WEBVTT 00:00.160 --> 00:04.200 And welcome back to cursor and welcome to week five directory. 00:04.240 --> 00:04.800 Here we go. 00:04.840 --> 00:06.440 We're going to the first lab. 00:06.560 --> 00:12.520 Week five day one Autogen Agent Chat, which is the main part of Autogen. 00:12.520 --> 00:14.560 That's sort of comparable with, say, crew. 00:15.040 --> 00:19.280 And a lot of what we do right now is going to look very familiar because it's very consistent with crew 00:19.280 --> 00:24.880 and OpenAI agents SDK, particularly this first thing which we always do, which is to load the env 00:25.160 --> 00:26.400 as usual. 00:26.680 --> 00:27.160 All right. 00:27.160 --> 00:29.600 So the first concept then is the model. 00:30.040 --> 00:34.320 And the model which is similar to concepts like LM that we've had before. 00:34.360 --> 00:37.800 It's like a wrapper around calling a large language model. 00:38.000 --> 00:45.760 And here we import something called OpenAI Chat completion client, which is the wrapper for for the 00:45.760 --> 00:48.640 LM we'll be using, which is GPT four mini. 00:48.720 --> 00:52.280 And this is how you create your model client as it's called. 00:52.320 --> 00:54.000 And it's very simple indeed. 00:54.240 --> 00:55.960 And you just pass in the name of your model. 00:56.200 --> 00:57.400 And so let's run that. 00:57.760 --> 01:03.400 And also I just want to show you here that you could do the same thing with Obama to run a local model 01:03.400 --> 01:04.640 like llama 3.2. 01:04.680 --> 01:06.840 It's just exactly the same idea. 01:06.840 --> 01:11.440 You could run that and you could continue all of this in exactly the same way, running locally instead 01:11.440 --> 01:13.520 of using GPT four or mini. 01:14.160 --> 01:16.920 Okay, that is the first concept. 01:17.160 --> 01:19.400 The second concept is the message. 01:19.480 --> 01:23.840 This is something which is a different concept for Autogen agent chat. 01:23.880 --> 01:27.360 It's this idea that you create an object called text message. 01:27.360 --> 01:29.440 In this case that has the content. 01:29.440 --> 01:31.760 I'd like to go to London is my message right now. 01:32.040 --> 01:34.760 And the source is the user me. 01:35.120 --> 01:38.360 And so if we run that and print it, we see it's a text message. 01:38.360 --> 01:39.480 The source is the user. 01:39.680 --> 01:42.800 And the, uh, there's the content. 01:42.920 --> 01:45.760 And now that is all there is to it. 01:45.760 --> 01:47.240 That is the message. 01:47.480 --> 01:49.960 The third concept is the agent. 01:49.960 --> 01:52.760 And it's very similar to things we've seen in the past. 01:53.160 --> 01:56.400 The the thing that we import is called assistant agent. 01:56.400 --> 01:58.920 That is the uh, you'll see this many times. 01:58.920 --> 02:06.080 This is the kind of most, the most fundamental class that will work with in in Autogen agent chat. 02:06.080 --> 02:08.960 So we create a new instance of assistant agent. 02:09.000 --> 02:15.960 We give it a name, airline agent, we give it a model client, the underlying LM, we give it a system 02:15.960 --> 02:19.480 message rather like the instructions in OpenAI. 02:19.880 --> 02:21.760 You are a helpful assistant for an airline. 02:21.760 --> 02:24.240 You give a short humorous answer. 02:24.240 --> 02:26.600 So let's give it that to see what that does to it. 02:26.840 --> 02:32.920 Uh, and uh, model client stream is how we say that we want it to stream back the results. 02:33.120 --> 02:38.640 Uh, and so it's something that we've, we've already done from time to time, but that is an agent 02:38.640 --> 02:40.360 that has been created. 02:40.680 --> 02:45.320 And then the thing that brings it all together is something called on messages. 02:45.600 --> 02:51.000 That is what we call on an agent to pass in a bunch of messages, which we do right here. 02:51.000 --> 02:52.880 We just put our single message in a list. 02:52.880 --> 02:54.560 We pass that into all messages. 02:54.560 --> 02:59.790 You also have to pass in this thing called cancellation token, which is how it knows when the messages 02:59.830 --> 03:00.390 are finished. 03:00.390 --> 03:05.630 That's a sort of fiddly thing about about agent chat, but but I wouldn't worry about it. 03:05.630 --> 03:07.190 You just call, uh. 03:07.350 --> 03:08.950 And of course it's an async. 03:08.990 --> 03:11.470 It's a coroutine, so we have to await it. 03:11.630 --> 03:15.350 Await agent on messages pass in the list. 03:15.390 --> 03:20.830 You pass in this cancellation token, which is how the framework is going to know when this agent is 03:20.830 --> 03:21.550 completed. 