WEBVTT 00:00.110 --> 00:02.040 Instructor: By the end of those couple of videos, 00:02.040 --> 00:03.870 we'll be implementing ourself 00:03.870 --> 00:05.970 the create react agent function, 00:05.970 --> 00:08.820 and specifically the agent executor. 00:08.820 --> 00:11.910 So what is the agent executor? 00:11.910 --> 00:16.560 It has a very fancy word and it sounds very, very cool 00:16.560 --> 00:20.850 and very complex, but essentially it's a while. 00:20.850 --> 00:24.780 So this is a fake implementation of the agent executor. 00:24.780 --> 00:26.700 Now this is just the pseudo code 00:26.700 --> 00:30.240 and it's essentially a while loop that keeps running, 00:30.240 --> 00:33.270 takes the user input, it takes this prompt 00:33.270 --> 00:35.880 and augments it with all the information we have 00:35.880 --> 00:38.730 about the tool for the LLMs disposal, 00:38.730 --> 00:41.363 and then sends everything to the LLM. 00:41.363 --> 00:44.040 Now the LLM returns as an answer, 00:44.040 --> 00:47.610 and if the answer contains that we need to run a tool. 00:47.610 --> 00:50.040 Then we figure out, and we'll talk about 00:50.040 --> 00:52.290 how this is done, which tool to run, 00:52.290 --> 00:56.760 we then run this tool, and then we go over and over again. 00:56.760 --> 00:58.170 So when do we stop? 00:58.170 --> 01:01.320 When the LLM decides that it has defined an answer 01:01.320 --> 01:04.950 or by default if we limit the number of iterations 01:04.950 --> 01:07.200 or number of messages we want to send. 01:07.200 --> 01:09.270 This is pretty much it, a fancy while loop 01:09.270 --> 01:11.820 where we are keep making LLM calls, 01:11.820 --> 01:15.120 then interpreting what the LLM has responded 01:15.120 --> 01:17.160 and then choosing the correct tool to use, 01:17.160 --> 01:19.983 running it, then running it over and over again. 01:23.790 --> 01:26.400 Now if we think about it a little bit deeper, 01:26.400 --> 01:28.800 we'll figure out there is a little more stuff 01:28.800 --> 01:31.770 that we need to perform during this entire process. 01:31.770 --> 01:34.620 So we first take the user query to the agent, 01:34.620 --> 01:37.500 and we then make this special LLM call. 01:37.500 --> 01:40.290 Now let's say the LLM is able to determine 01:40.290 --> 01:43.590 that we want to use the Get Text Length tool, 01:43.590 --> 01:45.450 so it gives us an answer back, 01:45.450 --> 01:47.460 but the LLM gives us text, 01:47.460 --> 01:51.330 so we need to parse this text in order to work with it. 01:51.330 --> 01:54.330 After we do that, we choose the correct tool 01:54.330 --> 01:57.840 the LLM chose for us, and we execute this tool 01:57.840 --> 01:59.580 and we get a response. 01:59.580 --> 02:02.790 Then we take this response alongside with our query 02:02.790 --> 02:04.710 and we run another iteration. 02:04.710 --> 02:06.810 So now the agent can decide whether 02:06.810 --> 02:09.660 to use the tool or another tool, 02:09.660 --> 02:13.080 or whether it has enough information to return the answer. 02:13.080 --> 02:14.400 So that's pretty much it, 02:14.400 --> 02:16.290 what's happening over and over again. 02:16.290 --> 02:18.780 And in the next couple of videos we'll be implementing 02:18.780 --> 02:20.640 all of this ourselves. 02:20.640 --> 02:22.380 I'm going to head up to desktop 02:22.380 --> 02:24.150 and I'm going to create a new directory 02:24.150 --> 02:26.670 called React Langchain. 02:26.670 --> 02:28.770 I'm going to cd into this directory 02:28.770 --> 02:31.170 and I'm going to create a new virtual environment 02:31.170 --> 02:33.210 by writing pipenv shell. 02:33.210 --> 02:34.620 The environment is created, 02:34.620 --> 02:38.266 and I'm going to install with pipenv langchain, 02:38.266 --> 02:41.490 OpenAI since we'll be using GPT-3.5, 02:41.490 --> 02:43.800 I'm going to install black formatting tool 02:43.800 --> 02:47.910 and Python.n package to help us load environment variables. 