WEBVTT 00:00.020 --> 00:00.680 Hey, welcome. 00:00.680 --> 00:05.660 In this video, you're going to cover how you can get structured data out directly from OpenAI using 00:05.660 --> 00:06.920 their SDK package. 00:06.920 --> 00:09.050 We'll look at how you can do this with tool cooling. 00:09.050 --> 00:13.010 Also how you can do it using the response format parameters inside of OpenAI. 00:13.040 --> 00:17.420 Firstly, structured outputs give almost 100% accuracy. 00:17.420 --> 00:19.250 This is a massive improvement. 00:19.250 --> 00:23.180 Rather than relying on prompting to give a structured data back, we're going to look at two different 00:23.180 --> 00:27.740 ways that you can get structured outputs, both using function calling, which is using tools, and 00:27.740 --> 00:32.660 also then as well as that using a new option called the response format parameter. 00:32.690 --> 00:36.950 Firstly, we're going to import a couple of libraries and show you how to do it using the response format. 00:37.130 --> 00:40.190 So we will import Pydantic base model the field. 00:40.220 --> 00:47.180 We'll also import the OS and we'll also import the OpenAI package. 00:48.530 --> 00:53.870 Now once you've done this you will need to set your environment key so you can do that directly here. 00:53.870 --> 00:57.470 Or you can use an environment variable file or do it on your terminal. 00:57.500 --> 01:02.070 Also we're going to create a Pydantic model that will be used to extract some structured data. 01:02.070 --> 01:08.820 So we'll have a class of key points, which is a base model that will take key points, which is a list 01:08.850 --> 01:10.410 of strings. 01:10.410 --> 01:19.080 And each of these will be a field with a description equal to a list of key points extracted from the 01:19.080 --> 01:19.890 text. 01:20.010 --> 01:20.490 Okay. 01:20.520 --> 01:23.550 So we've given the description of what that field should do. 01:23.550 --> 01:26.370 And then we're going to set up an open AI client. 01:27.390 --> 01:30.360 And then once we've done that we're then going to create a completion request. 01:30.390 --> 01:32.790 Now to do the response format. 01:32.790 --> 01:37.710 At the moment at this current time of filming, you will need to do the client beta chat completion 01:37.980 --> 01:38.790 pass. 01:40.950 --> 01:46.380 And then you will then provide a model that will allow for structured outputs via the API. 01:46.410 --> 01:47.940 So we'll use GPT for a mini. 01:47.970 --> 01:52.050 We will also do a Messages Python list. 01:52.050 --> 01:56.610 And we will have one message in here, which is a role of user. 01:56.730 --> 01:59.340 And the content of that message will be. 02:00.130 --> 02:08.080 You know, I will say James is awesome and likes to learn code and new things. 02:09.820 --> 02:14.050 And then notice here we're going to add something new that you might not have seen before in OpenAI, 02:14.080 --> 02:15.880 which is this response format. 02:15.880 --> 02:18.370 And we're providing a format here. 02:18.370 --> 02:22.240 When we get our result back, we get a past chat completion. 02:22.240 --> 02:24.250 So let's go and have a look at what this is. 02:25.090 --> 02:27.460 You can see we've got this dot key points. 02:27.460 --> 02:29.920 So if we go in here we have a look. 02:29.920 --> 02:36.250 Notice how we've got the structured data that would allow us to form that Pydantic model. 02:36.250 --> 02:37.540 So I could do something like this. 02:37.570 --> 02:43.000 Key points dot model dot validate JSON. 02:43.000 --> 02:46.810 And you can see here we've then created a Key Points object. 02:46.810 --> 02:50.590 And we could store that directly inside of a variable. 02:50.590 --> 02:52.210 So we could put like key points. 02:53.080 --> 02:54.730 And then we can then print that out. 02:54.730 --> 03:00.440 And now we've got our structured data that we've got back which is returning a key point Pydantic model. 03:00.440 --> 03:07.700 So this is really great if you want to easily extract structured data from text, and you'll use it 03:07.700 --> 03:12.560 with this response format as a parameter that goes into the parse function. 03:12.590 --> 03:15.020 Be aware that you can inherit from other models. 03:15.020 --> 03:21.470 So we could have key points and summary which could inherit from the key points. 03:21.830 --> 03:28.070 And the key points could also then have a summary which will then have a description of. 03:28.160 --> 03:30.950 We'll put a summary of this text. 03:31.760 --> 03:37.520 And this can be quite useful because then we can change the response format to have key points and summary. 03:37.550 --> 03:39.470 Go in here into the response format. 03:39.470 --> 03:40.760 And then you'll see. 03:40.790 --> 03:45.920 Now what we would do is we would validate that model with the key points and summary. 03:45.950 --> 03:48.530 And you can see we've got both the key points. 03:48.530 --> 03:52.280 And we also have a summary of that specific piece of text. 03:52.280 --> 03:57.920 This is the quickest way to extract structured outputs using OpenAI's package. 