1 00:00:03,469 --> 00:00:16,129 Okay, so in this chapter we are going to build a simple agent using language that is the problem we 2 00:00:16,129 --> 00:00:17,060 want to solve. 3 00:00:17,680 --> 00:00:26,770 We want our LM application to be able to decide which tool is the most suitable to answer the question 4 00:00:26,770 --> 00:00:27,850 we pose. 5 00:00:29,170 --> 00:00:36,190 And we want our application in this case, to be able to search on Google and to search for information 6 00:00:36,190 --> 00:00:38,200 on a specific URL. 7 00:00:38,200 --> 00:00:43,270 So we are going to build a very simple agent with these two tools. 8 00:00:43,570 --> 00:00:50,470 He's going to be able to search on Google, and he's going to be able to search for information on a 9 00:00:50,470 --> 00:00:51,910 specific URL. 10 00:00:52,610 --> 00:01:02,900 And as an agent, we want him to decide on autonomously which tool is he going to use, depending on 11 00:01:02,900 --> 00:01:04,519 the question we have. 12 00:01:04,550 --> 00:01:14,510 So the agent is going to decide if he wants to search on Google, or if he wants to search on a particular 13 00:01:14,510 --> 00:01:20,510 URL, depending on on the question we provide to to the application. 14 00:01:20,510 --> 00:01:29,180 So what we are going to do is we are going to create an agent that knows how to decide between two tools 15 00:01:29,180 --> 00:01:34,730 doing a search on Google or connecting to a specific web page. 16 00:01:34,970 --> 00:01:37,850 This is the process we are going to follow. 17 00:01:37,880 --> 00:01:41,240 We have the code here on the right. 18 00:01:41,660 --> 00:01:50,120 So first, as always, we load the OpenAI credentials from our dot env file. 19 00:01:50,120 --> 00:01:54,770 This is the confidential file we use to store our passwords, etc., etc.. 20 00:01:56,910 --> 00:01:58,080 After that. 21 00:01:58,080 --> 00:02:06,060 What we do is load OpenAI, the module OpenAI, in order to create an instance of an LM. 22 00:02:06,690 --> 00:02:12,810 And once we have that, we are going to load the two. 23 00:02:13,900 --> 00:02:15,790 Tools we are going to use. 24 00:02:15,790 --> 00:02:19,360 The first tool we are going to use is Google Search. 25 00:02:19,900 --> 00:02:29,380 In order to make use of Google Search in your LM application, first you want to get a couple of private 26 00:02:29,380 --> 00:02:31,150 keys in Google. 27 00:02:31,150 --> 00:02:40,810 So here you have the links to go to Google in order to sign up to create your Google API key and your 28 00:02:40,810 --> 00:02:50,080 Google CC ID, and without any question you may have in the process of getting these keys, you know 29 00:02:50,080 --> 00:02:54,730 where to go ChatGPT, Google, StackOverflow, etc. it's very easy. 30 00:02:54,730 --> 00:02:56,590 It's free and it's very easy. 31 00:02:56,710 --> 00:03:03,250 Once you have these keys, you store these keys in your dot m file. 32 00:03:03,280 --> 00:03:09,790 And in this case, what we are doing here is we are getting these keys from our dot m file. 33 00:03:10,390 --> 00:03:20,290 Then in order to, uh, create a tool that is able to search on Google, we need to install the Google 34 00:03:20,290 --> 00:03:26,260 API, Python client, plugin or module to do that. 35 00:03:26,260 --> 00:03:32,290 If we want to do that from the notebook, we write this and execute this command. 36 00:03:32,530 --> 00:03:39,190 If we want to do it from the terminal, we execute this command from the terminal. 37 00:03:39,190 --> 00:03:41,350 It's different if you do it from the notebook. 38 00:03:41,350 --> 00:03:47,080 From the notebook you add the exclamation mark and from the terminal you just write this down. 39 00:03:47,080 --> 00:03:52,150 Since we have already installed this, we comment this and we don't want to do anything. 40 00:03:52,720 --> 00:03:57,310 So then we are going to create the variable. 41 00:03:57,310 --> 00:04:03,130 In this case, we name the variable search in order to make use of that functionality. 42 00:04:03,130 --> 00:04:08,770 So now we are able to use this functionality in the variable search. 43 00:04:09,070 --> 00:04:16,930 So first thing cover we now are able to search on Google from our LM application. 