02:47.910 --> 02:48.743 Amazing. 02:48.743 --> 02:51.600 So now that we have all the packages installed, 02:51.600 --> 02:53.220 let's go and head up to PyCharm 02:53.220 --> 02:55.170 and you can use VS Code if you want. 02:55.170 --> 02:56.550 And I'm simply going to open the 02:56.550 --> 02:58.173 directory that we just created. 02:59.370 --> 03:01.470 So notice that PyCharm automatically 03:01.470 --> 03:03.480 detected our pipenv environment. 03:03.480 --> 03:04.803 So we're all good. 03:07.950 --> 03:11.610 And now let's create a new file and we'll call it .env 03:11.610 --> 03:14.610 and it's going to hold our environment variables, 03:14.610 --> 03:18.453 which we're going to store it with the OpenAI API key. 03:20.640 --> 03:21.473 Cool. 03:21.473 --> 03:23.550 So now that we've done that, let's go 03:23.550 --> 03:27.030 and create a new file and we'll call it main.py 03:27.030 --> 03:29.910 and here we'll be writing all of our code. 03:29.910 --> 03:33.450 And we want to start by loading our environment variables, 03:33.450 --> 03:36.390 which right now is the open AI API key 03:36.390 --> 03:38.820 that are stored in our .env file 03:38.820 --> 03:40.770 into our Python environment. 03:40.770 --> 03:43.830 So for that we're going to be used the Python dotenv package 03:43.830 --> 03:45.360 that we installed before. 03:45.360 --> 03:48.090 And we're going to do it with two simple commands. 03:48.090 --> 03:50.730 We're going to first import from dotenv 03:50.730 --> 03:54.060 the function load dotenv and we're going to call it 03:54.060 --> 03:56.670 and it's going to take all the values from the dotenv 03:56.670 --> 03:59.190 and load it into our environment variables. 03:59.190 --> 04:01.410 So now let's run some boilerplate code. 04:01.410 --> 04:04.620 I'm going to write if name equals main. 04:04.620 --> 04:07.770 And for our beginning print I'm going 04:07.770 --> 04:12.770 to write, print("Hello ReAct LangChain!") 04:14.670 --> 04:16.920 I'll go and click the green play button, 04:16.920 --> 04:19.170 which is currently configured on the current file, 04:19.170 --> 04:21.180 which is our main.py file. 04:21.180 --> 04:23.490 And this will run our script. 04:23.490 --> 04:25.560 So let's run it for a sanity check 04:25.560 --> 04:27.750 to see if that everything is working, 04:27.750 --> 04:31.590 and we can see the print, and we can now start coding. 04:31.590 --> 04:34.710 We'll first begin by writing our first tool, 04:34.710 --> 04:37.230 and for that we're going to be defining a function, 04:37.230 --> 04:40.050 which is called Get Text Length. 04:40.050 --> 04:42.600 And this function is going to receive text, 04:42.600 --> 04:45.480 which is a string, and it's going to return an integer. 04:45.480 --> 04:47.790 Now we'll start by defining 04:47.790 --> 04:49.920 what this function is going to do. 04:49.920 --> 04:52.680 So the description here is very important 04:52.680 --> 04:56.190 because this is going to help the LLM decide 04:56.190 --> 04:58.650 if it's going to use this tool or not, 04:58.650 --> 05:00.180 in its reasoning engine. 05:00.180 --> 05:02.880 And we'll soon see how this actually works. 05:02.880 --> 05:05.610 So I'm going to write the description of this function. 05:05.610 --> 05:09.390 I'm going to write that it returns the length of a text 05:09.390 --> 05:12.990 by characters, and the implementation is very simple, 05:12.990 --> 05:15.450 it's simply returning the len function 05:15.450 --> 05:18.300 when we inputed the text. 05:18.300 --> 05:20.610 So it will count the number of characters. 05:20.610 --> 05:22.890 For sanity checks, let's simply run it, 05:22.890 --> 05:27.633 and let's put it on a simple input of text equals to dog. 05:29.520 --> 05:33.120 And we can see that dog indeed has three characters in it. 05:33.120 --> 05:35.043 So we can see the function is working.