WEBVTT 00:00.450 --> 00:03.930 -: Okay, I'm gonna teach you how to use Instructor. 00:03.930 --> 00:05.700 It's a really lightweight package. 00:05.700 --> 00:09.300 If you've tried LangChain and really hated it, 00:09.300 --> 00:11.220 fallen in and out of love with it 00:11.220 --> 00:13.710 'cause it was too complicated, and too much overhead, 00:13.710 --> 00:17.070 Instructor is really, it's something you can't argue with 00:17.070 --> 00:19.590 because adds a minimum amount of overhead. 00:19.590 --> 00:23.190 It's basically the same as using the OpenAI API, 00:23.190 --> 00:25.170 except it gives you a lot of the benefits 00:25.170 --> 00:26.400 that you get with LangChain. 00:26.400 --> 00:28.710 You can drop in different LLMs, 00:28.710 --> 00:32.160 you can get automatic retries, and do a lot of cool stuff. 00:32.160 --> 00:36.270 First you wanna open install Instructor, really simple. 00:36.270 --> 00:37.680 And then this is how you use it. 00:37.680 --> 00:40.470 So you can see that this is exactly the same 00:40.470 --> 00:42.450 as the OpenAI API. 00:42.450 --> 00:45.210 And one thing that is a little bit different 00:45.210 --> 00:47.970 with the Instructor is that you pass it a model, 00:47.970 --> 00:49.410 a Pydantic model. 00:49.410 --> 00:52.920 So if you know anything about object-oriented programming 00:52.920 --> 00:56.010 or types or whatever, you probably familiar with this. 00:56.010 --> 00:57.930 But even if you aren't, it's pretty straightforward. 00:57.930 --> 01:01.530 You basically just make a structure, like a class, 01:01.530 --> 01:05.310 that tells the model what type of response you want. 01:05.310 --> 01:09.210 So inherits from the base model, which is from Pydantic. 01:09.210 --> 01:11.640 And then we've called it user info. 01:11.640 --> 01:14.970 And we just say that user info has a name and an age. 01:14.970 --> 01:16.980 And it's a string and an integer. 01:16.980 --> 01:19.170 And so that's all the information we have to give it. 01:19.170 --> 01:22.110 And we create a client with OpenAI 01:22.110 --> 01:23.910 and then we pass in that response model. 01:23.910 --> 01:27.060 What Instructor does under the hood is it feeds that 01:27.060 --> 01:28.740 into the system instructions. 01:28.740 --> 01:33.740 And that way we get back what the actual Pydantic model 01:33.840 --> 01:34.800 wants you to give back. 01:34.800 --> 01:37.980 If I run this, then we're gonna get back 01:37.980 --> 01:39.570 this user info model. 01:39.570 --> 01:42.240 You can see I just said user_info.name. 01:42.240 --> 01:43.447 It is John Doe. 01:43.447 --> 01:46.440 User_info.age is 30, and it's pulled that 01:46.440 --> 01:48.360 from the messages here. 01:48.360 --> 01:52.353 And this is great for image for, sorry, retake. 01:53.280 --> 01:56.610 This is great for if you want to extract information 01:56.610 --> 01:58.200 from large amounts of text. 01:58.200 --> 02:00.210 This is just a short thing here, 02:00.210 --> 02:02.520 but we could get quite a lot of information. 02:02.520 --> 02:03.540 And it's also great 02:03.540 --> 02:05.970 because you don't have to remember, okay, 02:05.970 --> 02:08.280 hold on, what was the thing again? 02:08.280 --> 02:11.400 How do we get the responses back 02:11.400 --> 02:13.020 and like how do we pass the response? 02:13.020 --> 02:15.120 And if it's JSON, what happens if it comes back 02:15.120 --> 02:16.200 and it's wrong? 02:16.200 --> 02:17.130 This is really good. 02:17.130 --> 02:19.800 And you can do automatic retries based on if it comes back 02:19.800 --> 02:20.850 the wrong thing. 02:20.850 --> 02:24.210 You can also drop in other LLMs. 02:24.210 --> 02:25.860 They're basically only line you have to change 02:25.860 --> 02:30.300 is like from anthropic if you want to drop in Claude. 02:30.300 --> 02:32.490 And they have all of that in those tutorials. 02:32.490 --> 02:34.350 And there's one other really cool thing 02:34.350 --> 02:37.140 I wanted to show you, which is streaming. 02:37.140 --> 02:41.820 So it can actually bring back structured data in midstream. 02:41.820 --> 02:44.700 So that doesn't make any sense to you probably, 02:44.700 --> 02:46.050 because you've never seen it before. 