03:57.960 --> 04:03.570 We'll have a look at a secondary way that you can do this by using OpenAI's tools, which allow you 04:03.570 --> 04:09.090 to do this whilst allowing you to answer original questions alongside structured outputs. 04:09.120 --> 04:13.650 We're going to build a database query analyzer, which will extract structured data. 04:13.650 --> 04:16.410 In order to get back that we can operate on our database. 04:16.440 --> 04:20.850 We'll give OpenAI some information about the different types of tables that we have. 04:20.850 --> 04:26.370 And also then we're going to build some dynamic pydantic models and enums so that we can encapsulate 04:26.370 --> 04:27.780 all of this with tool use. 04:27.810 --> 04:30.150 To start with, we need to import the enum. 04:31.500 --> 04:34.740 We also need to import typing modules for union. 04:35.730 --> 04:43.320 And then you'll also then going to need to import Pydantic and import OpenAI. 04:44.880 --> 04:48.630 And from OpenAI we're also going to import OpenAI. 04:48.930 --> 04:52.440 Now we're going to set up a client if you don't already. 04:52.470 --> 04:54.720 And I'm just going to put up here. 04:54.990 --> 05:01.540 Import OS If you don't already have an environment variable of your API key, you'll need to set this 05:01.540 --> 05:01.780 here. 05:01.780 --> 05:04.750 So remember to set your API key. 05:06.100 --> 05:10.630 I already have one set, so I'm going to leave this cell and I'm not going to execute it. 05:10.660 --> 05:15.880 The next thing that we're going to do is set up the specific type of table that we want OpenAI to query. 05:15.910 --> 05:20.590 The first thing that we're going to do is we're going to build our table and column enums, so that 05:20.590 --> 05:24.070 OpenAI knows easily what table and column values it's working with. 05:24.100 --> 05:28.360 Firstly, we'll create a new class called table which is a string of enum. 05:28.360 --> 05:32.350 And in there we're just going to put fruits is equal to fruits. 05:37.240 --> 05:42.490 Then we'll create a second class called column which is also going to be a string enum. 05:43.690 --> 05:48.430 And inside of that we're going to have a range of different column names such as the ID column name 05:48.430 --> 05:49.300 color. 05:49.930 --> 05:52.210 We're also going to have ripeness. 05:55.540 --> 05:57.710 We're going to have a picture. 06:01.730 --> 06:03.440 We're also going to have sold art. 06:06.260 --> 06:07.220 Stored art. 06:09.530 --> 06:10.700 And discarded art. 06:10.730 --> 06:15.020 Next you're going to set up a different type of enum, which is called an operator. 06:15.020 --> 06:20.450 And this is going to give the ability for you to extract what types of operations might need to be done 06:20.450 --> 06:22.910 to filter the data from the SQL database. 06:22.910 --> 06:27.650 So we're going to have a new class called operator which is a string of enum. 06:27.650 --> 06:31.820 And then we're going to list some of the normal edge cases that you would find when using a SQL database, 06:31.820 --> 06:38.210 such as equal to greater than is equal to greater than, we'll have less than is equal to less than. 06:38.210 --> 06:40.250 We'll also have less. 06:40.460 --> 06:43.250 Less than or equal to is equal to this. 06:43.430 --> 06:47.480 And we'll also have greater than or equal to is equal to this string operator. 06:47.480 --> 06:50.780 And we'll also have a not equal to operator as well. 06:50.780 --> 06:55.760 Now that we have our operator we're also going to create a order by string enum. 06:56.450 --> 06:59.570 And in that we're going to have two different types of keys. 06:59.570 --> 07:02.540 For our enum we're going to have the ask which stands for ascending. 07:03.320 --> 07:07.100 And we're also going to have the desk which stands for descending. 07:08.450 --> 07:09.110 Okay cool. 07:09.110 --> 07:13.220 So now that we've got the order done, the next thing that we're going to do is we're going to allow 07:13.220 --> 07:14.990 for a dynamic value. 07:14.990 --> 07:19.940 And that is basically just going to be a base model. 07:19.940 --> 07:24.770 So a Pydantic based model, and they're going to have a column name which will be equal to a type of 07:24.770 --> 07:25.550 string. 07:25.730 --> 07:26.000 Okay. 07:26.030 --> 07:28.010 So we've set up a dynamic value. 07:28.160 --> 07:33.170 The final two pydantic models that we need is a condition which will be a base model. 07:33.170 --> 07:36.620 And this is where we start integrating things which we've written before. 07:36.620 --> 07:38.780 So we have a column which will be a string. 07:38.780 --> 07:41.090 But we then also use the operator. 07:41.090 --> 07:44.570 So the operator here will be equal to a type of operator. 07:44.810 --> 07:51.200 And then as well as that we're going to have a value for this condition which will be either a string 07:53.000 --> 07:55.170 int or dynamic value. 07:58.440 --> 07:59.010 Okay. 07:59.010 --> 08:01.110 Now we can write the final thing which is the query. 08:01.110 --> 08:02.640 So we're going to have a query. 