WEBVTT 00:00.360 --> 00:04.920 Let me now create a new file and I'm going to call it schemas dot p y. 00:05.880 --> 00:11.840 And this file is going to hold the pydantic classes that we want to get structured output of. 00:12.080 --> 00:18.680 And by the way when we installed Linkchain we implicitly installed Pydantic because Linkchain uses Pydantic 00:18.680 --> 00:20.640 a lot in our implementation. 00:21.360 --> 00:23.120 So just to give you a quick overview. 00:23.360 --> 00:28.680 Pydantic is a Python library for data validation using Python typings. 00:28.960 --> 00:35.520 So it automatically validates the data, converts types when possible, and provides clear error messages 00:35.520 --> 00:37.880 when the data doesn't match our descriptions. 00:37.960 --> 00:45.280 And it turns out using Pydantic objects to describe structured output for Llms is very, very effective. 00:45.480 --> 00:46.640 And we're going to see that. 00:46.640 --> 00:50.960 And we're going to discuss at the end of this video why this is happening and why this is working so 00:50.960 --> 00:51.760 well together. 00:52.240 --> 00:52.560 All right. 00:52.560 --> 00:54.240 Let me start with the imports. 00:54.280 --> 00:58.960 So let me first import the type annotation of list. 00:59.240 --> 01:01.770 And I also want to import from pedantic. 01:01.810 --> 01:08.970 The base model and field and base model is the core pedantic class that we're going to inherit from 01:09.130 --> 01:12.730 in order to create our own data models and data objects. 01:12.930 --> 01:19.690 And when we create the class that is going to inherit from base model, we get automatic validations, 01:19.690 --> 01:23.930 realization from JSON and ID auto complete. 01:23.930 --> 01:25.290 And this is very useful. 01:25.730 --> 01:33.970 Now field is a function that lets us add extra validation and metadata to our model attributes. 01:34.210 --> 01:41.570 And we can set constraints like minimum maximum maybe descriptions and default values. 01:41.570 --> 01:46.130 And this is very very useful when we want to create those data objects. 01:47.770 --> 01:49.970 All right let's go and see this in action. 01:49.970 --> 01:54.170 So let's create now a new class which will inherit from base model. 01:54.290 --> 01:56.650 And let's call this class source. 01:57.330 --> 02:05.770 Now source is going to have one attribute, which is going to be a URL, and its type is going to be 02:05.770 --> 02:06.330 string. 02:06.730 --> 02:13.210 Now I'm using now the field function to add a description to mention that this is the URL of the source. 02:13.410 --> 02:20.290 Now notice here I also wrote a short description that this is the schema for a source used by the agent. 02:20.450 --> 02:26.410 So you can imagine that we are going to propagate all of this information eventually to an LLM. 02:26.570 --> 02:29.730 So this is what is going to help us get structured output. 02:30.130 --> 02:30.610 All right. 02:30.610 --> 02:36.410 Let's now create our second class which is going to be our agent response class. 02:36.490 --> 02:39.930 So it's also going to inherit from the base model. 02:40.130 --> 02:45.890 And here the description is going to be that this is the schema for the agent response with answer and 02:45.890 --> 02:46.530 sources. 02:46.570 --> 02:47.130 Right. 02:47.170 --> 02:50.610 So we want now the agent to return us two fields. 02:50.610 --> 02:54.410 One is going to be the field answer which is going to be a string. 02:54.490 --> 02:59.060 And here you can see we added the description of this is the agent's answer to the query. 02:59.260 --> 03:05.420 And if we'll take our example from before, where we asked for job postings for AI engineers in Lake 03:05.420 --> 03:06.460 Chain in the Bay area. 03:06.700 --> 03:09.700 So the answer of the agent should be here in this field. 03:09.700 --> 03:14.620 So it's going to be here the list of three job openings that we looked for. 03:15.180 --> 03:19.020 So the second thing is it's going to have an attribute of sources. 03:19.300 --> 03:20.900 Now notice here the type hinting. 03:20.900 --> 03:25.940 This is going to be a list of the source class that we created before. 03:26.300 --> 03:31.420 And here in the description is going to be the list of sources used to generate the answer. 03:32.180 --> 03:39.460 And the default factory list tells pedantic to call the list function to create a new empty list for 03:39.460 --> 03:42.100 each instance where no value is provided. 03:42.100 --> 03:46.500 So this is going to ensure that all of the instances get their own independent lists. 03:48.380 --> 03:48.780 Cool. 03:48.780 --> 03:52.060 So I also wanted to add to this example a nested class. 