WEBVTT 00:00.160 --> 00:01.200 Hey there Eden here. 00:01.200 --> 00:04.600 And in this video we're going to be discussing structured output. 00:04.960 --> 00:10.600 So most of the time when we're going to be building agents and AI applications, we are going to be 00:10.600 --> 00:14.120 using lens and lens respond in text. 00:14.280 --> 00:20.160 Now if we want to take this text and we want to create a structure to it, and we want to downstream 00:20.160 --> 00:26.600 it into an application, maybe to serialize it, maybe to render it in the user interface, then we 00:26.600 --> 00:28.520 need to have structured output. 00:28.520 --> 00:36.120 So we don't just want text, we want maybe a JSON object or maybe a Pydantic object that we can go and 00:36.120 --> 00:41.560 we can downstream to our application and we can programmatically parse and work with. 00:41.600 --> 00:44.600 And this is what we're going to be doing in this video. 00:44.880 --> 00:49.520 And LinkedIn provides us a very easy way to do this with the create agent function. 00:49.560 --> 00:53.440 We need to simply add an argument which is called response format. 00:53.440 --> 00:57.520 And we're going to provide to this argument the schema that we want. 00:57.560 --> 00:59.000 It may be a JSON schema. 00:59.200 --> 01:01.280 It may be a Pydantic object. 01:01.560 --> 01:07.810 And then magically the response that the agent is going to output us, is going to adhere to that scheme 01:07.810 --> 01:09.530 or to that data object. 01:09.530 --> 01:14.530 So in this video, we're going to be creating a Pydantic object, and we're going to be creating a nested 01:14.530 --> 01:18.530 Pydantic object just to add a bit of complexity to what we're building. 01:18.530 --> 01:22.010 So the Pydantic object is going to be our agent response. 01:22.010 --> 01:24.850 And it's going to be the answer of the agent. 01:24.850 --> 01:26.610 So this is going to be one field. 01:26.610 --> 01:33.730 And it's also going to have a sources field which is going to be a list of sources Pydantic object as 01:33.730 --> 01:38.530 well, which is going to represent the URL where the agent took the answer from. 01:38.530 --> 01:41.850 So this is going to be where the agent grounded its answer from. 01:41.850 --> 01:46.610 And I want to remind you that in this video we're just reviewing interface. 01:46.890 --> 01:48.930 So we're going to see how we can use it. 01:48.930 --> 01:50.650 And to use the technology. 01:50.650 --> 01:54.530 And in the next section we're going to see how this is implemented under the hood. 01:54.850 --> 01:56.010 So let's go to the code. 01:58.530 --> 02:04.530 Let me go and import from typing the list type int. 02:05.290 --> 02:09.020 And I also want to import from pedantic. 02:09.580 --> 02:15.100 The base model class and the field class. 02:16.860 --> 02:23.740 And for those of you who are not familiar with pedantic, pedantic base model is going to give us a 02:23.740 --> 02:29.300 base class so we can inherit from in order to define a structured data scheme. 02:29.540 --> 02:37.980 So it's going to provide functionality like data parsing and serialization and automatic type validations. 02:38.180 --> 02:44.900 And the field class is going to allow us to add metadata to our models attribute. 02:44.900 --> 02:53.140 So we can add descriptions, which is going to be helpful for the Llms to understand what to put in 02:53.180 --> 02:53.860 that field. 02:54.580 --> 02:55.020 All right. 02:55.020 --> 02:59.180 So let me now go and create my pedantic object. 02:59.180 --> 03:06.180 And I want to start by writing an object to describe the sources that the agent use to get the answers. 03:06.540 --> 03:09.740 So I'm going to define a new class and I'm going to call it source. 03:10.180 --> 03:13.680 It's going to inherit from the pedantic base model. 03:14.080 --> 03:17.080 And let me write here in the description. 03:17.800 --> 03:21.600 Scheme for a source used by the agent. 03:22.360 --> 03:25.480 So now let me go and define the field. 03:25.480 --> 03:27.200 Let's call this field URL. 03:27.520 --> 03:29.360 And it's going to be a string. 03:29.640 --> 03:32.360 And here I'm going to use the field class. 03:32.360 --> 03:34.720 And I can add here a description. 03:34.960 --> 03:38.680 And here in the description I'm going to write the URL of the source. 03:40.600 --> 03:45.560 So the point of this class here is to represent the source of the answer. 03:45.960 --> 03:55.280 Now I want this source pedantic class to be a nested field in another pedantic class, which is going 03:55.280 --> 03:57.080 to be called agent response. 03:57.480 --> 04:04.920 So agent response is going to have a an answer, and it's going to have a list of a sources. 04:05.320 --> 04:05.840 All right. 04:05.840 --> 04:09.440 So let's go and create this agent response class. 04:09.640 --> 04:12.680 So I'm going to write class agent response. 04:12.920 --> 04:15.770 And it's going to also inherit from base model. 04:16.690 --> 04:20.690 And let's go and add the description for this class. 04:20.930 --> 04:26.170 So this is going to be the scheme for the agent response. 04:26.330 --> 04:30.810 So it's going to have an answer and it's going to have the sources. 04:31.410 --> 04:36.250 So let me now add the attribute of an answer which is going to be a string. 04:36.730 --> 04:45.250 And let me write in the description that this is going to be the agent's answer to the query. 