1 00:00:02,990 --> 00:00:08,600 So in this chapter, we are going to talk a little bit more about agents in long Chain. 2 00:00:09,470 --> 00:00:14,330 For me, the most important thing of this chapter is a. 3 00:00:15,370 --> 00:00:19,990 To tell you a few warnings regarding agents. 4 00:00:20,140 --> 00:00:30,130 So it is important for you to be aware that agents are not still a mature functionality in launching. 5 00:00:30,940 --> 00:00:38,740 So they are still in in a experimental phase, and until they improve, it is not advisable to use them 6 00:00:38,740 --> 00:00:40,960 in a professional application. 7 00:00:40,960 --> 00:00:43,810 So they are very popular. 8 00:00:43,810 --> 00:00:51,700 They created a lot of excitement at the beginning, but now we are seeing that in many occasions they 9 00:00:51,700 --> 00:00:56,080 are failing when we are trying to do serious things with them. 10 00:00:56,080 --> 00:01:05,200 So one thing is like, you know, like a toy application, like a game application, like a non serious 11 00:01:05,200 --> 00:01:08,500 application and like a hobby application. 12 00:01:08,500 --> 00:01:15,850 Another thing is a professional application where a customer is paying you to develop it or whatever, 13 00:01:15,850 --> 00:01:17,770 right, or for your startup or whatever. 14 00:01:17,770 --> 00:01:26,770 So careful with agents are appealing, are interesting, but they are still not 100% mature. 15 00:01:27,840 --> 00:01:33,600 Maybe they will soon, but at this time they are still not 100% mature. 16 00:01:33,600 --> 00:01:41,280 So using agents A also is relatively expensive in terms of OpenAI. 17 00:01:41,310 --> 00:01:50,430 This is another thing you need to be aware of, uh, when you, uh, make a different operations with 18 00:01:50,430 --> 00:01:57,480 agents in OpenAI, with ChatGPT, this is going to be more expensive than usual operations. 19 00:01:57,840 --> 00:02:07,020 A and finally, agents tend to break or get lost in infinite loops if they are not ask a question related 20 00:02:07,020 --> 00:02:09,810 to the tools they have access to. 21 00:02:09,810 --> 00:02:12,420 We are going to see an example now. 22 00:02:12,600 --> 00:02:15,360 But this is very important because. 23 00:02:16,140 --> 00:02:24,900 If your application goes into an infinite loop and you have this application connected to OpenAI or 24 00:02:24,900 --> 00:02:30,240 any other private LM model, that is going to cause you a lot of money. 25 00:02:30,240 --> 00:02:32,490 So you have to be extremely careful with that. 26 00:02:32,490 --> 00:02:39,690 And every time you work with agents, you need to limit the maximum number of iterations. 27 00:02:39,690 --> 00:02:47,100 We are going to see that in the example that they can make in order to prevent a possible disasters, 28 00:02:47,100 --> 00:02:49,380 like the one I told you. 29 00:02:49,380 --> 00:02:56,790 So in the exercise of today, A, we are going to play around a little bit with agents. 30 00:02:56,790 --> 00:03:04,380 Again, we want to create an LM application that autonomously autonomous, autonomously decides which 31 00:03:04,380 --> 00:03:08,310 tool it will use to solve the problem we present to to it. 32 00:03:08,310 --> 00:03:12,480 So this is the keyword autonomously okay. 33 00:03:12,720 --> 00:03:17,040 So we are going to use a long chain agents. 34 00:03:17,040 --> 00:03:22,260 Initially we are going to create a very simple one very similar to one you already know. 35 00:03:22,260 --> 00:03:29,430 And then we will create a customized agent based on a very simple drag application. 36 00:03:29,430 --> 00:03:38,430 So in the right side of the screen first you have our classic, you know, a m file connection. 37 00:03:38,430 --> 00:03:44,850 So we can, uh, communicate with the OpenAI API and use ChatGPT. 38 00:03:44,970 --> 00:03:50,730 The initial part of this exercise is just a very simple drag application. 39 00:03:51,240 --> 00:03:53,460 So you are already familiar with it. 40 00:03:53,730 --> 00:03:57,180 Uh, we we create the instance of the LM model. 41 00:03:57,480 --> 00:04:02,940 We load a text file, which is going to be our private document. 42 00:04:03,030 --> 00:04:09,690 We divide this document into small chunks with a text splitter. 43 00:04:09,690 --> 00:04:19,769 And finally we are going to load these, uh, chunks of text converted into numbers to embeddings. 44 00:04:19,769 --> 00:04:23,070 We are going to load the embeddings into a vector database. 45 00:04:23,920 --> 00:04:25,030 So. 46 00:04:27,220 --> 00:04:32,470 Once we have that here, you have our typical chain. 47 00:04:33,230 --> 00:04:39,500 The retrieval QA chain in order to make questions to to this private document. 48 00:04:39,500 --> 00:04:40,310 Okay. 49 00:04:40,970 --> 00:04:50,330 So now we are going to see how we can do that with an agent instead of using the default chain we were 50 00:04:50,330 --> 00:04:51,140 using before. 