WEBVTT 00:00.160 --> 00:00.600 Okay. 00:00.640 --> 00:03.320 So now we're going to work on another agent. 00:03.320 --> 00:07.840 And we're going to make heavier use of structured outputs than we did last time. 00:08.080 --> 00:12.080 So the agent in particular that we're going to work on, we're going to call the planner agent. 00:12.320 --> 00:19.640 This is the one that's going to be responsible for taking a query and coming up with a handful of searches 00:19.640 --> 00:24.440 that it should run based on that query in order to do some deep research. 00:24.720 --> 00:31.320 Now, I'm going to keep the number of searches small to three searches because as I say, it's two and 00:31.320 --> 00:32.800 a half cents for each search. 00:32.840 --> 00:35.200 And I don't want to add up to some big old bill. 00:35.200 --> 00:37.400 When I first did this, I had that at ten. 00:37.400 --> 00:39.560 So you certainly can have it at bigger numbers. 00:39.560 --> 00:44.280 You obviously get better results if you do that, and it's very much a personal preference thing. 00:44.320 --> 00:46.000 Run it with three to start with. 00:46.040 --> 00:48.080 See you know your cost $0.10, $0.20. 00:48.080 --> 00:54.040 And if you're enjoying yourself then you know splash out, spend half a dollar and get one really good 00:54.040 --> 00:55.760 comprehensive result. 00:56.440 --> 00:57.040 Okay. 00:57.360 --> 00:59.400 So we're going to keep it on three. 00:59.560 --> 01:03.110 So the instructions what is the system prompt we're going to use here? 01:03.350 --> 01:05.670 You are a helpful research assistant. 01:05.670 --> 01:13.310 Given a query, come up with a set of web searches to perform to best answer that query output. 01:13.510 --> 01:17.070 And then the number three terms to query for. 01:18.030 --> 01:20.150 So that hopefully is super simple. 01:20.150 --> 01:26.470 We're just building an agent that is going to, given some sort of a query, come up with a bunch of 01:26.470 --> 01:31.190 things you would search the internet for in order to research that query, and we're calling that the 01:31.190 --> 01:31.870 planner. 01:32.110 --> 01:36.150 So keep that in mind because then everything is going to click into place. 01:36.670 --> 01:38.150 So now we use structured outputs. 01:38.190 --> 01:38.870 Remember this. 01:38.870 --> 01:45.310 This is where you have a class that is a subclass of this thing called base model from pedantic. 01:45.470 --> 01:47.870 And that means that we're going to use it as like a schema. 01:47.910 --> 01:54.470 We're going to use it to describe a structure, which we will later ask a model to return information 01:54.470 --> 01:54.750 in. 01:55.350 --> 02:01.930 And we in the case of a web search item, we're going to have a reason and a query. 02:02.850 --> 02:09.010 And the reason is a string, and the query is a string, and the doc string that we have right here 02:09.010 --> 02:12.730 underneath this is this is super important. 02:12.850 --> 02:20.850 And by by adding this documentation, by annotating these fields with this comment right underneath 02:20.850 --> 02:26.850 them, we ensure that that that information is actually provided as part of structured outputs. 02:26.850 --> 02:29.530 That is part of the information that's provided to the model. 02:29.530 --> 02:32.930 So it knows why it's populating these fields. 02:33.050 --> 02:34.530 So this is another good trick. 02:34.530 --> 02:35.530 We didn't do that yesterday. 02:35.530 --> 02:39.570 So so now you've you've got this in your toolkit as well. 02:40.330 --> 02:42.250 Now one other small point. 02:42.970 --> 02:47.690 There's various reasons why you might ask for a reason as well as just the query. 02:47.690 --> 02:54.930 But one of them is that this does actually result in sort of reasoning style behavior. 02:55.170 --> 02:58.170 This is very similar to to what people call chain of thought prompting. 02:58.170 --> 03:05.280 I guess it is chain of thought teleprompter, but by asking the model to output its rationale for why 03:05.280 --> 03:11.640 that search is important, and actually by asking for that first before the query itself, you encourage 03:11.640 --> 03:18.400 the model to be sort of apparently thinking through, and as a result, you tend to get better outputs, 03:18.400 --> 03:23.880 you get better queries, just as a side effect of asking for a rationale first. 03:24.320 --> 03:30.080 And I've said it before, but you always have to come back to remembering that while we try and treat 03:30.080 --> 03:36.760 models like they're like there's a bit of humanity to them, this is just a strange side effect of great 03:36.800 --> 03:41.320 next token prediction, because it's predicting the most likely next tokens. 03:41.320 --> 03:43.000 That's all these models are doing. 03:43.600 --> 03:48.840 If you ask it to start predicting tokens reflecting a reason, then when it starts to predict tokens 03:48.840 --> 03:55.600 reflecting the query, it is more likely to be consistent with that reason, and it ends up with this 03:55.600 --> 03:59.190 very coherent output that just makes more sense. 03:59.190 --> 04:02.590 That has a higher likelihood of being what you're after. 04:02.870 --> 04:05.270 So another those little sidebars. 