WEBVTT 00:00.160 --> 00:02.320 Well, I'm extremely excited about today. 00:02.320 --> 00:09.160 This is of course, day four of week two, and we're getting on to our first large, juicy project in 00:09.160 --> 00:11.760 the form of deep research. 00:11.880 --> 00:18.560 This is one of the classic use cases of Agentic AI, the case where you have an agent that is able to 00:18.600 --> 00:24.800 go off, search the internet and do some research, look at different links and try and research a topic 00:24.800 --> 00:25.720 that you give it. 00:25.720 --> 00:33.280 And we know this so well because many of the Frontier Labs offer this agent via their online chat tools. 00:33.280 --> 00:38.560 And for example, on OpenAI, you can go to GPT and you can press the deep research button. 00:38.560 --> 00:44.320 And that runs a model in deep research mode, which is basically running an agentic workflow. 00:44.520 --> 00:45.880 So we are going to do that. 00:45.920 --> 00:51.480 We are going to give our agents the ability to do deep research, and we're going to use a few things 00:51.480 --> 00:52.440 that we've already learned about. 00:52.440 --> 00:53.360 We're going to use tools. 00:53.360 --> 00:55.760 Of course we're going to use structured outputs again. 00:55.800 --> 00:58.880 We encountered it briefly last time a little bit more this time. 00:59.080 --> 01:04.340 And we're going to use something called hosted tools for the first time, using tools that are running 01:04.340 --> 01:05.220 remotely. 01:05.500 --> 01:07.620 All right, let's get straight to it. 01:08.100 --> 01:09.660 Okay, so here we are in Quito. 01:09.700 --> 01:12.820 We're going to open up the directory for week two. 01:12.860 --> 01:17.300 And we're now going to lab for deep research. 01:17.580 --> 01:24.420 And we are going to start again in a notebook like this in one of the notebooks running within cursor. 01:24.660 --> 01:28.820 And I know that a lot of you would prefer to be working with real Python modules. 01:28.820 --> 01:29.860 And fear not. 01:29.900 --> 01:31.220 We will get to that tomorrow. 01:31.260 --> 01:35.300 We will do this in Python code and proper modules. 01:35.540 --> 01:41.420 The lab, this kind of format, these notebooks, they're so good for talking through things for experimenting. 01:41.420 --> 01:46.060 And it gives you the power to be changing things, playing around while we go through this. 01:46.060 --> 01:47.860 So I think it's good just for today. 01:48.340 --> 01:53.340 So this is of course a enormous classic agentic use case. 01:53.340 --> 01:59.160 And what's great about it is it applies so broadly you can apply this to any business area. 01:59.200 --> 02:05.440 You can build a deep research agent that's focused on your business area, or you can use this as something 02:05.440 --> 02:07.400 which is like a sidekick for yourself. 02:07.440 --> 02:09.600 We're going to build a different kind of sidekick later in this course. 02:09.600 --> 02:13.800 But but this is one kind already and it should be useful for you immediately. 02:13.920 --> 02:19.320 So I think it's terrific that we're doing this, and I hope you're as excited as me to be building your 02:19.320 --> 02:23.720 own deep research agent, something which the Frontier Labs have done themselves. 02:23.960 --> 02:26.040 Okay, so let's get started. 02:26.080 --> 02:27.600 Enough of my prattling away. 02:27.600 --> 02:29.040 We do some imports. 02:29.040 --> 02:31.800 We do one of these usual load dot envs. 02:32.160 --> 02:32.800 All right. 02:32.800 --> 02:38.280 And now it's time for me to talk about the big the big change this week, which is that we are going 02:38.280 --> 02:44.520 to use some things called hosted tools, which means that they are tools that OpenAI is going to run 02:44.520 --> 02:45.360 for us. 02:45.400 --> 02:49.640 There only are three of them that OpenAI provides, at least as of now, and we're going to use one 02:49.640 --> 02:50.120 of them. 02:50.120 --> 02:54.200 The three are the web search tool, the file search tool, and the computer tool. 02:54.440 --> 03:01.220 The web search tool All lets us do what we're going to do today, which is ask OpenAI to run a search 03:01.260 --> 03:02.220 on our behalf. 