03:52.060 --> 03:56.700 So notice the agent response holds within it the class of source. 03:56.740 --> 04:01.870 So we have here two pedantic objects, which we compose one on another. 04:02.230 --> 04:10.110 And LM these days can easily handle this kind of composition and to output those kinds of nested objects. 04:10.230 --> 04:13.110 And a fun fact, this wasn't always the case. 04:13.310 --> 04:18.550 So in the past LMS, even top tier LMS struggled with this task. 04:18.590 --> 04:20.070 However, this is not the case today. 04:20.110 --> 04:21.950 We can easily do those kinds of things. 04:22.670 --> 04:25.950 All right so let's go now and format the code. 04:28.910 --> 04:30.190 Let me write here black. 04:31.190 --> 04:33.190 And I sort all right. 04:33.430 --> 04:41.230 So all of this information here on the answers that we expect to do, they need somehow to be plugged 04:41.230 --> 04:44.590 into the prompt that we send the large language model. 04:44.750 --> 04:48.830 So somewhere somebody needs to plug this information in. 04:49.110 --> 04:51.350 So this is what we're going to do now. 04:51.510 --> 04:55.990 So let's go and create a new file and let's call it prompt py. 04:56.230 --> 04:59.310 And here we're going to write our own special prompt. 04:59.710 --> 05:03.470 I'm going to call it the react prompt with format instructions. 05:03.630 --> 05:08.150 And this prompt is going to be the react prompt that we already saw from before. 05:08.350 --> 05:13.310 But we're going to transform it a bit and add here our format instructions. 05:13.310 --> 05:18.390 So the information that we want to give the LLM for us to get a structured response. 05:18.870 --> 05:23.910 So let me go to the prompt in the prompt tab and let me copy it and paste it. 05:27.550 --> 05:27.950 All right. 05:27.950 --> 05:30.510 So this is the exact prompt that we used before. 05:31.030 --> 05:34.670 And now we want to do a bit of prompt engineering. 05:34.790 --> 05:37.870 And here in this final answer clause here. 05:37.870 --> 05:40.310 So I'm going to edit this prompt a bit. 05:40.350 --> 05:47.830 And here I want to tell the LLM to use the format instructions that I'm going to plug it in in the final 05:47.830 --> 05:48.390 answer. 05:48.710 --> 05:54.470 And let me now read this line, the final answer to the original input question formatted according 05:54.510 --> 05:59.240 to the format Instructions and then the placeholder format instructions. 05:59.240 --> 06:04.720 And you can probably guess that we're going to plug in some dynamic value of the objects that we want 06:04.760 --> 06:06.040 our LM to structure. 06:06.680 --> 06:07.280 Alrighty. 06:07.280 --> 06:09.840 So let's go now to the main file. 06:10.080 --> 06:14.200 And we want to use now this prompt in our reasoning chain. 06:16.440 --> 06:19.400 So we want to get rid of this prompt here. 06:19.400 --> 06:23.360 And here we want to put our new and revised prompt that we just wrote here. 06:23.960 --> 06:24.480 Cool. 06:24.480 --> 06:29.600 So let me go now and let me import some classes that we'll be needing here. 06:30.560 --> 06:32.520 So let's go to the top of the file. 06:33.160 --> 06:41.240 And I'm going to paste in those couple of imports here I want to import the pedantic output parser. 06:41.680 --> 06:44.720 And let's have a sneak peek on the implementation here. 06:45.160 --> 06:49.240 And this class is a link chain output parser. 06:49.520 --> 06:53.240 And it's going to take the response that we get from the LM. 06:53.520 --> 06:59.450 And it's going to parse that response into a Pydantic model object. 06:59.570 --> 07:02.890 Now, we're not going to dive into the implementation here. 07:02.970 --> 07:05.090 It's really some ugly parsing here. 07:05.090 --> 07:07.570 And really nobody wants to see this. 07:08.090 --> 07:16.410 And the assumption here is that the response which the LM is going to return is going to be a JSON response, 07:16.490 --> 07:20.890 which is going to be in the Pydantic object format. 07:21.090 --> 07:28.330 So this class shouldn't have any hard time doing it, because we're leaving all of the output and structuring 07:28.330 --> 07:29.730 the data to the LM. 07:29.770 --> 07:32.690 So the LM is going to do all of the heavy lifting. 07:32.810 --> 07:35.770 And this class is simply going to take the response. 07:35.770 --> 07:38.890 And it's going to format it nicely as a Pydantic object. 07:38.930 --> 07:39.610 Alrighty. 07:39.610 --> 07:41.330 So let's go back to the imports. 07:41.330 --> 07:46.090 And we also imported prompt template which we know from previous lessons. 07:46.330 --> 07:50.370 And we imported a class which is called Runnable Lambda. 07:50.570 --> 07:53.450 Now this is a part of the link chain expression language. 