the weather 05:26.460 --> 05:27.630 like in Paris today?" 05:27.630 --> 05:32.250 And we can tell ChatGPT that it has access to these tools. 05:32.250 --> 05:34.740 And what you'll see is ChatGPT, 05:34.740 --> 05:37.923 if we look at the response.output, 05:39.510 --> 05:42.930 you'll see that it decided it wants to call a tool, 05:42.930 --> 05:45.240 and then you as the developer will get access 05:45.240 --> 05:46.740 to the tool call, 05:46.740 --> 05:49.200 the arguments that it wants to call it with, 05:49.200 --> 05:51.420 which will be under the .arguments 05:51.420 --> 05:54.120 under this output[0], 05:54.120 --> 05:56.130 and you can see here these are the arguments 05:56.130 --> 05:58.950 that it wants to use, the latitude and longitude. 05:58.950 --> 06:00.660 And then we call the function. 06:00.660 --> 06:02.490 And after calling the function, 06:02.490 --> 06:04.830 we add onto the input messages 06:04.830 --> 06:07.110 the original tool call that it wanted to make, 06:07.110 --> 06:09.570 and also the result of that tool call. 06:09.570 --> 06:11.973 Then we can call ChatGPT again. 06:14.460 --> 06:17.137 After calling it again, it will then respond in text, 06:17.137 --> 06:20.310 "The current weather in Paris is 16 degrees Celsius." 06:20.310 --> 06:23.190 So this basically allows your AI models 06:23.190 --> 06:24.870 to have access to functions 06:24.870 --> 06:28.410 so that they can do work outside of textual environments. 06:28.410 --> 06:30.450 The next thing we have is reasoning models. 06:30.450 --> 06:33.030 So you can use a more powerful model 06:33.030 --> 06:35.400 if you're working on a very complicated task. 06:35.400 --> 06:36.450 The only difference here is 06:36.450 --> 06:38.100 that we've just switched the model name 06:38.100 --> 06:39.930 and we're using o4-mini here. 06:39.930 --> 06:41.970 Now, I will say this will take a little bit longer 06:41.970 --> 06:45.120 to come back, and you won't get immediate responses. 06:45.120 --> 06:48.720 So depending upon how fast you need to respond to your users 06:48.720 --> 06:50.790 and what cost you need to do, 06:50.790 --> 06:53.100 you might want to have a look at the trade-offs 06:53.100 --> 06:54.720 between using reasoning models 06:54.720 --> 06:56.730 versus using standard chat models. 06:56.730 --> 07:00.690 You can see this took around 15.5 seconds to come back. 07:00.690 --> 07:03.000 We also have this idea called embeddings, 07:03.000 --> 07:05.850 which is basically a numerical representation 07:05.850 --> 07:07.200 of what text is. 07:07.200 --> 07:09.150 You can see we have a bunch of words 07:09.150 --> 07:12.930 and then we can have a client.embeddings.create, 07:12.930 --> 07:15.540 choosing a embedding model that we want to pick, 07:15.540 --> 07:19.590 and the words that we want to turn into numerical numbers. 07:19.590 --> 07:21.360 We then extract the embeddings 07:21.360 --> 07:23.400 for each individual piece of embedding. 07:23.400 --> 07:26.610 We then have a function to calculate the .product, 07:26.610 --> 07:29.580 and we basically then create a similarity matrix 07:29.580 --> 07:32.100 for all of the .products between all of the numbers 07:32.100 --> 07:33.630 that get created. 07:33.630 --> 07:35.790 And if you run this, for example, 07:35.790 --> 07:38.580 you will see exactly what's happening here, 07:38.580 --> 07:41.550 where we have a similarity matrix using the .product, 07:41.550 --> 07:44.040 which is the cosine similarity, 07:44.040 --> 07:49.040 and each embedding has around 3,072 dimensions. 07:49.530 --> 07:50.820 There are 10 embeddings, 07:50.820 --> 07:53.640 and this is the similarity matrix between all of them. 07:53.640 --> 07:54.840 You can see, for example, 07:54.840 --> 07:57.810 so king and king is obviously completely the same, 07:57.810 --> 08:01.680 king and man has a 0.41 similarity. 08:01.680 --> 08:06.680 You'll see if we look at, for example, banana versus orange, 08:06.690 --> 08:09.030 that has a 45 similarity. 08:09.030 --> 08:11.910 It has quite a good one with apple as well, 08:11.910 --> 08:13.110 orange and apple, 08:13.110 --> 08:16.050 but it has less similarity against men and women, 08:16.050 --> 08:19.080 and therefore you can use this to see what type 08:19.080 --> 08:22.800 of similar words, you know, are similar to other phrases. 08:22.800 --> 08:27.210 So you can see king and queen are similar with 0.55 08:27.210 --> 08:30.450 versus apple and banana that has a lower similarity. 08:30.450 --> 08:33.690 King and apple even, again, has a lower similarity. 08:33.690 --> 08:35.190 So embeddings are useful 08:35.190 --> 08:37.560 for when we're doing different types of work 08:37.560 --> 08:40.710 where we need to go beyond just using the text. 08:40.710 --> 08:44.040 This is where we will turn the text into a numerical number 08:44.040 --> 08:46.500 and then when we use that for downstream tasks. 08:46.500 --> 08:48.630 Cool, I know this is just a whistle-stop tour 08:48.630 --> 08:51.390 and we're gonna go into more of the features in detail, 08:51.390 --> 08:53.460 but I just thought I have one notebook 08:53.460 --> 08:55.770 that kind of summarizes all of the core functionality 08:55.770 --> 08:57.450 that I would recommend looking at 08:57.450 --> 08:59.433 when using OpenAI's platform.