03:02.740 --> 03:09.380 The file search tool is how we can run searches across the vector stores that you can run. 03:09.420 --> 03:16.020 You can upload and have remotely with OpenAI and computer tool allows things like taking screenshots 03:16.020 --> 03:17.660 and clicking around a computer. 03:17.700 --> 03:19.180 Now we're not going to be using either of those two. 03:19.180 --> 03:20.580 We're going to be doing the web search tool. 03:20.580 --> 03:25.260 But fear not, later in the course we will have basically our own computer tool, but we won't be running 03:25.260 --> 03:25.860 it at hosted. 03:25.860 --> 03:28.900 We'll be running it on our own computers, as you will see. 03:28.940 --> 03:32.780 In case that sounds intriguing, but for now we're going to be using the web search tool. 03:32.780 --> 03:38.780 And I want to make an important point, which is that OpenAI's web search tool is not that cheap for 03:38.780 --> 03:41.580 some reason, which seems not entirely clear to me. 03:41.580 --> 03:50.540 It actually costs 2.5 cents right now for me to run this for, even at the low end of the cheapest model 03:50.540 --> 03:56.040 at the lowest end, 2.5 cents per call, which may not sound like it's a huge amount, But typically 03:56.040 --> 03:59.440 if we're running a deep research, we might be doing like ten different searches. 03:59.440 --> 04:05.520 And that means it's $0.25 a pop, which means that you can quite quickly add up to 2 to $3, which is 04:05.520 --> 04:07.160 what I spent the first time I did this. 04:07.160 --> 04:09.560 I think the way I have it set up now would be more like $1. 04:09.760 --> 04:11.920 Still, that's maybe not to be sniffed at. 04:11.920 --> 04:17.520 So I would suggest monitor this and you can decide to skip doing the searches, or I'll show you where 04:17.520 --> 04:20.840 you can turn it all the way down so that you're only spending a few cents. 04:21.000 --> 04:24.400 And then later in future weeks, we're going to be doing this a bit differently. 04:24.400 --> 04:27.560 And we'll have other ways to do this which will cost a fraction of this. 04:27.560 --> 04:27.880 So. 04:27.880 --> 04:28.640 So fear not. 04:28.640 --> 04:30.800 But for now, it's nice for us to use OpenAI's. 04:30.800 --> 04:32.560 And OpenAI's is actually really good. 04:32.600 --> 04:32.960 So. 04:32.960 --> 04:35.280 So maybe maybe it's worth a 2.5 cents. 04:35.280 --> 04:36.000 I don't know. 04:36.560 --> 04:38.920 The costs themselves are on this link here. 04:38.920 --> 04:39.320 So. 04:39.320 --> 04:43.960 So it's always worth double checking because it might be different for your region, or OpenAI might 04:43.960 --> 04:45.200 change the pricing all the time. 04:45.240 --> 04:51.960 So go there and you can look we are going to be using GPT four mini low search okay. 04:52.200 --> 04:55.580 With that, let's get started with our first search. 04:56.180 --> 04:56.620 Okay. 04:56.660 --> 05:00.220 It's now time for us to create our first agent for this product. 05:00.220 --> 05:04.460 And the first agent is called the search agent. 05:04.700 --> 05:06.300 So here are the instructions. 05:06.300 --> 05:07.940 The system prompt for this agent. 05:08.220 --> 05:09.780 You are a research assistant. 05:09.820 --> 05:13.060 Given a search term, you search the web for that term. 05:13.060 --> 05:20.060 You produce a concise summary 2 to 3 paragraphs, capture the main points, write succinctly, and no 05:20.060 --> 05:23.260 need to have good grammar because it will be consumed by someone else. 05:23.340 --> 05:24.980 Synthesizing a report. 05:25.340 --> 05:30.740 It's a very clearly written set of instructions right here, and I would love to take credit for it. 05:30.740 --> 05:37.140 But I must confess that I just took this verbatim from OpenAI's documentation for how to do web searches 05:37.140 --> 05:42.940 with that tool, so we can probably take it on good authority that this is a well written prompt, since 05:42.940 --> 05:45.060 it's written by the people that built the model. 05:45.300 --> 05:47.140 So, um, we're going to use this prompt. 