44 00:04:16,930 --> 00:04:22,780 The second functionality we want to be able to connect with an external URL. 45 00:04:22,780 --> 00:04:32,590 In order to do that, we are going to load or import a module called Text Requests wrapper from the 46 00:04:32,590 --> 00:04:34,870 Lang Chain Utilities library. 47 00:04:35,110 --> 00:04:39,310 And once we have this module, we can create a new variable. 48 00:04:39,310 --> 00:04:40,840 We will call it request. 49 00:04:40,840 --> 00:04:45,460 You can call it whatever you want in order to make use of this new functionality. 50 00:04:45,460 --> 00:04:53,650 Once we have these two functionalities in our in our application, we are going to set the tools that 51 00:04:53,650 --> 00:04:55,900 we that will use our agent. 52 00:04:55,900 --> 00:04:59,920 So the first thing we do is we import the tool module. 53 00:04:59,920 --> 00:05:06,940 And with this tool module we are going to create a new variable called tools that is going to store 54 00:05:06,940 --> 00:05:08,770 our two tools okay. 55 00:05:08,770 --> 00:05:11,770 So this is the first tool which is the Google search. 56 00:05:11,770 --> 00:05:16,000 And this is the second tool which is the request what we have called requests. 57 00:05:16,000 --> 00:05:21,910 What is doing this is going to connect with a particular URL on the internet okay. 58 00:05:22,180 --> 00:05:31,960 So once we have the tools defined we are going to initialize and configure our agent okay. 59 00:05:31,960 --> 00:05:40,240 In order to do that we are going to load three more modules initialize agent agent type and load tools. 60 00:05:40,240 --> 00:05:43,030 And we are going to create our agent. 61 00:05:43,120 --> 00:05:47,860 So we are going to call our agent my agent you can call it whatever you want. 62 00:05:47,860 --> 00:05:52,000 And we are going to initialize it with this functionality. 63 00:05:52,000 --> 00:05:57,190 And we are going to tell them okay, you are going to use these tools we have created. 64 00:05:57,190 --> 00:06:02,680 We are going to use this LM instance we have created here. 65 00:06:04,850 --> 00:06:05,540 Okay. 66 00:06:05,870 --> 00:06:11,780 And, uh, this is the type of agent we want to use from long chain. 67 00:06:11,780 --> 00:06:19,040 As you will see, if you go to the long chain documentation, long chain provides a few, uh, types 68 00:06:19,040 --> 00:06:20,480 of agent you can use. 69 00:06:20,480 --> 00:06:26,870 This is by far the most popular, the one we use, like, 99% of the time. 70 00:06:26,870 --> 00:06:29,030 But you have other options as well. 71 00:06:29,480 --> 00:06:38,330 We are going to define this new agent with the option verbose as true because we want to see what is 72 00:06:38,330 --> 00:06:44,300 happening, you know, on the on the background whenever we execute a command we'll see how very soon. 73 00:06:44,300 --> 00:06:54,770 And this is very important, we are setting a max iterations to three because we want to avoid here 74 00:06:54,770 --> 00:06:56,540 one of the main. 75 00:06:58,600 --> 00:07:05,380 One of the most common troubles you are going to find when working with agents, which is that in some 76 00:07:05,380 --> 00:07:13,930 cases, if the agent is not able to find an answer from your questions right away, he's going to loop 77 00:07:13,930 --> 00:07:17,290 over and over trying to get the response. 78 00:07:17,290 --> 00:07:26,530 So if you do not limit the number of attempts your agent is going to make in order to solve your questions, 79 00:07:26,530 --> 00:07:34,450 your agent can be looping, you know, forever, and your OpenAI bill is going to be high. 80 00:07:34,480 --> 00:07:36,730 So be very careful with this. 81 00:07:36,910 --> 00:07:40,390 Okay, so this is the note we make here. 82 00:07:40,390 --> 00:07:46,840 Operations with agents are more expensive than regular operations in OpenAI. 83 00:07:46,960 --> 00:07:51,160 And especially if the number of iterations is high okay. 84 00:07:51,160 --> 00:07:52,930 So be careful with this. 85 00:07:53,380 --> 00:07:58,750 So once we have defined our agent we can start asking questions. 86 00:07:58,750 --> 00:08:04,180 And we will see how our agent decides which tool to use. 87 00:08:04,360 --> 00:08:15,310 So we are going to start asking our agent which team won the 2010 soccer World Championship. 88 00:08:16,790 --> 00:08:18,020 Is the question. 