WEBVTT 00:00.360 --> 00:01.440 Pressing on. 00:01.760 --> 00:06.040 So what we're now going to do is call this. 00:06.080 --> 00:10.040 It looks like I'm going to need to import async IO there. 00:10.080 --> 00:10.640 Hang on. 00:10.680 --> 00:11.640 Let's do that. 00:12.640 --> 00:14.240 Missed an import at the start. 00:14.640 --> 00:15.960 Actually let's do this properly. 00:15.960 --> 00:18.360 I'm going to take this out just so that it's nice and clean for you. 00:18.360 --> 00:23.240 I'm going to put that async import right up at the top there and rerun that cell. 00:23.560 --> 00:25.120 I like to do my imports at the top. 00:25.120 --> 00:27.520 I know people have different views on this stuff. 00:27.960 --> 00:33.520 Okay, so we are now going to say we want to write a cold sales email. 00:33.520 --> 00:35.240 That's going to be our instruction. 00:35.240 --> 00:38.720 And what we're going to do now is we're going to call this async IO. 00:39.920 --> 00:40.640 Do you remember this? 00:40.640 --> 00:46.800 I showed this to you back when we had the async briefing, that this is where you start to see the power 00:46.800 --> 00:47.800 of async IO. 00:47.920 --> 00:52.520 We're going to be running three different runs at the same time. 00:52.720 --> 00:54.680 And of course again it's not multithreading. 00:54.680 --> 00:58.200 It's not like the CPU is actually time slicing between them in some ways. 00:58.400 --> 01:03.900 But but rather it's going to be using an event loop, which will run each one, and whenever it's pausing 01:03.900 --> 01:06.420 waiting on I o it will let another one run. 01:06.420 --> 01:11.820 And because most of what these things are doing is pausing on I o that will allow all three of them 01:11.820 --> 01:20.820 to sort of alternate and be talking to to an LLM to open AI in parallel, and then we will collect back 01:20.860 --> 01:21.860 the results. 01:21.900 --> 01:27.300 What comes back from this await gather is a collection of results. 01:27.300 --> 01:30.660 And we will call final output on each one and print it. 01:30.980 --> 01:32.820 So we will let that run. 01:33.020 --> 01:39.220 And you can see that I've wrapped all of this in a trace, which means that we should see all of this 01:39.260 --> 01:40.420 in one trace. 01:40.460 --> 01:43.500 We will do in a little bit, but no surprise. 01:43.500 --> 01:49.420 Hopefully what you'll see down here is I have to click here to get a scrollable element, and you'll 01:49.420 --> 01:53.660 see that we can scroll through and read a bunch of emails. 01:54.460 --> 02:00.260 And the third one, of course, is very concise because we asked our third email writer to be concise. 02:01.040 --> 02:05.280 Okay, so so far all we've really done is multiple calls to LMS. 02:05.320 --> 02:07.640 And now the final piece of the puzzle. 02:07.640 --> 02:11.680 This is just a simple workflow of calling LMS. 02:11.720 --> 02:17.520 See if you can remember which one of the patterns this is, but it's a sort of variety of one of them. 02:17.560 --> 02:20.560 And we are going to create a new agent called sales picker. 02:20.840 --> 02:26.240 And sales picker picks the best cold sales email from the given options. 02:26.560 --> 02:27.720 Imagine you're a customer. 02:27.720 --> 02:30.400 Pick the one you're most likely to respond to. 02:30.560 --> 02:31.920 Don't give an explanation. 02:31.920 --> 02:35.000 Reply with the email you select only. 02:35.040 --> 02:37.080 It's a classic kind of prompting. 02:37.240 --> 02:40.040 You know that you often end with the constraint. 02:40.040 --> 02:44.480 With any limitation is a good way to end a system prompt. 02:44.920 --> 02:45.800 All right. 02:45.920 --> 02:50.600 So that's the message the user prompt writer called sales email. 02:50.840 --> 02:53.080 We again we have this same construct. 02:53.080 --> 02:59.560 We're going to run three agents in parallel to craft three messages. 02:59.560 --> 03:04.820 We're going to gather the outputs and then we just simply call another agent. 