04:46.170 --> 04:49.970 I have a typo here, but don't worry, the LLM can handle it easily. 04:50.850 --> 04:57.250 So the second attribute of this class is going to be the sources. 04:57.530 --> 05:00.490 So let me write here sources. 05:02.650 --> 05:05.090 And sources is going to be a list. 05:06.250 --> 05:09.050 And it's going to be a list of source. 05:09.530 --> 05:11.730 Now let me create a field class. 05:11.730 --> 05:17.170 And here let me write that the default factory is going to be list. 05:17.330 --> 05:21.620 And this will make that when we create an agent response class. 05:21.620 --> 05:27.220 If we don't provide the sources, then the sources attribute is going to be populated with an empty 05:27.220 --> 05:27.700 list. 05:28.140 --> 05:30.180 And let's add the description here. 05:30.500 --> 05:35.460 And here is going to be the list of sources used to generate the answer. 05:44.220 --> 05:44.860 Alrighty. 05:44.860 --> 05:53.100 So now we want our agent to not return us a string like we saw before, but we want our agent to return 05:53.100 --> 05:58.540 us an agent response object so we can then downstream it into an application. 05:58.540 --> 06:04.220 Maybe we want to serialize it and then return it in a response from a server. 06:04.220 --> 06:07.940 And then we want to render this response in the front end. 06:08.100 --> 06:10.980 So we want our agent to return this data structure. 06:11.180 --> 06:18.460 And to do that all we need to do is add to the create agent function the argument of response format. 06:18.660 --> 06:24.230 And we want the response format to be the agent response class that we created earlier. 06:24.550 --> 06:25.830 And that's it. 06:25.870 --> 06:27.510 It's going to work like magic. 06:27.710 --> 06:32.990 And the agent response now is going to be of type agent response here. 06:32.990 --> 06:35.790 So it's going to have all the fields that we requested. 06:35.790 --> 06:38.190 And this is going to work really really well. 06:38.190 --> 06:40.470 And we can downstream it into an application. 06:40.510 --> 06:40.990 Okay. 06:41.030 --> 06:42.590 So let's go and run it now. 06:42.590 --> 06:45.070 And let me show you how it's going to be looking like. 06:45.430 --> 06:47.750 So I'm going to run it in debug mode. 06:47.750 --> 06:51.310 And I'm going to put a breakpoint right before we print it. 06:52.070 --> 06:53.750 And let me fast forward it a bit. 06:54.150 --> 07:01.030 And now in the result we have now a new key here which is called structured response here. 07:01.190 --> 07:05.510 And the type of structured response is going to be agent response. 07:05.710 --> 07:12.070 So it's going to have the answer field and it's going to have the sources field which is going to be 07:12.070 --> 07:13.470 a list of sources. 07:14.030 --> 07:14.310 Right. 07:14.310 --> 07:19.350 So let me go now to the debug console and let me show you this in another eye. 07:19.670 --> 07:28.210 So let's get the type of results in the key of Structured response. 07:30.330 --> 07:32.490 So it's going to be agent response class. 07:32.770 --> 07:34.290 So let me go and print it. 07:34.290 --> 07:35.890 So let's go and sorry. 07:35.930 --> 07:41.090 Let's go and print this um this variable here. 07:41.650 --> 07:45.810 So now we can see all the information we have here the answer field. 07:46.090 --> 07:47.730 So let's go check this out. 07:47.730 --> 07:49.010 And this is the answer. 07:53.810 --> 07:59.810 And let's go and check out the sources so we can see the sources is a list. 07:59.810 --> 08:02.730 So let's go and check one element of the list. 08:03.170 --> 08:04.690 Let's go to that URL. 08:04.690 --> 08:05.770 Let's open it. 08:08.210 --> 08:12.250 We can see this is a job application which is not actually accepting candidates. 08:12.250 --> 08:15.650 But still it's looking for link chain developers. 08:16.930 --> 08:23.410 So this is how to use the response format argument in the link chain agent. 08:23.650 --> 08:26.650 So it's going to return us a structured response here. 08:27.050 --> 08:27.290 Right. 08:27.330 --> 08:34.020 So right now this works like a black box, so we do not know how it's working under the hood. 08:34.060 --> 08:37.220 We are going to figure this out in the next section. 08:37.420 --> 08:43.060 So right now I want to open the trace and show you how it's going to look there. 08:43.580 --> 08:49.620 So let me go and open blacksmith and let me go to the last trace here. 08:50.140 --> 08:51.780 And we can see. 08:51.780 --> 08:56.580 Let me go to the very bottom here to the last, last result here. 08:56.580 --> 08:58.620 So we can see we have this additional field. 08:58.620 --> 09:00.660 We have a structured response. 09:02.180 --> 09:04.460 So it's going to have the answer field. 09:04.460 --> 09:06.420 And this is going to contain the answer. 09:06.620 --> 09:12.620 And then we're going to have the sources field which is going to be a list of source objects. 09:12.860 --> 09:18.940 And each object has the URL field which is going to contain the LinkedIn URL for the job posting. 09:18.980 --> 09:19.300 Okay. 09:19.340 --> 09:22.540 So very similar to what we saw in the debugger. 09:22.540 --> 09:24.980 So now we're seeing it in the trace. 09:25.260 --> 09:27.220 Now how is this working under the hood? 09:27.260 --> 09:28.620 I'm going to give you a quick hint. 09:28.620 --> 09:30.580 It's going to be using function calling. 09:30.580 --> 09:33.460 But more on this on the next couple of sections.