04:05.310 --> 04:08.830 And I do have a bunch of YouTube videos that clarify that more. 04:08.830 --> 04:13.790 But it's so important to keep that in your mind because it really helps figure out what works well with 04:13.790 --> 04:15.030 these kinds of prompts. 04:15.230 --> 04:15.910 All right. 04:16.110 --> 04:20.710 So this is a pedantic object with a reason and a query. 04:21.150 --> 04:26.070 And now we've got another one called Web Search plan. 04:26.110 --> 04:28.110 So this was web search item. 04:28.390 --> 04:34.350 A web search plan is a list of web search items in a field searches. 04:34.670 --> 04:38.070 So this is an object which has one field searches. 04:38.070 --> 04:40.470 And that is a list of these guys. 04:40.470 --> 04:47.350 And you can see it says they're a list of web searches to perform to best answer the query, a list 04:47.350 --> 04:50.110 of web searches to perform to best answer the query. 04:50.230 --> 04:51.630 And that is what goes there. 04:51.630 --> 04:55.390 And this annotation is associated with this field searches. 04:56.110 --> 04:56.710 Okay. 04:56.750 --> 05:03.050 Hopefully you are with me So then we now create our planner agent. 05:03.170 --> 05:05.450 It is an agent that's named planner agent. 05:05.570 --> 05:07.810 Its instructions are these ones right here. 05:08.010 --> 05:11.570 The model is GPT four mini and the output type. 05:11.610 --> 05:14.890 Remember this is where we specify structured outputs. 05:14.930 --> 05:20.250 We tell the planner agent that it shouldn't output in text like these things normally do. 05:20.250 --> 05:26.490 It shouldn't just create generate natural language in the form of text, but rather it should build 05:26.490 --> 05:33.450 an object of type web search plan, which means it's an object with one field searches, which has a 05:33.490 --> 05:36.090 list of items that looks like this. 05:36.850 --> 05:41.290 And remember, as another sidebar, keep in the back of your mind when I say that we're going to tell 05:41.290 --> 05:43.290 the model to create an object. 05:43.330 --> 05:44.930 Maybe that all sounds a bit magical. 05:44.970 --> 05:50.730 Like how does a model, how does a large language model, a transformer architecture, know to generate 05:50.890 --> 05:51.450 objects? 05:51.450 --> 05:52.690 Python objects. 05:52.730 --> 05:58.520 Remember, any time you hear magical stuff like that, it usually means we're talking about JSON behind 05:58.520 --> 05:59.080 the scenes. 05:59.080 --> 06:00.800 And that's exactly what this is. 06:00.800 --> 06:06.160 What happens is that this is turned into a bunch of JSON, and the model is prompted and it says your 06:06.160 --> 06:09.120 response should conform to this JSON. 06:09.160 --> 06:10.640 Here is how it needs to look. 06:10.680 --> 06:17.120 And as a result, the model generates that JSON and it gets populated into an object of this type. 06:17.360 --> 06:18.480 That's that's the end of it. 06:18.480 --> 06:20.200 There's nothing magical about it. 06:20.920 --> 06:21.640 All right. 06:21.880 --> 06:24.000 We should now run this and see it working. 06:24.560 --> 06:25.080 Here we go. 06:25.120 --> 06:25.960 Let's do it. 06:26.000 --> 06:29.880 We run this, and now we're going to say the same. 06:29.880 --> 06:34.960 The same thing that before we just did an internet search on latest AI frameworks in 2025, what we're 06:34.960 --> 06:36.240 now going to do is carry out. 06:36.240 --> 06:37.440 We're not going to do any searching at all. 06:37.440 --> 06:43.880 We're just going to use this planner agent to come up with three web searches that would be relevant, 06:43.880 --> 06:48.080 and we would want it to respond with an object of type web search plan. 06:48.120 --> 06:51.200 So let's see what we get if we run it and print the result. 06:52.440 --> 06:58.260 So what we get back is an object which has a field searches Like that. 06:58.580 --> 07:00.620 And that field is a list. 07:00.660 --> 07:01.260 Good. 07:01.500 --> 07:04.660 And it's a list of things called web search items. 07:04.860 --> 07:05.580 Great. 07:05.740 --> 07:11.940 And a web search item should have a reason and a query reason to find the most recent developments and 07:11.940 --> 07:14.820 releases in AI agent frameworks for 2025. 07:14.940 --> 07:18.300 And the query is just latest AI agent frameworks 2025. 07:18.460 --> 07:19.660 Okay, fine. 07:19.940 --> 07:23.180 That doesn't seem particularly profound, but it seems perfectly reasonable. 07:23.580 --> 07:27.620 The next one to explore industry specific or academic research. 07:27.780 --> 07:31.100 And the query is emerging AI agent frameworks. 07:31.220 --> 07:36.220 And the next one is to be updated with major conferences, publications or announcements. 07:37.020 --> 07:38.060 AI frameworks. 07:38.060 --> 07:38.620 Conferences. 07:38.620 --> 07:40.220 Announcements 2025. 07:40.260 --> 07:41.020 Fair enough. 07:41.380 --> 07:44.420 Seems like a perfectly plausible, reasonable way of doing it. 07:44.460 --> 07:46.260 I'm not sure I would have done that last one, but. 07:46.260 --> 07:52.020 But who's to say whether I would have done it better or GPT four or mini is going to do it better. 07:52.020 --> 07:57.660 But that is the planner at work and you're seeing structured outputs used very well there.