WEBVTT 00:00.990 --> 00:02.640 -: Hi, I'm gonna walk you through 00:02.640 --> 00:06.450 how to do the ReAct pattern, or the reason and act. 00:06.450 --> 00:09.180 This is the very first agentic framework 00:09.180 --> 00:11.520 and so, it's really important to know. 00:11.520 --> 00:14.520 Simon Wilson has a really great writeup of this 00:14.520 --> 00:15.570 and I think that the code 00:15.570 --> 00:17.970 that he uses here just explains it really well. 00:17.970 --> 00:20.070 So, I'm just gonna walk through this code 00:20.070 --> 00:21.750 rather than type it out together. 00:21.750 --> 00:23.910 And it's just gonna help you understand 00:23.910 --> 00:25.680 how every agent is created 00:25.680 --> 00:27.570 because this is the basic pattern 00:27.570 --> 00:30.330 that all of them were implemented with, right? 00:30.330 --> 00:34.890 So, I've updated the code for the latest modules. 00:34.890 --> 00:37.767 So, we're just gonna install openai, 00:37.767 --> 00:41.370 and we need to get an OpenAI API key. 00:41.370 --> 00:44.190 So, just gonna go to the OpenAI account, 00:44.190 --> 00:46.470 and then you can create a new key, 00:46.470 --> 00:49.620 testing udemy, 00:49.620 --> 00:51.000 call it that. 00:51.000 --> 00:51.873 Create a key. 00:54.240 --> 00:55.410 I'm just gonna copy that. 00:55.410 --> 00:57.600 Now, you don't wanna share this key with anyone 00:57.600 --> 01:00.300 because it's really important, it's a password, 01:00.300 --> 01:02.490 and you could lose all of your money 01:02.490 --> 01:04.170 if somebody else gets access to it 01:04.170 --> 01:07.380 'cause they could make calls to your GPT account. 01:07.380 --> 01:09.330 So, I'm just gonna submit that in there 01:10.260 --> 01:12.750 and that's where the OpenAI key goes in, 01:12.750 --> 01:15.090 we're using get getpass here to do this. 01:15.090 --> 01:19.080 And then, we're gonna instantiate the OpenAI client 01:19.080 --> 01:20.850 and we're gonna create a ChatBot. 01:20.850 --> 01:24.240 So we have the system message, it's gonna go in here, 01:24.240 --> 01:25.920 and then the user message, 01:25.920 --> 01:29.340 and then we're just gonna run the ChatBot 01:29.340 --> 01:30.990 and then get the responses. 01:30.990 --> 01:33.720 We're using gpt-4o for this one. 01:33.720 --> 01:36.570 And now, that's not really that important. 01:36.570 --> 01:38.460 The most important thing is the prompt 01:38.460 --> 01:41.130 and it's the prompt where all of the magic happens. 01:41.130 --> 01:42.300 This is how it works. 01:42.300 --> 01:45.360 It's asking the agent to run in a loop 01:45.360 --> 01:48.150 of Thought, Action, PAUSE, Observation. 01:48.150 --> 01:50.970 So. it's gonna have a think about what to do. 01:50.970 --> 01:52.620 It's gonna decide on an action 01:52.620 --> 01:54.630 and I'll show you what that means in a second. 01:54.630 --> 01:55.463 Then, it's gonna pause 01:55.463 --> 01:57.960 and wait for you to send it something back, 01:57.960 --> 02:01.350 and that thing that it gets back is an observation. 02:01.350 --> 02:02.370 And then at the end of the loop, 02:02.370 --> 02:04.200 it's gonna output an answer. 02:04.200 --> 02:06.450 So, it's saying to use Thought to describe your thoughts 02:06.450 --> 02:07.470 about the question you've been asked. 02:07.470 --> 02:09.840 There's a chain of Thought being implemented there. 02:09.840 --> 02:10.740 And then, use Action 02:10.740 --> 02:12.750 to run one of the actions available to you, 02:12.750 --> 02:14.490 so this is function calling 02:14.490 --> 02:17.430 or tool calling depending on how you wanna talk about it. 02:17.430 --> 02:19.050 And then, Observation is just a result 02:19.050 --> 02:20.820 of running those actions. 