07:53.450 --> 07:56.530 And we'll be seeing and learning it very, very soon. 07:56.530 --> 07:58.650 So I don't want to talk about it right now. 07:59.290 --> 08:06.610 And let's go and import the prompt that we wrote and the agent response class from our schema.py file. 08:07.130 --> 08:11.050 Let me just see that the imports are referring to the right code. 08:11.610 --> 08:14.410 So let me go here and let me click here. 08:14.770 --> 08:16.970 And yeah so this seems to be right. 08:16.970 --> 08:17.610 Imports. 08:18.250 --> 08:18.850 All right. 08:18.850 --> 08:22.090 So let me go now and paste this line over here. 08:22.170 --> 08:26.970 And this is going to be the output parser that we're going to create. 08:27.010 --> 08:30.210 It's going to be an object of the pedantic output parser. 08:30.450 --> 08:35.330 And it's going to receive as an input the agent response object. 08:35.450 --> 08:40.970 So we want the output parser to parse objects of the type agent response. 08:42.090 --> 08:46.370 And now I want to create the react prompt with format instructions. 08:46.370 --> 08:50.810 So it's going to be an object of the prompt template. 08:52.090 --> 08:55.900 And the template is going to be what we wrote in the prompt.py file. 08:55.900 --> 08:59.980 So it's going to be the value of the react prompt with format instructions. 09:00.260 --> 09:03.460 So this is now the new template we're going to be using. 09:03.740 --> 09:08.340 So it's going to have the same input variables like before. 09:08.860 --> 09:12.060 So it's going to have a tool names. 09:12.100 --> 09:14.940 It's going to have the input the agent scratchpad. 09:15.180 --> 09:21.140 And by the way notice here I forgot actually to write the tools input variable. 09:21.700 --> 09:24.700 And funny enough this is still going to work. 09:24.700 --> 09:26.500 And I've debugged it for a bit. 09:26.500 --> 09:34.620 And it turns out that in the prompt template logic the login object, they go through the prompt and 09:34.660 --> 09:41.180 they extract everything between those curly brackets and they implicitly add them as input variables. 09:41.340 --> 09:47.820 So this is why the missing tools input variable, which is missing here, is still going to be populated 09:47.820 --> 09:48.420 over here. 09:48.780 --> 09:54.270 And I highly suggest you do those kinds of exercises where if something doesn't add up, simply take 09:54.270 --> 09:56.150 a look at the implementation. 09:56.150 --> 09:59.030 You will have all the answers there anyways. 09:59.030 --> 10:01.510 So up until now we did everything that we know. 10:01.510 --> 10:05.630 We populated the prompt template, but there is something else we need to do here. 10:05.630 --> 10:08.670 So we need to provide those format instructions here. 10:08.670 --> 10:15.350 So we're going to do that with the partial method which is going to allow us to partially format a form 10:15.350 --> 10:21.350 template by pre-filling some of its variables to create a new prompt template that only requires the 10:21.350 --> 10:23.790 remaining variables to be provided later. 10:23.790 --> 10:30.870 And this is useful when we have some variables available early, but others later in the workflow. 10:31.310 --> 10:37.990 So I want to partially populate first the format instructions placeholder, and the value I'm going 10:37.990 --> 10:40.990 to give it is I'm going to give it the output parser. 10:41.030 --> 10:45.230 It has a function or a method which is called get format instructions. 10:45.630 --> 10:49.390 And it's going to get us the format instructions where we want to plug in. 10:49.390 --> 10:52.720 And we're reviewing those format instructions A very very soon. 10:53.280 --> 10:59.120 And anyways, last thing I want to do is to switch the prompt when creating the reasoning chain with 10:59.120 --> 11:01.600 the prompt with format instructions. 11:02.000 --> 11:07.080 And let me run this in debug, because I want to show you the get format instructions here. 11:09.720 --> 11:15.360 So once it's going to get to this breakpoints, let me go and let me copy this expression over here. 11:15.640 --> 11:18.480 Let's paste it here in the debug console. 11:18.680 --> 11:22.000 And here we can see those are the formatting instructions. 11:22.800 --> 11:27.720 The output should be formatted as a JSON instance that conforms to the JSON schema below. 11:28.000 --> 11:30.280 And in example blah blah blah blah. 11:30.560 --> 11:36.440 And at the bottom here we're going to have all the metadata of our Pydantic objects that we wrote. 11:36.640 --> 11:40.600 So eventually this is going to be propagated to the LLM. 11:40.800 --> 11:48.480 And the LLM is going to return us a JSON response that is adhering here to those fields and to those 11:48.480 --> 11:49.080 requirements. 