WEBVTT 00:00.400 --> 00:02.120 What we're going to do now is pretty cool. 00:02.200 --> 00:07.880 We're going to use the tools that we worked with last week in long chain. 00:08.280 --> 00:15.320 Autogen has a really easy way to wrap long chain tools so that you can call them directly from within 00:15.320 --> 00:16.040 Autogen. 00:16.200 --> 00:22.400 So when we were working with Landgraff last week, we were using long chains tools, and now in Autogen, 00:22.440 --> 00:24.880 we're also going to be able to use long chains tools. 00:24.920 --> 00:30.400 And that's great because Long Chain has a huge tools community and lots to choose from, and that's 00:30.440 --> 00:30.720 great. 00:30.720 --> 00:35.120 And we've already used many of them, so we can just suddenly have access to all of them from within 00:35.120 --> 00:36.920 the Autogen ecosystem. 00:37.320 --> 00:44.080 So within Autogen, there is this long chain tool adapter, and you can use that to wrap any long chain 00:44.080 --> 00:46.080 tool and it becomes an autogen tool. 00:46.760 --> 00:52.280 So for example, uh, this, this uh, this is importing some of long chains classes. 00:52.280 --> 00:53.720 We've already used these before. 00:53.880 --> 00:59.520 The Google Serp API wrapper is a way to call the Google Surfer API that we've we've used that We've 00:59.560 --> 01:02.880 got access to the file management toolkit. 01:03.040 --> 01:08.360 Was that great set of tools for reading and writing files with from a particular directory. 01:08.760 --> 01:12.080 And then this is just the generic Lang chain tool class. 01:12.840 --> 01:16.280 So just before we do that, let me tell you the prompt that we're going to have for an LM. 01:16.400 --> 01:22.920 Your task is to find one way nonstop flights from New York's John F Kennedy to London's Heathrow Airport 01:22.920 --> 01:23.760 in June. 01:24.200 --> 01:29.400 Search online, write the details to a file and select the best one. 01:29.880 --> 01:31.680 So then this code. 01:31.680 --> 01:36.960 These two lines here are taken identically from what we did last week. 01:37.160 --> 01:39.960 This is creating a Google search API wrapper. 01:39.960 --> 01:46.960 And this is the Lang chain code to create a lang chain tool for searching the internet with a description 01:46.960 --> 01:48.360 based on this function. 01:48.360 --> 01:53.360 And we could also do the same thing for things like sending a push notification, and for running Python 01:53.360 --> 01:55.960 code and everything else that we did last time. 01:56.400 --> 02:00.310 And so we end up with, with, with this long chain tool. 02:00.630 --> 02:04.150 And now I can call long chain tool adapter. 02:04.150 --> 02:08.590 Create a new instance of that passing in the long chain tool. 02:08.710 --> 02:14.550 And by passing that in this adapter adapts that long chain tool to become an autogen tool. 02:14.830 --> 02:16.670 It's just sort of a wrapper around it. 02:17.030 --> 02:19.550 And so I put that in a little list called Autogen tools. 02:20.070 --> 02:26.750 And then I've also collected some long chain file tools by getting the file management toolkit giving 02:26.750 --> 02:31.070 it a directory sandbox I've created an empty directory sandbox on the left. 02:31.350 --> 02:32.830 And we call Get tools. 02:32.830 --> 02:34.310 And we get a bunch of tools. 02:34.550 --> 02:42.150 And for each of those, I'm going to add them to my autogen tools by appending a Lang chain tool adapter 02:42.150 --> 02:48.470 and Autogen piece of Autogen code that adapts Lang Chains tool to be an Autogen tool. 02:48.630 --> 02:51.110 That's what this does right here. 02:51.590 --> 02:53.030 And then I'm going to print each one out. 02:53.030 --> 02:55.950 So we get to see what is this set of tools that we've got. 02:56.470 --> 02:57.230 And then. 02:57.550 --> 03:00.630 Then I'm simply going to call an autogen. 