WEBVTT 00:00.440 --> 00:03.680 Okay, next one is a boring one because you know it. 00:03.720 --> 00:09.480 Well, I've brought in the same send email tool that we used before. 00:09.680 --> 00:13.920 So this is we're using the decorator the function tool. 00:13.920 --> 00:18.680 If you remember you use that decorator if you want to convert a function into something that can be 00:18.680 --> 00:19.080 a tool. 00:19.080 --> 00:23.040 And it will generate all of that JSON gumph for you behind the scenes. 00:23.320 --> 00:29.320 So this is send email the subject and a body, and it sends out an email with the given subject in the 00:29.320 --> 00:29.560 body. 00:29.560 --> 00:31.680 We're going to use Send Grid as before. 00:31.720 --> 00:35.960 Hopefully you still have your account handy and you have it in your env file. 00:36.000 --> 00:38.400 Remember to change this email. 00:38.440 --> 00:44.280 Let me change this to your verified email. 00:44.880 --> 00:49.040 And yes, and please do also change this one. 00:49.040 --> 00:54.840 Again, don't send me lots of emails with AI generated content, but by all means send me emails with 00:54.840 --> 00:55.680 your questions. 00:55.720 --> 01:02.600 Of course I always welcome them, but I don't welcome AI generated search deep research things as much. 01:03.070 --> 01:11.150 Uh, so then uh, content would be, uh, the HTML version, and then we send the email exactly as before. 01:11.350 --> 01:12.550 That is our tool. 01:12.750 --> 01:15.630 You don't need to see this, but I'm going to show it to you anyway. 01:15.630 --> 01:19.030 You remember that if I just print this. 01:19.070 --> 01:20.110 What's it doing here? 01:20.150 --> 01:26.590 If I just print this out, we should see that we get a function tool with a description and so on. 01:26.590 --> 01:30.750 And if you're not sure why this has happened, then do look up decorators. 01:30.750 --> 01:35.870 Make sure you understand how decorators work and how they're able to convert a function like this. 01:36.630 --> 01:37.310 Okay. 01:37.350 --> 01:41.390 With that in mind, this is an email agent. 01:41.390 --> 01:42.910 This is all old hat to you. 01:42.910 --> 01:44.590 Now we've done this several times. 01:44.790 --> 01:50.870 The instructions are you're able to send a nicely formatted HTML email based on a detailed report. 01:51.350 --> 01:56.270 You'll be provided with with a with a detailed report, you should use your tool to send one email, 01:56.310 --> 02:03.350 providing the report converted into a clean, well presented HTML with a subject line. 02:03.670 --> 02:06.190 So you can see I'm no longer trying to break things out. 02:06.300 --> 02:11.820 I'm giving this agent full discretion to come up with a subject line, to rewrite the email, to do 02:11.820 --> 02:18.620 the whole lot using GPT four or mini, and we're providing it with the one tool to actually send the 02:18.620 --> 02:19.260 email. 02:19.260 --> 02:21.180 So that is our email agent. 02:21.180 --> 02:23.580 And that is something which is old news to you. 02:23.580 --> 02:24.500 You know, it's super. 02:24.500 --> 02:26.220 Well, okay. 02:26.620 --> 02:28.940 We're about we're about to get real. 02:29.140 --> 02:36.100 And so I will, uh, take a pause for a second and then we will get into the real meat of the project. 02:36.740 --> 02:37.460 Okay. 02:37.740 --> 02:43.060 So we have another agent again that's going to use structured outputs. 02:43.380 --> 02:45.740 So this is the researcher. 02:46.300 --> 02:51.260 So you are a senior researcher tasked with writing a cohesive report for a query. 02:51.420 --> 02:56.940 You'll be provided with the original query and some initial research done by a research assistant. 02:56.980 --> 03:01.860 You come up with an outline for the report that describes the structure and flow of the report, and 03:01.860 --> 03:08.260 then generate the report and return that as your final output should be in markdown, it should be lengthy 03:08.260 --> 03:13.050 The detailed 5 to 10 pages of content of at least a thousand words. 03:13.370 --> 03:16.530 Okay, so that has framed the situation. 03:16.970 --> 03:18.770 Now we're going to use structured outputs. 03:18.810 --> 03:24.530 Again we're going to have a report data pedantic object a subclass of base model. 03:24.650 --> 03:30.010 It has a short summary which we begin with a short 2 to 3 sentence summary. 03:30.210 --> 03:35.610 The full markdown final report and then some follow up suggestions. 03:35.770 --> 03:38.570 Suggested topics to research further. 03:39.370 --> 03:41.410 And so this is an agent. 