03:21.950 --> 03:25.350 And then I'm going to print the chat message content. 03:25.350 --> 03:26.670 So we're passing in this message. 03:27.030 --> 03:32.310 Um, the messages I'd like to go to London and the assistant is, uh, your helpful assistant for an 03:32.310 --> 03:32.590 airline. 03:32.590 --> 03:34.870 You give short, humorous answers. 03:34.870 --> 03:36.590 Let's see what happens if we do this. 03:37.070 --> 03:37.870 Great choice. 03:37.870 --> 03:41.430 Just remember, if it starts raining, it's not a sign to panic. 03:41.470 --> 03:43.470 It's just London welcoming you. 03:43.710 --> 03:44.030 Ha! 03:44.270 --> 03:49.870 A nice humorous answer, of course, from our agent, and it's worth pointing out that we're having 03:49.870 --> 03:53.030 a same kind of moment as with, uh, OpenAI agents SDK. 03:53.190 --> 03:56.710 It's so easy to do this, to package it up and to make this call. 03:56.710 --> 03:59.870 It's really a nice, lightweight abstraction. 03:59.870 --> 04:05.470 There's not a lot of heaviness to this, just a lightweight abstraction around calling LMS. 04:05.670 --> 04:07.110 Well, let's take that a little bit further. 04:07.110 --> 04:09.030 Let's of course we have to bring in tools. 04:09.030 --> 04:10.430 It's always about tools. 04:10.430 --> 04:13.390 Let's bring in a tool, do something more interesting right now. 04:13.430 --> 04:16.350 Now we're going to make a tool that's going to get ticket prices. 04:16.350 --> 04:19.830 And we're going to arm our agent with the ability to look up ticket prices. 04:19.830 --> 04:24.990 And we might as well use like a SQLite database because people often like to think, say, say, okay, 04:25.030 --> 04:30.070 what would it be like if we had, uh, our agents being able to query the database? 04:30.070 --> 04:34.990 Now, there's there's sophisticated ways of doing it to actually write a SQL tool that gives agents 04:34.990 --> 04:36.270 the ability to write SQL. 04:36.310 --> 04:37.710 But in this case, it's perfectly simple. 04:37.710 --> 04:41.670 We can just write a tool that can just look up in the database. 04:41.670 --> 04:42.790 So let's do that right now. 04:42.790 --> 04:45.550 So we're going to import Sqlite3. 04:45.870 --> 04:50.710 We're going to create and we're going to delete tickets database if it already exists because we ran 04:50.710 --> 04:51.230 this before. 04:51.550 --> 04:57.590 And then once it's been deleted, we will then connect to a new database and create a table called cities, 04:57.590 --> 05:01.270 which has a city name and a round trip price. 05:01.470 --> 05:08.270 And uh, people that are familiar with this will know perfectly well that it's created a DB database 05:08.270 --> 05:09.910 that will now be empty. 05:09.910 --> 05:16.350 And we are going to populate our database with a bunch of tickets for to London, Paris, Rome, Madrid, 05:16.390 --> 05:18.190 Barcelona and Berlin. 05:18.510 --> 05:19.990 And that's been done. 05:20.150 --> 05:26.950 And we're going to write a simple query function get city price that takes a name, and it will get 05:26.950 --> 05:30.150 the round trip price to travel to the city. 05:30.550 --> 05:35.790 It will simply connect to the database and it will run a little select statement passing in the city 05:35.830 --> 05:39.390 name and return the result. 05:39.430 --> 05:44.670 And yes, for security conscious people, there's perhaps more things that I should do to make sure 05:44.710 --> 05:48.110 to validate this and make sure the city name is a city name and all the rest of it. 05:48.110 --> 05:50.550 But this is just a toy example for now. 05:50.830 --> 05:51.990 So we run this. 05:52.030 --> 05:53.870 We've got get get city price. 05:53.870 --> 05:54.830 Let's just check it out. 05:54.830 --> 05:56.790 Let's try get city price for London. 05:56.790 --> 05:58.030 It's populated in the database. 05:58.030 --> 05:59.350 We get back to 99. 05:59.350 --> 06:02.470 Let's do Rome and we get back for 99. 06:02.510 --> 06:04.110 That does appear to be working. 06:04.390 --> 06:06.230 And now check this out. 06:06.630 --> 06:08.590 This is the same as before. 06:08.830 --> 06:11.270 This, uh, creating an assistant agent. 06:11.270 --> 06:15.150 I'm calling it smart agent now because it's going to be smarter than before. 06:15.590 --> 06:17.350 Uh, otherwise that's the same. 06:17.350 --> 06:17.990 This is the same. 06:17.990 --> 06:23.590 We're passing in the same underlying client, the same underlying lm the system message. 