03:00.950 --> 03:02.750 Agent with this prompt. 03:02.750 --> 03:07.190 Find me a flight, use your tools and I'm going to give it. 03:07.230 --> 03:07.790 Of course. 03:07.830 --> 03:08.590 Tools. 03:08.750 --> 03:09.310 Here they are. 03:09.350 --> 03:14.670 I pass in the autogen tools so that that is what it can use. 03:14.790 --> 03:16.950 And then I will print out the inner messages. 03:16.950 --> 03:21.150 So we see it doing its thing and then get it to display the outcome. 03:21.550 --> 03:22.990 All right let's run this. 03:22.990 --> 03:24.430 Let's see what it actually does. 03:25.630 --> 03:30.790 So what you'll see is that it uh it's printed out the different tools it has access to. 03:30.830 --> 03:35.110 It has access to do an internet search to file move things around. 03:35.190 --> 03:36.870 So what's it actually done? 03:37.310 --> 03:43.030 So it's done a function call to to do a query, an internet query for finding the non-stop flights. 03:43.590 --> 03:47.830 Uh, it then, uh, got the results and it sound. 03:47.870 --> 03:51.030 It's replied with, I have found several promising deals. 03:51.030 --> 03:54.790 I will now write the details to a to a file called. 03:54.990 --> 04:01.310 Now, one of the interesting things about Autogen is the way that it handles the interactions that agents 04:01.310 --> 04:04.070 typically, uh, will then stop. 04:04.230 --> 04:07.070 Um, particularly in this kind of mode of working with them. 04:07.110 --> 04:13.830 And so what we now do is that I'm just going to send the next message of, okay, proceed, send a second 04:13.830 --> 04:17.070 message to this agent in the same way. 04:17.990 --> 04:19.670 And we'll see what it comes back with. 04:22.790 --> 04:26.750 So it's thinking about that which which probably means it's streaming back something big. 04:28.830 --> 04:29.550 And here we go. 04:29.630 --> 04:32.030 The details of the flights have been saved to flights. 04:32.790 --> 04:34.630 Now I will select the best flight. 04:34.630 --> 04:36.630 The selected flight is all of this. 04:36.670 --> 04:41.470 This flight offers a good balance of price and service, making it a great option for travel. 04:41.670 --> 04:42.190 Sounds good. 04:42.230 --> 04:44.310 It's with Virgin Atlantic, a great airline. 04:44.470 --> 04:45.430 And there's details. 04:45.430 --> 04:51.190 And this of course, has actually been, uh, grabbed from the true internet search. 04:51.190 --> 04:53.470 And it is like a useful piece of information. 04:53.790 --> 04:57.230 And of course, I know you're wondering, did it write a file? 04:57.230 --> 04:58.550 Let's go and have a look. 04:58.550 --> 05:01.310 We open up sandbox and there's a file flight MD. 05:01.350 --> 05:03.910 Let's let's open a preview of that file. 05:04.230 --> 05:05.670 And here we go. 05:05.670 --> 05:12.310 Here are a bunch of different options nicely formatted with a bunch of different, uh, things. 05:12.590 --> 05:19.110 Uh, does seem interestingly as if there was a cheaper option, but it has selected, uh, Virgin, 05:19.110 --> 05:22.790 which is one of my favorite airlines, but is the most expensive of the options. 05:22.790 --> 05:26.110 But I guess we didn't necessarily say the cheapest, uh, did we? 05:26.150 --> 05:31.030 I'll go back and check out the the prompt, um, the best flight. 05:31.190 --> 05:35.430 So, uh, um, a good choice of food and drinks and on board. 05:35.430 --> 05:38.230 Let's see, uh, promising deals. 05:38.230 --> 05:38.390 Yes. 05:38.390 --> 05:40.590 So we don't necessarily say that it needs to be the cheapest. 