03:41.570 --> 03:43.010 It will take these instructions. 03:43.010 --> 03:46.010 It will use GPT four or mini the output type. 03:46.010 --> 03:49.330 This is where you specify that we're using structured outputs. 03:49.370 --> 03:52.410 The output type is the report data. 03:52.650 --> 03:56.290 And there we have our our report writer. 03:56.610 --> 03:58.370 And this is our writer agent. 03:58.370 --> 04:05.010 And this is the final building block before we get to the actual flow through this project. 04:05.890 --> 04:06.490 Okay. 04:06.530 --> 04:07.530 So it's crunch time. 04:07.530 --> 04:10.810 We have five functions for me to show you. 04:10.810 --> 04:12.440 But they're relatively simple. 04:12.440 --> 04:14.080 They're pretty straightforward. 04:14.080 --> 04:18.120 So here they are the I'm going to show you three and then two. 04:18.560 --> 04:25.840 The first three, the ones that are going to execute a search using the two agents that we defined before 04:25.880 --> 04:28.320 the planner agent and the search agent. 04:28.320 --> 04:30.560 And these functions, you can see they're very short. 04:30.560 --> 04:31.480 They're very simple. 04:31.480 --> 04:34.920 Let me tell you what they each do plan searches. 04:34.960 --> 04:35.400 All right. 04:35.400 --> 04:41.520 This very simply calls runner run for our planner agent passing in the query. 04:41.960 --> 04:45.200 And we make a query and that gets passed into the planet agent. 04:45.200 --> 04:49.880 Remember the planner agent is the one that will then come back with a number of searches. 04:50.280 --> 04:52.960 So we'll then count how many searches there are and print that. 04:52.960 --> 04:59.600 And we will return that final output, which is in the form of that search list. 04:59.640 --> 05:00.760 That's what formatted it. 05:00.760 --> 05:06.640 Again, it is this thing here, the web search plan, a list of web search items. 05:07.080 --> 05:07.760 All right. 05:07.800 --> 05:08.680 Back we go. 05:09.280 --> 05:10.000 Okay. 05:10.600 --> 05:16.310 So then the next function is perform searches. 05:16.310 --> 05:19.950 This is when we're actually going to do the searches for each of these. 05:19.990 --> 05:23.990 Now do you remember I showed you that asyncio a while ago? 05:24.030 --> 05:26.670 That was the great way that you can run things in parallel. 05:26.870 --> 05:28.550 Perfect time for it right now. 05:28.870 --> 05:38.150 So we're going to create a bunch of async IO tasks to search for each item, for each item in the the 05:38.190 --> 05:39.670 results of the searches. 05:39.670 --> 05:45.910 This this this is the Pydantic object that searches feel gives you the list of the search items. 05:45.910 --> 05:48.750 And for each item we were going to call search searches. 05:48.790 --> 05:51.750 The next function we'll talk about we're going to call search. 05:51.910 --> 05:52.430 We're going to. 05:52.470 --> 05:56.310 This gives you a coroutine so we can turn that into a into a task. 05:56.510 --> 06:02.470 And then we can use this gather approach that I told you about async gather. 06:02.470 --> 06:09.990 If you then await that, it will run all of these tasks in parallel and the results will go into results. 06:10.750 --> 06:12.470 And then we will put finished. 06:12.990 --> 06:14.790 So hopefully I didn't lose you there. 06:14.790 --> 06:19.900 Basically all I'm saying is we take the results from the planner and we just search for all of these 06:19.900 --> 06:23.900 things in parallel using the search agent. 06:24.340 --> 06:26.780 And this is how we actually use the search agent. 06:26.940 --> 06:28.420 We give it an input. 06:28.620 --> 06:31.300 And what we're actually doing is we don't just tell it the query. 06:31.340 --> 06:33.780 We also tell it the reason that we're searching for that. 06:33.820 --> 06:37.580 Remember I said there's more than one one reason why we asked for a reason. 06:37.620 --> 06:40.260 It's also so we can give some context here. 06:40.500 --> 06:47.580 And then we, uh, we we run the search agent and then we return the final output of the search. 06:48.340 --> 06:49.060 Okay. 06:49.540 --> 06:51.860 Hopefully you followed these three. 06:51.900 --> 06:57.420 If not, come back and look at these as such short functions that that it should make sense. 06:57.940 --> 07:00.940 And then finally some housekeeping. 07:01.100 --> 07:02.300 Two functions. 07:02.420 --> 07:05.500 One of them to write the report, one of them to send the email. 07:05.780 --> 07:09.060 The writing the report will use a writer agent. 