02:20.820 --> 02:22.530 So, this is actually before tool calling 02:22.530 --> 02:24.780 and we got us working through a lot of cajoling, 02:24.780 --> 02:27.300 but (laughs) we've given it a couple of actions here. 02:27.300 --> 02:31.380 Calculate and then also search Wikipedia, 02:31.380 --> 02:33.960 and we just said always look up things on Wikipedia 02:33.960 --> 02:35.160 if you have the opportunity to do. 02:35.160 --> 02:36.810 So, the question we're gonna ask here 02:36.810 --> 02:38.520 is what is the capital of France? 02:38.520 --> 02:39.930 And then, the thinking 02:39.930 --> 02:42.090 is that I should look up France in Wikipedia, 02:42.090 --> 02:43.170 and then it's gonna take an action. 02:43.170 --> 02:44.220 and then pause, 02:44.220 --> 02:46.020 and then you'll be again called with this. 02:46.020 --> 02:48.390 Observation, France is a country, the capital is Paris. 02:48.390 --> 02:50.250 Then, you'll output the answer is Paris. 02:50.250 --> 02:52.320 So, this is a few short example, 02:52.320 --> 02:56.940 so showing it how to use these tools that it has in place 02:56.940 --> 02:59.790 and how to run through the loop that we talked about. 02:59.790 --> 03:02.010 Cool, so then, we're just gonna use Regex 03:02.010 --> 03:03.780 to pull out the Action at the end 03:03.780 --> 03:05.340 and that gives us our answer. 03:05.340 --> 03:06.173 Now we have that, 03:06.173 --> 03:08.700 we have a function that instantiates the ChatBot 03:08.700 --> 03:10.170 and then does a number of turns. 03:10.170 --> 03:12.240 So, here is five turns, 03:12.240 --> 03:13.560 but it's just gonna keep running 03:13.560 --> 03:15.930 until the ChatBot decides it's done and gives an answer. 03:15.930 --> 03:19.140 And if there's an Action detected, 03:19.140 --> 03:21.480 so the Action here, 03:21.480 --> 03:22.770 if there's an Action in the prompt, 03:22.770 --> 03:24.870 then it's gonna do that Action. 03:24.870 --> 03:28.650 And if not, then it's just gonna return. 03:28.650 --> 03:31.050 So when we do an Action, 03:31.050 --> 03:33.120 we actually have a function that we do, 03:33.120 --> 03:35.700 so we have written a function to call Wikipedia. 03:35.700 --> 03:38.340 We just take the query that ChatGPT gives us 03:38.340 --> 03:40.260 and then send it to Wikipedia. 03:40.260 --> 03:43.350 And then, we have a function to eval. 03:43.350 --> 03:47.010 So, this is like a Python function to evaluate a formula 03:47.010 --> 03:49.560 so it can run calculations. 03:49.560 --> 03:52.050 Okay, so I'm just gonna run all of this. 03:52.050 --> 03:54.690 Oh yeah, and we need to set the API key. 03:54.690 --> 03:57.030 So if you run into that error, 03:57.030 --> 03:59.253 then you just need to set that here. 04:04.740 --> 04:06.510 Okay, and takes... 04:06.510 --> 04:07.343 Oh, yeah. 04:16.180 --> 04:17.580 Okay, so that's working, 04:17.580 --> 04:20.220 and now we can see it in action, hopefully. 04:20.220 --> 04:22.020 So, what is the capital of England? 04:22.020 --> 04:24.600 It thinks, I should look up England on Wikipedia. 04:24.600 --> 04:26.820 It does the action, it pauses, 04:26.820 --> 04:28.860 and then we send it back this, right? 04:28.860 --> 04:31.440 So, we sent back England is a country 04:31.440 --> 04:32.520 that's part of the United Kingdom, 04:32.520 --> 04:33.960 is located on the island of Great Britain, 04:33.960 --> 04:35.640 which covers 60%. 04:35.640 --> 04:39.270 So, this is from the Wikipedia call that it made 04:39.270 --> 04:41.430 and then it can give the answer, right? 