05:40.590 --> 05:45.990 So it chose, for whatever reason, to recommend the Virgin flight out of those flights. 05:46.310 --> 05:47.830 And that is what it did. 05:47.870 --> 05:54.420 But by all means you can rerun it asking for the cheapest and hopefully Gpt4 mini can handle that task. 05:55.060 --> 06:00.660 Okay, so we've shown how you can wrap Lang chain tools and access them from Autogen. 06:00.660 --> 06:05.100 And a little mini side mission for you right now is to go and find some more tools. 06:05.140 --> 06:10.220 Bring in the Python one, and have it be able to multiply pi by three and try some other tools. 06:10.220 --> 06:15.580 There are so many tools in the Lang chain ecosystem that now you can be unleashed to make Autogen do 06:15.580 --> 06:17.620 all sorts of cool things for you. 06:17.940 --> 06:24.780 Uh, and, um, okay, next up we have, uh, team interactions. 06:25.100 --> 06:26.980 So check this out. 06:26.980 --> 06:30.900 I'm going to take you through this pretty quickly because it's fairly self-explanatory. 06:31.220 --> 06:34.100 Uh, so you can create multiple assistants. 06:34.100 --> 06:39.540 So in this case, we're going to have, um, an assistant agent, uh, called primary and one called 06:39.540 --> 06:40.620 evaluator. 06:41.020 --> 06:43.660 Any guesses what pattern I might be about to work with here? 06:43.980 --> 06:49.300 So the prompt is your task is to find a one way non-stop flight from JFK to London London-heathrow uh, 06:49.500 --> 06:49.940 first. 06:49.980 --> 06:53.380 Search online for promising deals, then reply with the best option you found. 06:53.620 --> 06:56.860 Each time you're called, you should reply with a different option. 06:57.300 --> 06:59.260 So then we. 06:59.300 --> 07:01.860 Create a primary agent, which is the, uh. 07:02.020 --> 07:03.700 It's only going to have access to the. 07:03.740 --> 07:04.740 Sherpa to search. 07:04.740 --> 07:06.940 We're no longer writing files at this point. 07:07.260 --> 07:12.060 Uh, so it's the one that's going to be doing internet searches and it's system messages. 07:12.060 --> 07:15.300 You're a helpful AI research assistant who looks for promising deals. 07:15.340 --> 07:16.900 Respond only with one option. 07:17.220 --> 07:19.460 And, uh, make sure it's different from others. 07:19.780 --> 07:26.060 And then we have an evaluator check whether it looks like the assistant has given a very promising recommendation. 07:26.060 --> 07:26.660 Respond. 07:26.700 --> 07:28.500 Approve when you're satisfied. 07:28.540 --> 07:30.700 If you've only seen one reply, then don't approve. 07:30.740 --> 07:31.860 You need to see more. 07:32.060 --> 07:37.860 You can tell slightly artificially making this so that we get some sort of interactions between them. 07:37.860 --> 07:43.020 If this were a real, uh, proper challenge you were trying to set, you might be perfectly happy with 07:43.020 --> 07:46.820 the first option if the evaluator agent is satisfied with it. 07:47.140 --> 07:48.380 Um, okay. 07:48.420 --> 07:51.180 So this this is the, uh, the way it works. 07:51.380 --> 07:57.220 Um, and then we create a team, and there's various ways to create teams. 07:57.220 --> 08:02.900 This is a bit like a crew and crew, but it's, uh, it's really, uh, somewhat simpler. 08:02.900 --> 08:03.420 Really. 08:03.540 --> 08:10.220 This is a round robin group chat, which means, like, one after the other, obviously. 08:10.420 --> 08:15.100 Um, and, uh, that is the simplest way that you could have some kind of relationship between them. 08:15.500 --> 08:20.860 And we pass in a list of agents to talk to each other, primary agent, and then an evaluation agent 08:20.980 --> 08:24.380 and a termination condition that tells it. 08:24.420 --> 08:26.980 When do you know that enough is enough? 