07:09.100 --> 07:15.180 It will, uh, run that agent that is going to reconstitute the output in HTML. 07:15.460 --> 07:20.570 And then finally we will write an email using the email agent. 07:20.890 --> 07:23.930 And that will then complete the process. 07:24.410 --> 07:29.050 So I should run these two to define these two functions. 07:29.090 --> 07:30.690 These two these five functions. 07:30.890 --> 07:34.370 And now we should be ready for showtime. 07:35.050 --> 07:35.330 All right. 07:35.330 --> 07:40.250 And if you're one of the people that doesn't like the way that I just execute code and talk to it, 07:40.290 --> 07:43.290 it's sort of lazy, then you're going to be happy. 07:43.290 --> 07:48.050 This time I will type the code so that we get to do this part together. 07:48.050 --> 07:50.210 So it is a moment of truth. 07:50.210 --> 07:57.010 We're going to say, like with trace research, Trace Curse is already filling it in. 07:57.130 --> 07:58.130 I'm not. 07:58.170 --> 07:59.170 No, no luck for me. 07:59.170 --> 08:01.570 Well, I'm going to ignore what curse is saying here. 08:01.850 --> 08:03.450 I'm not looking at that. 08:03.450 --> 08:06.610 So we're going to print a message like starting research. 08:06.970 --> 08:08.570 Sometimes cursor is too good. 08:09.130 --> 08:10.650 All right, let's get rid of cursor. 08:10.690 --> 08:11.730 No cheating. 08:14.810 --> 08:17.210 I'm not looking I'm not looking at what cursor is saying. 08:17.250 --> 08:17.610 Search. 08:17.770 --> 08:19.290 What if I type something different? 08:20.090 --> 08:22.050 Confuse cursor. 08:22.050 --> 08:22.080 sir. 08:22.600 --> 08:23.080 All right. 08:23.120 --> 08:24.040 Now search. 08:24.600 --> 08:25.440 That doesn't work. 08:25.440 --> 08:25.800 It's still. 08:25.800 --> 08:27.040 It still figures it out. 08:27.080 --> 08:28.560 All right, but there we go. 08:28.600 --> 08:29.240 It knows. 08:29.240 --> 08:34.160 So the first thing we're going to do is have a variable called search plan, which is going to await 08:34.200 --> 08:35.560 planning the searches. 08:35.560 --> 08:38.040 So if you come up here there is planning the searches. 08:38.040 --> 08:39.520 So that's the first step. 08:39.520 --> 08:40.400 We do that. 08:40.560 --> 08:41.040 All right. 08:41.040 --> 08:43.880 The next step is the search results. 08:43.880 --> 08:46.040 We're going to perform the searches. 08:46.040 --> 08:50.160 So we come up here we perform the searches passing in the search plan. 08:50.200 --> 08:53.360 That's the next step of this which we also await. 08:53.360 --> 08:57.000 The next thing we need to do is actually run the report. 08:57.000 --> 09:02.600 So we await writing the report, passing in the query and the search results. 09:03.040 --> 09:05.680 And cursor knows well as do you. 09:05.760 --> 09:09.920 You already saying what we need to do next is send the email. 09:09.920 --> 09:12.040 And that really is it. 09:12.200 --> 09:16.600 I can take out the and then I can print. 09:17.040 --> 09:22.960 All right we won't we will do something different to what cursor suggests to show that we have some 09:22.960 --> 09:24.000 autonomy here. 09:24.000 --> 09:25.910 We're not just ruled by the machines. 09:25.950 --> 09:27.510 Okay, there we go. 09:27.550 --> 09:28.350 Let's kick this off. 09:28.350 --> 09:29.470 Let's have this running. 09:29.910 --> 09:31.790 So it's starting the research. 09:31.830 --> 09:33.270 It's planning the searches. 09:33.270 --> 09:34.870 It will perform through searches. 09:34.870 --> 09:35.990 So it's. 09:36.030 --> 09:37.310 And it's already finished. 09:37.310 --> 09:37.590 So. 09:37.590 --> 09:41.950 So it used an agent to figure out what three searches to do. 09:41.990 --> 09:49.030 And then for each of those, it used a remote, a hosted, uh, tool from OpenAI to actually carry out 09:49.070 --> 09:49.990 that search. 09:50.030 --> 09:51.710 It got back all of the results. 09:51.710 --> 09:58.310 We're now using an agent to synthesize the results and turn it into a report using. 09:58.550 --> 10:05.150 It's going to generate a report in markdown, and then it's going to turn that into an HTML email. 10:05.670 --> 10:08.790 So um, we will let it do its thing. 10:09.150 --> 10:10.230 Let's just come back up. 10:10.230 --> 10:14.910 The final report is is is in this format which has markdown as one of the, the parts of it. 10:14.910 --> 10:18.070 And that's the part we're now doing the writing email. 10:18.070 --> 10:23.990 And it's in writing the email that it will both come up with a subject line, and it will also write 10:23.990 --> 10:25.830 it as an HTML email. 10:25.870 --> 10:28.270 And in a second, hopefully I will receive that email.