04:41.430 --> 04:43.750 So, we could also test what is two plus two 04:44.910 --> 04:47.070 and it should run the evaluation. 04:47.070 --> 04:49.170 Yeah, here it's saying, calculate two plus two. 04:49.170 --> 04:50.820 So, we're gonna pass in the two plus two 04:50.820 --> 04:52.170 and we're gonna evaluate that. 04:52.170 --> 04:54.540 So now, it has access to the calculator. 04:54.540 --> 04:56.820 And when you string these two things together, 04:56.820 --> 04:57.840 we could give it more tools. 04:57.840 --> 05:00.300 Like you could give it a tool with access to your database 05:00.300 --> 05:02.610 to save data or to run queries. 05:02.610 --> 05:05.700 You could give it a tool access to an API 05:05.700 --> 05:07.110 that you have access to, 05:07.110 --> 05:08.520 really whatever it is you need. 05:08.520 --> 05:09.607 Like the weather API, right? 05:09.607 --> 05:11.460 "What is the weather tomorrow?" you could ask it 05:11.460 --> 05:12.293 and it could do that. 05:12.293 --> 05:13.590 There's a lot of potential 05:13.590 --> 05:18.450 in just giving access to the AI to specific tools 05:18.450 --> 05:20.820 and then telling it what it needs to do 05:20.820 --> 05:21.930 in order to get there. 05:21.930 --> 05:24.990 So, you can give it a longer running query. 05:24.990 --> 05:28.440 You could maybe ask it, let's just say, 05:28.440 --> 05:31.270 query what is the population 05:32.160 --> 05:35.040 of London times... 05:35.040 --> 05:37.680 Oh, we've already done London, let's do Paris. 05:37.680 --> 05:42.650 What is the population of Paris times the population, 05:44.830 --> 05:48.510 or what is the population of Paris times two 05:48.510 --> 05:52.210 plus the population 05:55.920 --> 05:57.240 of London? 05:57.240 --> 05:58.950 You know, whatever it's you want to do. 05:58.950 --> 06:00.960 But this should do a couple of hops now. 06:00.960 --> 06:02.730 So, it's looking up the Paris and London, 06:02.730 --> 06:04.530 population of Paris first. 06:04.530 --> 06:06.150 It gets the population there, 06:06.150 --> 06:08.490 then it looks up the population of London, 06:08.490 --> 06:11.820 and then it comes through and runs into an error. (laughs) 06:11.820 --> 06:13.347 So, it's that you can only concatenate tuple 06:13.347 --> 06:14.820 and not int, right? 06:14.820 --> 06:17.520 What it's done here is it's run into some sort of error 06:17.520 --> 06:19.470 when doing this calculation, 06:19.470 --> 06:22.230 and this is pretty common with AI. 06:22.230 --> 06:24.120 We could run it again and maybe it won't get stuck, 06:24.120 --> 06:26.490 but when you're using an agent, 06:26.490 --> 06:29.520 they do tend to make errors, they tend to get stuck. 06:29.520 --> 06:30.720 And here, you can see, this time, 06:30.720 --> 06:32.400 you know, didn't put the commas in 06:32.400 --> 06:35.070 and therefore, it performed accurately, right? 06:35.070 --> 06:36.877 So, that's something you could put in your prompt and say, 06:36.877 --> 06:38.820 "Don't put commas when you're calculating." 06:38.820 --> 06:42.240 But whatever it is that you do in terms of observability 06:42.240 --> 06:44.130 and identifying when it gets stuck 06:44.130 --> 06:46.500 and training it how to do better things, 06:46.500 --> 06:48.330 that is an exercise left to the reader. 06:48.330 --> 06:50.850 But you can see here that population of Paris times two 06:50.850 --> 06:53.880 plus the population of London is this, 06:53.880 --> 06:55.260 and we can check whether that's correct. 06:55.260 --> 06:57.180 Yeah, hopefully that gives you a good intuition 06:57.180 --> 06:58.950 for how agents work, 06:58.950 --> 07:02.133 and you'll be using this pattern in the future.