08:02.670 --> 08:07.470 And in our query what we're going to take is the table name which is the type of table. 08:07.500 --> 08:11.460 We are also going to take the columns which is a list of column. 08:12.210 --> 08:17.310 And then we also then have the conditions which is a list of conditions. 08:18.240 --> 08:23.610 And then we have a order by which is equal to an enum of order by. 08:25.260 --> 08:28.950 Once we've written all these Pydantic models, we can then query this directly. 08:28.950 --> 08:30.630 By using an open AI client. 08:30.660 --> 08:32.670 We'll create an open AI client here. 08:32.760 --> 08:36.420 Then the next thing that we're going to need to do is we're going to need to set up a completion. 08:36.420 --> 08:41.670 So I'm going to type completion is equal to client dot chat dot completions dot create. 08:41.670 --> 08:45.270 And then inside of here we're going to use the model GPT four zero. 08:46.170 --> 08:48.060 And then we're going to have messages. 08:49.830 --> 08:52.590 And then inside that we're going to have two Python dictionaries. 08:52.590 --> 08:56.410 The first one which is the role system setting the system message. 08:56.410 --> 09:02.830 And we have a content which is you are a helpful assistant that extracts information from a user request. 09:03.010 --> 09:06.820 We're going to have a second one which says Role User content. 09:06.820 --> 09:10.600 I want to query the fruits table to get all of the apples that are ripe and sold. 09:10.630 --> 09:13.150 After 2024 0101. 09:13.150 --> 09:16.450 We're then just going to print this message and see what comes out. 09:16.450 --> 09:19.570 And you can see we've got this information here. 09:19.600 --> 09:21.940 You'll notice is we haven't got the structured data back. 09:21.940 --> 09:25.420 And that's because we also have to add the tools to load the tools properly. 09:25.420 --> 09:26.920 We need to make a new tools. 09:26.920 --> 09:32.110 And then we'll do OpenAI dot Pydantic function tool. 09:32.110 --> 09:34.330 And then we put the query inside that. 09:34.960 --> 09:40.360 And then what you'll see is OpenAI is wrapping that Pydantic model and it's making it into a tool call. 09:40.360 --> 09:45.550 And then what you can see is if we look at the arguments, these are the arguments of the function. 09:45.550 --> 09:50.230 Also if we have a look at the tool call in general, you can see we've got things like the id, the 09:50.230 --> 09:53.320 function, the args that the function is going to be called. 09:53.390 --> 09:56.750 And also we can look at that as well in greater detail. 09:56.750 --> 09:59.330 So again you can see here these are the arguments. 09:59.330 --> 10:01.010 These are the bits of structured data. 10:01.010 --> 10:04.370 And returning that in a JSON format. 10:05.390 --> 10:10.610 Once we load the import JSON we can get these out by wrapping this with Json.loads. 10:10.610 --> 10:12.680 And you can see this is the structured data. 10:12.680 --> 10:15.080 So it's decided it needs to query the table name. 10:15.080 --> 10:18.320 It's going to use these columns and it's going to have these conditions. 10:18.320 --> 10:24.200 So the column which is a type of operator is equal to and the value is going to be an apple. 10:24.230 --> 10:29.450 It's also going to have another condition which is the ripeness operator is going to be equal to ripe. 10:29.450 --> 10:33.290 And then we've also got a column which is sold at. 10:33.290 --> 10:38.630 And the operator that we're going to use for that is greater than this value is 2024 0101. 10:38.630 --> 10:46.430 So you can quite easily see how we've made our own domain specific language using this Pydantic query, 10:46.430 --> 10:52.400 so that we're able to extract structured information from user requests about data in your back end. 10:52.400 --> 10:57.210 Now, the reason why you might want to use this way of extracting structured data specifically using 10:57.210 --> 11:03.450 Pydantic models is potentially you might use the same code to respond in a natural language way. 11:03.480 --> 11:06.270 So for example, if you copy the code above. 11:06.300 --> 11:08.250 Delete the client line. 11:08.280 --> 11:12.540 Then after that, all I want you to do is change the user to. 11:12.570 --> 11:14.370 Hi, how are you? 11:14.400 --> 11:21.450 So we're going to print the completion dot choices zero dot message. 11:22.530 --> 11:26.940 And what you'll see here is we get a chat completion message with the content. 11:26.940 --> 11:32.520 So I can then for example do the content and you'll see that we're getting a message back from OpenAI. 11:32.550 --> 11:39.570 So the fact is that you won't always call the tools, but these are able to be called if a specific 11:39.570 --> 11:41.460 query comes in that is useful for that. 11:41.460 --> 11:44.160 So that's why you would use the tools in this video. 11:44.160 --> 11:49.320 You've learned about both the response format and also how to do tool use, which can either respond 11:49.320 --> 11:53.010 to the user or it can extract structured data via tools.