08:27.260 --> 08:31.020 And this is, uh, you can see here how it's defined. 08:31.220 --> 08:33.340 Text mention termination. 08:33.580 --> 08:35.260 The word approve. 08:35.700 --> 08:37.380 So this is a little bit brittle. 08:37.380 --> 08:43.060 I'm relying on the fact that the assistant, the evaluator agent, will reply the word approve. 08:43.420 --> 08:46.770 Normally you would want something a little bit more profound than that. 08:46.770 --> 08:52.450 You would probably want to reply, have structured outputs here and use that to test. 08:52.450 --> 08:53.810 But this is perfectly good. 08:53.810 --> 08:54.210 For now. 08:54.210 --> 08:55.890 I just want to get you a feel for this. 08:55.890 --> 09:00.890 I don't necessarily want to give you a full bells and whistles solution. 09:01.290 --> 09:02.450 And then this. 09:02.450 --> 09:05.050 This is how it comes together. 09:05.050 --> 09:09.290 You call team run and you await team run. 09:10.050 --> 09:11.410 And it might be worth me mentioning. 09:11.450 --> 09:15.250 Actually I've been focusing a lot on the on messages. 09:15.650 --> 09:18.650 The on messages, uh, thing that you call agents here. 09:18.650 --> 09:21.050 You can also call Agent Run as well. 09:21.090 --> 09:22.890 An agent can be called with run. 09:22.890 --> 09:25.650 And then you just pass in this exactly the same thing. 09:25.650 --> 09:30.770 You pass in the task, um, task equals and a prompt. 09:30.770 --> 09:33.610 And what comes back is just going to be the final messages. 09:33.610 --> 09:35.730 So that is another way of doing it as well. 09:35.970 --> 09:37.890 Um, but this is what we will do now. 09:38.090 --> 09:38.330 Okay. 09:38.370 --> 09:42.610 I've actually just gone and tightened these prompts a bit and made them just shorter and simpler because 09:42.610 --> 09:42.730 it. 09:42.730 --> 09:43.450 It was going. 09:43.490 --> 09:45.050 It was thrashing around a bit. 09:45.090 --> 09:48.290 The prompt is just find a one way non-stop flight. 09:49.090 --> 09:53.970 The system message says you are a helpful AI research assistant who looks for promising deals on flights. 09:54.050 --> 09:56.690 Incorporate any feedback, the evaluator says. 09:56.690 --> 09:58.530 Provide constructive feedback. 09:58.530 --> 10:00.330 Respond with approve. 10:00.490 --> 10:04.850 And so with that, I'm now kicking off our message here. 10:04.850 --> 10:11.130 And it does take it's quite chatty I discovered and it goes backwards and forwards. 10:11.130 --> 10:15.650 And I'm sure there's ways you can tighten up these prompts to make the conversation be a little bit 10:15.650 --> 10:16.290 crisper. 10:16.650 --> 10:18.890 But I will let this thing run and let it go for a little bit. 10:18.890 --> 10:24.210 And when we are back in a sec, we'll see the response from the agents. 10:24.250 --> 10:28.770 Well, I'm going to be honest with you, this is my second shot at it that I'm showing you here. 10:28.810 --> 10:33.610 I kicked this off even with these shorter prompts, and it was running for more than a minute, like 10:33.610 --> 10:34.690 going backwards and forwards. 10:34.690 --> 10:36.730 And GPT four and mini is quick, so. 10:36.890 --> 10:39.490 So I imagine it was in quite a conversation there. 10:39.690 --> 10:41.410 Uh, and so I interrupted it. 10:41.410 --> 10:46.010 I restarted my my lab and then I ran it a second time and it took 10s. 10:46.010 --> 10:46.690 And it's done. 10:46.850 --> 10:53.890 So you may want to be if you if you run it like this, then then obviously interrupt the, the, uh, 10:53.970 --> 10:59.850 the notebook by pressing the, the, uh, the button here when it's running to stop it or restart, 10:59.890 --> 11:01.130 restart the kernel. 