03:04.980 --> 03:07.740 So this is just Python code step by step. 03:07.780 --> 03:09.340 Call three agents in parallel. 03:09.380 --> 03:10.260 Take the output. 03:10.300 --> 03:11.540 Call another agent. 03:11.580 --> 03:18.220 It's just a Python script that runs these things in order, and it will then print the best email at 03:18.220 --> 03:19.140 the end of it. 03:19.500 --> 03:26.780 And because we've got the whole thing under one trace, we hope to see all of this in a single trace. 03:27.700 --> 03:31.860 So this is probably going to take I think like 30s in total. 03:31.860 --> 03:35.060 So I'll have to keep the sentence running if we want to see this live. 03:35.380 --> 03:39.100 And at the end of it we'll be able to go into the trace is run. 03:39.140 --> 03:41.980 So here is a single email that's been returned. 03:42.140 --> 03:45.220 And it is apparently the best one. 03:45.380 --> 03:48.620 But let's now go and have a look at the trace to see if we believe it. 03:49.060 --> 03:49.500 Okay. 03:49.500 --> 03:51.020 So this was our email. 03:51.060 --> 03:54.540 Now this is a link to go to see the traces in OpenAI. 03:54.860 --> 03:55.820 Here it comes. 03:56.020 --> 03:59.540 So at the top trace is selection from salespeople. 03:59.540 --> 04:01.680 Let's go and have a look into this trace. 04:01.680 --> 04:07.160 What you'll see is that, sure enough, it called the professional sales agent, and you can read the 04:07.200 --> 04:10.840 actual prompts that went in it called the engaging sales agent. 04:10.880 --> 04:15.160 It called the Busy sales agent, and then it called the sales picker. 04:15.320 --> 04:18.280 And you can go through this in the trace and see what's going on. 04:18.840 --> 04:22.560 So it's worth pointing out that we've really not done anything very magical so far. 04:23.000 --> 04:23.680 It's nice. 04:23.680 --> 04:28.200 It's elegant that we've had these wrappers, these agent constructs around things, but they're very 04:28.200 --> 04:29.160 lightweight wrappers. 04:29.160 --> 04:32.720 We've essentially simply called a total of four LMS. 04:32.720 --> 04:38.200 We've called the three sales LMS, and then we've called the one at the end to pick the email it preferred. 04:38.200 --> 04:41.800 And it's sort of elegant that you can go into the traces and see what's going on. 04:41.800 --> 04:46.000 But now it's time to get a bit more agentic and start talking about tools. 04:46.840 --> 04:52.200 So as we get into tools, let me remind you, have a think back to what we did in week one, when we 04:52.200 --> 04:58.520 had to build some boilerplate JSON to describe a function that we wanted to use as like a tool. 04:58.800 --> 05:04.050 It's quite chunky JSON that described the parameters and various other things about it. 05:04.050 --> 05:10.050 And then we had to write a function called handle tool calls, which is the one that had the the hokey 05:10.090 --> 05:14.810 if statement that I then tried to write into something a bit more fancy, but which was still a fair 05:14.810 --> 05:17.650 amount of sort of gunk, to use a technical word. 05:17.690 --> 05:23.690 And what we're now going to see is how that whole stuff can be, can be super simplified without taking 05:23.690 --> 05:25.370 any flexibility away from us. 05:25.930 --> 05:28.530 So first of all, we're just going to recap. 05:28.530 --> 05:34.130 We've got these three sales agents the professional, the engaging and the busy ones. 05:34.130 --> 05:35.210 And there they are. 05:35.450 --> 05:38.850 And let me just look at one of these sales agents. 05:38.930 --> 05:41.850 So if I look into it you can see that it's got a name. 05:41.850 --> 05:43.610 It's got the instructions. 05:43.610 --> 05:50.850 And you can see that it has things like tool choice tools, and it has like a list of tools that it 05:50.850 --> 05:53.130 has access to, which right now is empty. 05:53.290 --> 05:57.370 So just to give you a sense of what's going on under the covers, it's perfectly simple. 