05:47.140 --> 05:49.260 We're going to create a search agent. 05:49.260 --> 05:51.460 It is an agent named search agent. 05:51.700 --> 05:53.680 This is the the instructions. 05:53.840 --> 05:55.600 And we're going to give it one tool. 05:55.600 --> 05:58.080 And that tool is called a web search tool. 05:58.080 --> 06:02.160 And this is one of OpenAI's three hosted tools. 06:02.160 --> 06:04.480 So we're we're providing that as our tool. 06:04.760 --> 06:06.520 There is an optional parameter. 06:06.520 --> 06:12.680 You can specify the search context size which can be low, medium or high where that reflects the price 06:12.680 --> 06:13.440 that you'll pay. 06:13.480 --> 06:16.280 Low is of course the cheapest and then medium is. 06:16.280 --> 06:19.160 The default is only a little bit more, but high I think is quite a lot more. 06:19.440 --> 06:21.040 You can also pick the model. 06:21.040 --> 06:24.160 It's a lot cheaper to pick GPT four or mini GPT four. 06:24.200 --> 06:25.200 Zero is a lot more. 06:25.360 --> 06:28.440 And um, finally there's there's this here. 06:28.440 --> 06:32.240 This is a nice little snippet of code to to to keep in mind. 06:32.280 --> 06:37.080 This is how you say that this agent is required to run the tool. 06:37.080 --> 06:38.280 It is mandatory. 06:38.320 --> 06:40.280 It doesn't have discretion over this. 06:40.280 --> 06:41.560 So this is this is the way you do it. 06:41.560 --> 06:45.120 You pass in this model settings with tool choices required. 06:45.120 --> 06:47.120 And that's a nice handy thing to have to one side. 06:47.120 --> 06:48.960 And of course in this case we want to do that. 06:48.960 --> 06:49.760 We want to. 06:49.800 --> 06:53.540 It's basically this agent is wrapping this tool call. 06:53.900 --> 06:59.580 And so when we're picking GPT for our mini and we're picking low, this is when as of now this costs 06:59.580 --> 07:03.580 two and a half cents every time we run this agent. 07:04.100 --> 07:05.620 So let's run this agent. 07:05.620 --> 07:10.940 Let's ask for the latest AI agent frameworks in 2025. 07:10.980 --> 07:17.060 We will run that search and we will display the results and see what happens. 07:17.540 --> 07:19.940 It's running and there we go. 07:19.980 --> 07:20.940 There are the results. 07:20.940 --> 07:23.060 It's come in nice markdown. 07:23.100 --> 07:30.060 Let's see it says Lang lang graph crew semantic kernel autogen OpenAI gym. 07:30.060 --> 07:34.940 That's uh, not exactly I guess that's a that's a reinforcement learning framework. 07:35.180 --> 07:36.260 It doesn't know about itself. 07:36.260 --> 07:38.420 It doesn't know about OpenAI agents SDK. 07:38.460 --> 07:45.820 Amusingly, uh, Rasa is an older framework and Jade for Java people and, uh, yeah. 07:45.820 --> 07:50.360 So, so, uh, an interesting set of results, I'd say mixed accuracy because I would. 07:50.360 --> 07:56.960 I would argue that long chain and semantic kernel aren't really agent frameworks, but certainly Landgraf, 07:57.000 --> 08:03.520 crew, and Autogen are very much there, and ones that we will be spending a week on each. 08:03.880 --> 08:04.400 All right. 08:04.400 --> 08:11.000 But this is a perfectly pleasant search result, and it is the beginning of our deep research agent, 08:11.560 --> 08:15.520 and it's always worth taking a quick look at the trace in OpenAI. 08:15.920 --> 08:18.320 So click on this little useful link here. 08:18.640 --> 08:20.240 And up it comes. 08:20.440 --> 08:21.200 And here it is. 08:21.200 --> 08:24.840 The search agent is our current trace. 08:24.840 --> 08:26.440 If I open it up we'll see. 08:26.480 --> 08:29.120 Sure enough that the search agent was called. 08:29.120 --> 08:33.320 And it tells us over here the tool it used was web search. 08:33.520 --> 08:37.400 And we get a bit of information about the total number of tokens that it did. 08:37.400 --> 08:39.560 That's the instructions, the system prompt. 08:39.560 --> 08:40.960 That was the user prompt. 08:40.960 --> 08:43.120 And this comes back is the output. 08:43.120 --> 08:46.960 Just as we would expect the trace has worked nicely. 08:47.120 --> 08:49.240 Now we can get on with the show.