11:01.250 --> 11:05.930 Um, if it's thrashing because you don't want to spend all of your, your, your tokens on this on this 11:05.930 --> 11:06.290 message. 11:06.290 --> 11:14.730 But I mean, I think this is a very eye opening sign of what can be troubling with these sorts of agent 11:14.730 --> 11:19.530 platforms that they can be quite hard to get them to perform the way you want. 11:19.810 --> 11:24.650 Um, you may have discovered when you're working with Landgraaf that there is a maximum, a recursion 11:24.650 --> 11:30.050 limit, as they call it, of 25, meaning that if you have more than 25 conversations between agents, 11:30.050 --> 11:34.010 then it will throw an exception unless you explicitly say you want more than that. 11:34.050 --> 11:38.090 And it appears that that there isn't the same kind of protection in an autogen. 11:38.090 --> 11:40.130 So it just kept and kept and kept on going. 11:40.490 --> 11:46.130 Uh, so, uh, be sure to to restart the kernel if that happens to you, or try and tighten up this 11:46.130 --> 11:47.810 prompt so that it doesn't happen. 11:47.970 --> 11:54.610 And observe that this is the kind of instability that is problematic, uh, with these kinds of agent 11:54.610 --> 11:58.410 frameworks, when there is autonomy to decide to keep going if they wish. 11:58.730 --> 12:00.330 But anyway, this time it took 10s. 12:00.330 --> 12:00.890 It was quick. 12:00.890 --> 12:02.490 And let's see what actually happened. 12:02.650 --> 12:06.130 The user said find a non-stop flight or a one way flight. 12:06.170 --> 12:07.330 A non-stop uh. 12:07.370 --> 12:12.690 The primary agent used the tools to do a query, and then it said it gave a response. 12:13.050 --> 12:18.090 Uh, with, with the the $402 flight there. 12:18.410 --> 12:23.690 And the evaluator responded, your response contains a lot of good information, but it lacks a specific 12:23.690 --> 12:24.970 answer to the question. 12:25.330 --> 12:27.570 And then it gives it's quite authoritative. 12:27.610 --> 12:32.490 Little does it know it's talking to itself, but it's quite, quite, quite stern. 12:32.650 --> 12:34.130 Focus on the user's request. 12:34.130 --> 12:37.760 Clarify dates, organize information, remove redundant. 12:37.760 --> 12:42.200 So it's really great to see the evaluator doing what it's told to do and evaluating. 12:42.480 --> 12:44.760 And then the primary says, thank you for the feedback. 12:44.760 --> 12:46.640 Here's a revised response. 12:46.640 --> 12:49.920 And it does indeed give a revised response with all that detail. 12:50.200 --> 12:54.600 And the evaluator sends approve which is the text that it has to respond. 12:54.640 --> 13:00.680 This is a little bit hacky, but it responds to the word approve because reply approve and your feedback 13:00.720 --> 13:05.080 is addressed and because it says approve, that means that the termination condition is met. 13:05.080 --> 13:06.800 And this completes. 13:06.920 --> 13:11.280 And that is how we arrive at our decent answer here. 13:11.280 --> 13:13.840 And this is an example of a team at work. 13:13.840 --> 13:17.840 And there's a lot more that you can do with teams, but I feel like it's pretty consistent with stuff 13:17.840 --> 13:18.600 we've done before. 13:18.680 --> 13:21.040 So this gives you enough of a flavor for it. 13:21.040 --> 13:25.240 Now, should you wish to build out your teams, you can look at some of the docs and build out something 13:25.240 --> 13:27.000 a bit more extensive than this. 13:27.400 --> 13:27.760 And. 13:27.800 --> 13:28.720 But just be aware. 13:28.760 --> 13:32.760 Be be aware of the fact that they can just just keep going. 13:32.800 --> 13:36.240 And you need to be you need to be ready to intercept if necessary.