05:57.370 --> 05:59.450 But these are the constructs we're working with. 05:59.950 --> 06:03.910 So what we're now going to do is have a tool. 06:04.350 --> 06:07.550 So this is a function send email. 06:07.550 --> 06:10.470 It takes an email body which is a string. 06:10.710 --> 06:13.830 And it has a like a doc string. 06:13.990 --> 06:17.830 Send out an email with the given body to all sales prospects. 06:18.150 --> 06:21.950 Now what I've got in here is the from email. 06:21.950 --> 06:28.710 And this email needs to be what's called the verified sender, the email address that you've set up 06:28.710 --> 06:31.230 in SendGrid that you're allowed to send from. 06:31.270 --> 06:33.030 Otherwise it won't actually send. 06:33.110 --> 06:36.230 So you'll need to update this and then two email. 06:36.230 --> 06:40.030 You should put that to be an email where you don't mind it receiving an email. 06:40.070 --> 06:42.910 I would prefer that you change it from my Gmail here. 06:42.910 --> 06:46.350 Otherwise I will get flooded with emails from people doing this. 06:46.350 --> 06:49.990 So maybe I'll change it before I commit this to GitHub. 06:49.990 --> 06:52.670 But yes, send that to another of your email addresses. 06:52.670 --> 06:54.430 Or you could probably just send it to yourself. 06:54.470 --> 07:00.290 I'm not sure, but send it to somewhere where you don't mind receiving cold sales emails, there'll 07:00.290 --> 07:00.970 only be 1 or 2. 07:00.970 --> 07:02.730 We're not going to send a mass of them. 07:03.210 --> 07:05.730 And then this is Sendgrid's API. 07:05.770 --> 07:07.170 It's perfectly simple. 07:07.170 --> 07:07.890 It's SendGrid. 07:10.290 --> 07:13.130 And you give it a body so you can look through this. 07:13.130 --> 07:20.890 But it's honestly it's just a boilerplate to send an email to a recipient and we respond with success. 07:21.130 --> 07:24.410 So this is very much like the tools that we wrote last time. 07:24.410 --> 07:25.570 It's like a function to do it. 07:25.610 --> 07:29.930 And if you remember, the next step was to write a ton of JSON to describe this. 07:29.930 --> 07:36.650 But all we need to do with OpenAI is put one of these decorators above it at function tool. 07:36.770 --> 07:41.850 If you're not sure how decorators work, again, that's in the guides, or just ask ChatGPT to explain 07:41.850 --> 07:42.290 it. 07:42.290 --> 07:48.290 But that allows OpenAI to write some code to sort of do something with this function. 07:48.330 --> 07:54.930 And but all we need to care about is we simply put that decorator above the function that we want to 07:54.970 --> 07:56.530 treat as a tool. 07:57.050 --> 08:03.110 And so we will just simply define that function doesn't look like we've done anything, but let's look 08:03.150 --> 08:05.150 at this thing called send email. 08:05.190 --> 08:07.270 Now we think it's a function, right? 08:07.510 --> 08:10.870 But no, it's not a function because we have this decorator. 08:10.870 --> 08:14.670 It's been converted into something called a function tool. 08:15.110 --> 08:17.390 It has a name send email. 08:17.710 --> 08:19.030 It has a description. 08:19.030 --> 08:24.070 And that description is taken from the docstring comment we had right here. 08:24.070 --> 08:24.870 Look at that. 08:24.870 --> 08:27.150 It's just turned that into a tool description. 08:27.150 --> 08:31.790 And then and this is the great stuff params JSON schema. 08:32.030 --> 08:37.990 This stuff is all of the boilerplate JSON that we had to write last time. 08:37.990 --> 08:42.190 It's just done it for us because it's read this and it's figured out what's needed. 08:42.390 --> 08:44.230 And and that's basically it. 08:44.390 --> 08:51.790 So just by simply adding that decorator function tool, all of the work of faffing around with these 08:51.790 --> 08:55.430 JSON objects has been taken away from us, which is great. 08:55.430 --> 08:56.950 That's what a good framework should do. 08:56.990 --> 08:59.990 Take out the boilerplate but leave us in full control.