11:49.080 --> 11:50.080 We write it over here. 11:50.200 --> 11:52.400 So this is actually very smart. 11:52.400 --> 11:55.160 And LMS now are very good at doing those things. 11:55.160 --> 12:00.880 So the output parser later is going to have a very easy time taking that JSON and turning that JSON 12:00.880 --> 12:02.640 into a Pydantic object. 12:03.120 --> 12:03.520 All right. 12:03.520 --> 12:08.080 So now let me run everything before we even go and parse it. 12:08.080 --> 12:12.600 And I want to show you the response that we eventually get from the agent here. 12:12.960 --> 12:15.120 So let's go and click run. 12:19.160 --> 12:21.200 And let me fast forward everything here. 12:26.600 --> 12:27.000 All right. 12:27.000 --> 12:28.480 So we got here a response. 12:28.720 --> 12:35.800 Now notice here that in the output field here we can see that we actually have here a JSON which has 12:35.800 --> 12:36.800 the answer key. 12:36.960 --> 12:39.800 And it also has the sources key with the URLs. 12:40.200 --> 12:43.240 So the LLM did what it was supposed to do. 12:43.720 --> 12:44.680 So let me go. 12:44.680 --> 12:50.810 And now put a breakpoint here and run it in debug and let me go and fast forward it again. 12:54.370 --> 12:55.970 So we got the result. 12:55.970 --> 12:57.850 And here is the output key. 12:58.130 --> 13:00.530 And this is here a JSON string. 13:01.050 --> 13:04.330 So let me go and even prove it to you. 13:04.330 --> 13:06.850 And let's go to the debug console. 13:07.010 --> 13:08.690 And let's go and check the types. 13:09.410 --> 13:14.170 So the type of a result is going to be a dictionary. 13:16.090 --> 13:20.970 And the type of result in the key of output is going to be a string. 13:21.090 --> 13:22.610 And this is the JSON string. 13:22.930 --> 13:26.890 And it has all of the properties of the Pydantic object that we want. 13:27.290 --> 13:32.570 So now let me show you how simple it is to create this Pydantic object. 13:32.690 --> 13:35.810 So I'm going to use the output parser of Linkchain. 13:35.970 --> 13:38.810 And it has a method which is called parse. 13:38.970 --> 13:42.050 And I'm going to give it this result here in the output key. 13:42.050 --> 13:47.050 And this is going to parse it into the answer agent object that we want to. 13:54.300 --> 14:02.460 And we can see now the result is of type agent response where we have the answer as the first attribute. 14:02.620 --> 14:09.460 And we also have the sources attribute which is going to be a list of sources. 14:09.500 --> 14:11.380 So this is a very cool. 14:11.580 --> 14:16.380 And now we actually have an object that we can work with programmatically. 14:16.500 --> 14:22.980 So if we want we can downstream it into an application maybe to a front end or expose it via API to 14:23.020 --> 14:23.740 serialize it. 14:23.740 --> 14:26.860 So we have lots of flexibility of what we can do here. 14:26.860 --> 14:32.900 And we can build really robust software leveraging Llms with these kinds of functionality. 14:33.020 --> 14:36.100 So this is a very important part of Linkchain. 14:36.100 --> 14:37.460 And let me show you what I mean. 14:37.460 --> 14:44.980 So let's go and take the parsed object and let's go for example in access the answer here. 14:44.980 --> 14:46.460 So this is the string answer. 14:46.700 --> 14:49.820 Let's go now and check out the sources. 14:50.260 --> 14:54.300 And this is here a list of sources where each source has a URL attribute. 14:54.300 --> 14:56.460 So we can check it for example like this. 14:56.780 --> 15:01.980 This is the URL of one source and we can continue to iterate through the URLs. 15:02.020 --> 15:07.820 Let me just open it and we can see that it opens a valid LinkedIn page. 15:09.180 --> 15:09.700 Hey there. 15:09.740 --> 15:10.980 Ethan here popping out. 15:10.980 --> 15:13.540 And yeah, in case you're wondering, those are new glasses. 15:13.580 --> 15:16.820 Turns out I need glasses in order to review the screen. 15:16.940 --> 15:18.980 You can see much better now, I can tell you that. 15:19.420 --> 15:24.500 And anyways, by now, I hope you have a better understanding of what is output parsing. 15:24.500 --> 15:25.580 Why do we need it? 15:25.660 --> 15:26.740 How does it work? 15:26.780 --> 15:28.260 It all boils down to a prompt. 15:28.260 --> 15:29.260 We send the LM. 15:29.300 --> 15:34.900 The LM structures the output according to our instructions, and then Linkchain simply takes it and 15:34.900 --> 15:38.100 simply formats it nicely to our Pydantic object. 15:38.140 --> 15:38.660 Right? 15:38.700 --> 15:46.380 So in the next video, I want to use the chain expression language to take the output of our agent and 15:46.380 --> 15:47.780 to format it nicely.