02:46.050 --> 02:48.300 But let me show you a demo, it's pretty cool. 02:48.300 --> 02:51.150 So here's a big textbook about some meeting. 02:51.150 --> 02:52.650 And we have some information there. 02:52.650 --> 02:55.830 Who are the people, what emails do we have, 02:55.830 --> 02:56.850 and things like that. 02:56.850 --> 02:58.200 And we set up a few different things. 02:58.200 --> 03:02.070 So we set up a user model instead of a meeting info model, 03:02.070 --> 03:05.190 which has a date, location, budget, deadline, et cetera. 03:05.190 --> 03:08.130 And then we have, we just create our extraction stream. 03:08.130 --> 03:10.320 And the only difference is we just say 03:10.320 --> 03:14.640 stream equals true, and we're giving it a partial. 03:14.640 --> 03:16.593 So it's saying we're getting back, 03:18.000 --> 03:21.990 so here's this partial, we're getting back a stream. 03:21.990 --> 03:24.300 And the stream is gonna be meeting info. 03:24.300 --> 03:26.940 And the meeting info has a list of users 03:26.940 --> 03:28.260 as well as the date location. 03:28.260 --> 03:30.750 And then each user will have a name, email, Twitter. 03:30.750 --> 03:31.860 Okay, we're gonna run this 03:31.860 --> 03:33.990 and we're gonna see this update in real time. 03:33.990 --> 03:36.570 So it's found John Smith and his email. 03:36.570 --> 03:38.370 It's found Jane Doe and her email, 03:38.370 --> 03:40.830 it's found Alice Johnson and her email. 03:40.830 --> 03:42.660 And it's starting to fill in these other things as well. 03:42.660 --> 03:43.493 And now it's done. 03:43.493 --> 03:45.690 Now you can see that this is really powerful 03:45.690 --> 03:49.020 because if you have a prompt that's pulling back 03:49.020 --> 03:51.420 and streaming back a bunch of structured data, 03:51.420 --> 03:53.580 you can actually act on that structured data 03:53.580 --> 03:54.990 before it's even ready. 03:54.990 --> 03:57.330 Which I think is really helpful. 03:57.330 --> 04:00.540 All right, one other thing is retry logic. 04:00.540 --> 04:03.570 And so, this is an example from the documentation. 04:03.570 --> 04:06.579 If you add like a field validator here, 04:06.579 --> 04:09.990 we're gonna do this just specifically to cause and error. 04:09.990 --> 04:12.270 And this is the field validator of name, 04:12.270 --> 04:15.390 and we're saying that name needs to be uppercase. 04:15.390 --> 04:17.400 So we have this here and it's gonna pass all that 04:17.400 --> 04:18.960 into the prompt as well. 04:18.960 --> 04:20.250 But we're gonna validate that. 04:20.250 --> 04:23.190 And we're using Tenacity here for retrying. 04:23.190 --> 04:27.210 Instructor has a max retries equals thing here. 04:27.210 --> 04:28.800 So we can just say, okay, 04:28.800 --> 04:30.900 we wanna retry three times, or whatever. 04:30.900 --> 04:33.540 And then we're just adding Tenacity into here 04:33.540 --> 04:34.830 to manage that for us. 04:34.830 --> 04:37.950 So, it's gonna print something before, 04:37.950 --> 04:39.720 it's gonna print something afterwards. 04:39.720 --> 04:42.900 And then it's gonna stop after three attempts. 04:42.900 --> 04:45.540 And I'm just gonna run this, and it's gonna extract 04:45.540 --> 04:48.690 the name John, but it's not gonna be all in uppercase. 04:48.690 --> 04:52.500 And we can assert here that it's not actually John, right? 04:52.500 --> 04:56.220 And you can see it tries one time, it tries another, 04:56.220 --> 04:59.790 it tries another time saying it's not passing the assertion. 04:59.790 --> 05:02.400 So that is really interesting. 05:02.400 --> 05:04.590 And this is a really helpful pattern. 05:04.590 --> 05:07.080 But Instructor makes that a little bit easier. 05:07.080 --> 05:09.510 All right, so yeah, hopefully that's useful for you. 05:09.510 --> 05:10.740 I really like this library. 05:10.740 --> 05:11.970 It's quite lightweight. 05:11.970 --> 05:13.530 We can get a lot of power out of it 05:13.530 --> 05:15.540 and a lot of the benefit that you would have 05:15.540 --> 05:16.950 from using LangChain, 05:16.950 --> 05:19.353 but without having to retool everything.