51 00:04:51,980 --> 00:04:57,560 Before doing that, we are going to create a simple agent. 52 00:04:57,590 --> 00:05:06,470 So in order to one of them, the most popular uses of agents, if you remember, is that we provide 53 00:05:06,590 --> 00:05:15,860 the agent with a number of tools, and the agent is going to decide, uh, which tool it will use depending 54 00:05:15,860 --> 00:05:18,080 on the question we make. 55 00:05:18,780 --> 00:05:26,580 So in this case, to simplify that to the maximum level, we are going to provide just one tool. 56 00:05:26,580 --> 00:05:31,950 So our tool is going to be the LM math which is like a math calculator tool. 57 00:05:31,950 --> 00:05:38,700 In order to to to be able to play with it, you need to install this, uh, plugin before. 58 00:05:38,700 --> 00:05:43,110 And once you have it, you know, you just define the, the tools. 59 00:05:43,110 --> 00:05:49,230 And once you have defined the tools, you can initialize the agent, as you remember with the LM, the 60 00:05:49,230 --> 00:05:56,220 tools, the LM model, etc., etc., etc. the agent type is going to be the most popular, one is going 61 00:05:56,220 --> 00:05:56,790 to be true. 62 00:05:56,790 --> 00:06:01,590 So we can see what is happening in the background and we are going to set. 63 00:06:02,230 --> 00:06:08,230 Max iterations to three, so we are preventing possible loops. 64 00:06:08,740 --> 00:06:10,810 So you will see one loop here. 65 00:06:10,810 --> 00:06:16,720 So first we are going to check this agent with a math question. 66 00:06:16,930 --> 00:06:20,260 If we make a math question the agent doesn't have any problem. 67 00:06:20,260 --> 00:06:23,110 Immediately is going to use the tool. 68 00:06:23,110 --> 00:06:31,090 You know the calculator is going to provide a good response in a very fast a way. 69 00:06:31,570 --> 00:06:37,780 But what happens if we make a question that is not related with math? 70 00:06:38,740 --> 00:06:46,570 So the agent is going to try to find a tool able to respond. 71 00:06:46,570 --> 00:06:48,790 Our question he is trying. 72 00:06:48,790 --> 00:06:55,540 If you see here he's trying to use the Google tool, but Google tool is not installed for this agent. 73 00:06:55,540 --> 00:07:02,320 So he's, you know, hitting errors, but you see that he's hitting error and trying again and trying 74 00:07:02,320 --> 00:07:02,830 again. 75 00:07:02,830 --> 00:07:12,280 And he will continue until the infinite if we haven't, uh, configured the max iteration number to 76 00:07:12,280 --> 00:07:12,580 three. 77 00:07:12,580 --> 00:07:17,110 So once it hits the third, uh, iteration, it stop. 78 00:07:17,350 --> 00:07:17,800 Okay. 79 00:07:17,800 --> 00:07:24,070 So here you have the error message agent stop due to iteration limit okay. 80 00:07:25,330 --> 00:07:31,810 So after this very simple example, just a reminder of how agents work. 81 00:07:31,900 --> 00:07:36,250 We are going to create a custom agent that is going to. 82 00:07:37,130 --> 00:07:47,270 A work as our, uh, default chain in the in the in the beginning of this, uh, exercise. 83 00:07:47,330 --> 00:07:55,760 So I am not going into the details, but the only point you need to keep in mind is that our tool, 84 00:07:55,760 --> 00:08:02,690 we are creating a custom tool here, this custom tool, what it does is it is going to search, you 85 00:08:02,690 --> 00:08:05,630 know, through our private document. 86 00:08:05,630 --> 00:08:12,560 And the most important thing to to keep in mind is that it is making use of our vector database. 87 00:08:12,560 --> 00:08:18,710 This store embeddings is the vector database we have created. 88 00:08:19,940 --> 00:08:23,330 Here in our rug application. 89 00:08:23,630 --> 00:08:27,050 So if we a. 90 00:08:28,610 --> 00:08:29,600 Uh, proceed. 91 00:08:29,600 --> 00:08:33,169 We define the agent, you know, with initialize agent. 92 00:08:33,200 --> 00:08:40,370 Uh, now, and we are using the tools with with the custom tool we have provided. 93 00:08:40,370 --> 00:08:43,940 So a maximum iteration three. 94 00:08:43,940 --> 00:08:56,630 If we ask a question, this, uh, agent is going to make use of the, uh, article search tool, which 95 00:08:56,630 --> 00:09:05,420 is the name of our custom tool and is going to find the answer we are looking for, which is a summary 96 00:09:05,420 --> 00:09:06,740 of this article. 97 00:09:07,040 --> 00:09:14,840 So apart from having this, uh, working code for you to play around and to experiment, the most important 98 00:09:14,840 --> 00:09:20,990 thing of this chapter is to remember the warnings around agents. 99 00:09:20,990 --> 00:09:29,480 So careful with agents are very interesting, a very appealing to most developers until you find out 100 00:09:29,480 --> 00:09:32,990 that in a production environment they can fail. 101 00:09:32,990 --> 00:09:35,150 So careful with them right now.