89 00:08:18,020 --> 00:08:23,060 And this is the way we run the question through our agent. 90 00:08:23,060 --> 00:08:23,780 Okay. 91 00:08:24,350 --> 00:08:24,950 And. 92 00:08:25,540 --> 00:08:30,520 We are going to store this response into this new variable we create. 93 00:08:30,820 --> 00:08:39,700 So here as we have defined variables through here, we can see what are the operations our agent is 94 00:08:39,700 --> 00:08:41,559 doing in the background. 95 00:08:41,830 --> 00:08:42,940 So. 96 00:08:44,020 --> 00:08:53,500 Our agent reads the question, and once he reads the question, he needs to decide which of the two 97 00:08:53,500 --> 00:08:58,060 tools is more, uh, adequate for him to use. 98 00:08:58,090 --> 00:09:06,130 So in this case, you can see that our agent is thinking and is is thinking, okay, I should look up 99 00:09:06,130 --> 00:09:07,540 the result online. 100 00:09:07,540 --> 00:09:11,800 So I'm going to use Google Search in order to solve this question. 101 00:09:12,130 --> 00:09:19,360 So what he is doing here is querying Google asking this to Google okay Google. 102 00:09:20,300 --> 00:09:23,570 Who is the world champion in in this year. 103 00:09:23,600 --> 00:09:23,990 Okay. 104 00:09:23,990 --> 00:09:30,080 So the first response gives the answer to our agent. 105 00:09:30,080 --> 00:09:32,780 So once he. 106 00:09:34,140 --> 00:09:36,030 Process this information. 107 00:09:36,030 --> 00:09:37,830 He knows the final answer. 108 00:09:37,830 --> 00:09:40,920 So here you have the thought of our agent. 109 00:09:40,920 --> 00:09:42,900 I know the final answer. 110 00:09:44,330 --> 00:09:48,020 So he has the final answer set. 111 00:09:48,320 --> 00:09:54,800 If we print the response, we now see how our agent is answering. 112 00:09:54,800 --> 00:09:58,790 Spain won the 2010 soccer World Championship. 113 00:09:59,660 --> 00:10:09,500 If we make another type of question asking for example, okay, go to this particular URL and tell me. 114 00:10:10,330 --> 00:10:14,140 What are the comments in this web page about? 115 00:10:14,140 --> 00:10:16,000 Okay, so. 116 00:10:16,850 --> 00:10:17,750 The. 117 00:10:18,660 --> 00:10:19,530 Agent. 118 00:10:19,710 --> 00:10:27,900 Once he starts processing our question, he immediately guessed that the best way to solve this question 119 00:10:27,900 --> 00:10:29,970 is to use the second tool. 120 00:10:30,000 --> 00:10:32,250 The tool we have named request. 121 00:10:32,250 --> 00:10:40,410 And as you know, the tool that allows our agent to connect with one particular URL, one particular 122 00:10:40,410 --> 00:10:42,720 web page on the internet. 123 00:10:42,720 --> 00:10:51,420 So he's connecting with this web page and he is downloading all the HTML, all the code that is in the 124 00:10:51,420 --> 00:10:53,340 background of this web page. 125 00:10:53,340 --> 00:10:54,600 And he's processing. 126 00:10:54,720 --> 00:10:57,300 He's learning from all this code. 127 00:10:57,330 --> 00:11:05,760 As you know, a web page on the background is HTML, but this HTML is wrapping all the text that is 128 00:11:05,760 --> 00:11:06,840 in a web page. 129 00:11:06,840 --> 00:11:11,430 So our agent is reading the website, the website, it is this web page. 130 00:11:11,430 --> 00:11:18,390 And you will see that after reading the information in this web page he says, I can see the comments 131 00:11:18,390 --> 00:11:22,170 in the web page which will give me the subjects. 132 00:11:22,320 --> 00:11:26,190 So he he processes the information, blah blah blah. 133 00:11:26,190 --> 00:11:33,840 And immediately he finds out that the subjects in the comments of the web page are related to OpenAI's 134 00:11:33,840 --> 00:11:34,950 embedding API. 135 00:11:34,980 --> 00:11:35,880 La la la la la. 136 00:11:35,910 --> 00:11:43,620 Okay, so if we print the response now we know, uh, what our agent has found found out, right? 137 00:11:43,620 --> 00:11:45,900 So as you have seen. 138 00:11:46,840 --> 00:11:53,590 We have learned how to build a simple agent, and we have also learned. 139 00:11:53,590 --> 00:12:00,940 Start learning about the limitations of agents and how you have to be careful with them, especially 140 00:12:00,940 --> 00:12:04,180 when you are working in a professional environment.