06:23.710 --> 06:28.510 Uh, I have actually added in, but I'm not sure if it's necessary to tell it that it has the ability 06:28.510 --> 06:30.230 to look for a round trip ticket. 06:30.710 --> 06:32.430 Um, we're streaming again. 06:32.470 --> 06:37.750 We're passing in this function as one of our tools in here. 06:37.870 --> 06:45.590 And there's also this slightly curious attribute reflect on tool use, uh, which is a way of indicating 06:45.590 --> 06:48.910 that we don't just want it to return the tools results. 06:48.910 --> 06:54.190 We do want it to be able to take that and continue processing even after the tool has returned. 06:54.190 --> 06:54.430 True. 06:54.470 --> 06:56.630 We we continue until it replies. 06:56.630 --> 07:00.430 So that's like it's rare that you wouldn't want that to be true. 07:00.870 --> 07:04.950 So you should always assume that that will be your default, I think. 07:05.230 --> 07:07.910 Um, and that's I'm not thinking of something obvious. 07:08.230 --> 07:09.670 So, so this is the way we do it. 07:09.670 --> 07:14.430 And there's something to point out here, which is a tiny difference from the other frameworks, a tiny 07:14.470 --> 07:19.710 way in which autogen maybe is a little bit better, a little bit superior, that you'll notice we're 07:19.710 --> 07:23.270 just passing in this Python function directly here. 07:23.390 --> 07:25.950 We didn't have any kind of decorator or anything. 07:25.990 --> 07:29.630 You remember in OpenAI agents SDK, we had to put in a decorator. 07:29.630 --> 07:34.390 We've had to in things like Landgraaf, you have to to, uh, wrap it in a, like a tool. 07:34.510 --> 07:36.630 We haven't had to do anything like that. 07:36.670 --> 07:39.270 It's, um, it's just really lightweight. 07:39.430 --> 07:44.350 And it's because, of course, they've just got a little bit of extra stuff in the abstraction code 07:44.550 --> 07:46.710 that sees this as a Python function. 07:46.750 --> 07:51.500 It uses this comment to figure out the description of the tool. 07:51.620 --> 07:53.140 So it just does a bit of that for you. 07:53.180 --> 07:55.460 It makes life a tiny bit easier. 07:55.700 --> 07:59.820 Removing some of the of the of the learning curve, which I think is really nice. 08:00.020 --> 08:06.260 Um, so what we're now going to do is we're just simply going to call smart agent this guy on messages 08:06.300 --> 08:07.820 passing in our same message. 08:07.820 --> 08:08.980 I want to go to London. 08:09.260 --> 08:16.140 Um, and then just just for the for fun, I'm going to print the inner messages. 08:16.340 --> 08:21.500 I mentioned that the message construct isn't just for the human to the agent. 08:21.500 --> 08:24.860 It's also for what's going on between agents and inside the agent. 08:24.860 --> 08:27.740 So we can see that by printing its inner messages. 08:27.740 --> 08:30.500 And then we will print the result and let's run that. 08:30.820 --> 08:32.540 And this is what happens. 08:32.540 --> 08:35.820 So you can see first of all that there's been a function call. 08:35.820 --> 08:40.180 This was an inner message with city name set to London. 08:40.340 --> 08:43.300 And it's a get city price is the function call. 08:43.740 --> 08:45.980 Uh the results came back. 08:46.020 --> 08:49.340 The result was 2.99 and there we go. 08:49.460 --> 08:52.540 This is linked back to there, as you would hope. 08:52.900 --> 08:58.660 And then this is the final response from the model are the city of T and top hats. 08:58.820 --> 09:01.900 A round trip ticket to London will set you back 2.99. 09:01.940 --> 09:06.900 Just remember, the only thing you should pack is your sense of humor, because the weather might require 09:06.900 --> 09:07.180 it. 09:08.900 --> 09:09.900 Very droll. 09:10.100 --> 09:10.860 There we go. 09:10.980 --> 09:12.460 Uh, so, uh. 09:12.620 --> 09:13.100 Great. 09:13.100 --> 09:13.980 Great answer. 09:13.980 --> 09:19.100 But of course, more important than the great answer is the fact that it was so simple to write that 09:19.100 --> 09:22.020 tool, to have it make a SQL call to a database. 09:22.020 --> 09:23.020 That's right. 09:23.060 --> 09:23.900 Sitting right there. 09:24.220 --> 09:31.340 Uh, and to have that run use the tool, um, and it just shows, uh, first of all, um, how quick 09:31.340 --> 09:36.420 and simple agent chat is, but secondly, how good you're getting at understanding this stuff because 09:36.460 --> 09:42.460 honestly, this is so familiar to you now that this is just in a less than ten minutes, you're already 09:42.500 --> 09:44.060 an expert at agent chat.