WEBVTT 00:00.110 --> 00:02.900 Now it's time to write our react agent. 00:02.900 --> 00:05.960 So let's define a variable called agent. 00:05.960 --> 00:08.180 And it's going to take our prompt. 00:08.180 --> 00:10.790 And it's going to plug it in into our LM. 00:10.790 --> 00:13.940 Now you might be wondering what are those pipe thingies. 00:13.940 --> 00:19.250 So this is me using the Lang chain expression language or LCL. 00:19.250 --> 00:26.360 So the Lang chain expression language is a way for us to define declaratively and compose our chains 00:26.360 --> 00:27.050 together. 00:27.050 --> 00:30.780 So there are lots of benefits of using the LCL. 00:30.780 --> 00:36.390 It makes our code more readable, more composable, and it really helps us understand what's happening 00:36.390 --> 00:38.640 under the hood when we dive into internals. 00:38.640 --> 00:45.810 And it has a lot more benefits like parallel processing, batch and streaming support fallbacks, and 00:45.810 --> 00:47.940 we'll dive into it later in this course. 00:47.940 --> 00:54.470 By the way, one of the reason it was created was because people complained that all the chains and 00:54.470 --> 00:58.910 agents were too abstract, and they didn't understand what was going on under the hood. 00:58.910 --> 01:04.190 But with the help of the LKL, we can understand exactly what's happening in each step. 01:04.190 --> 01:08.120 And what's the chronological order of those executions. 01:08.150 --> 01:13.130 Okay, let's go back to our agent and let's now describe what are we doing here. 01:13.130 --> 01:21.150 The pipe operator takes the output of its left side and plugs it in into the input of its right side. 01:21.150 --> 01:26.700 So here it's going to run the prompt step, which is going to output us a prompt value. 01:26.700 --> 01:32.340 And it's going to input it into the LM because LMS receives a prompt value. 01:32.340 --> 01:38.010 So intuitively you can think about it that we're taking our prompt and plugging it into the LM and then 01:38.010 --> 01:38.760 running it. 01:38.760 --> 01:44.540 You can see in this description the building blocks of length chain and what would they output and what 01:44.540 --> 01:46.040 input do they need to receive. 01:46.580 --> 01:50.390 Here we can see that prompt outputs us a prompt value. 01:50.390 --> 01:53.390 And this is something that the LM can receive. 01:53.390 --> 01:56.480 It receives strings and it can receive a prompt value. 01:56.870 --> 02:02.810 But this prompt is not complete yet because we haven't supplied it with our question that we want to 02:02.810 --> 02:03.440 ask our agent. 02:03.440 --> 02:06.410 To do so, we need to plug it in its input. 02:06.410 --> 02:07.640 And what does it input? 02:07.640 --> 02:12.920 It comes in the form of a dictionary with the keyword, which is going to replace the placeholders, 02:12.920 --> 02:17.090 and the value is going to be what's going to replace those placeholders. 02:17.090 --> 02:24.050 So because we partially initialize this prompt with the tools and the tool names, we only need to populate 02:24.050 --> 02:25.790 the input placeholder. 02:25.790 --> 02:32.300 So we're going to input our prompt with a dictionary that's going to contain the key of input. 02:32.300 --> 02:38.480 And its value is going to hold eventually the question that we're going to be asking our agent. 02:38.690 --> 02:42.860 So because we want everything to be dynamic, we won't be passing right here. 02:42.860 --> 02:48.140 The prompt that we're going to send to the agent, we're going to pass it when we invoke all of this 02:48.140 --> 02:48.620 chain. 02:48.620 --> 02:55.120 And when we invoke the chain, we need to supply it with a dictionary of the keywords that we're going 02:55.120 --> 02:56.560 to replace the placeholders. 02:56.560 --> 03:01.810 So because of that, instead of putting here the input, we're going to write a lambda function that 03:01.810 --> 03:06.910 is receiving a dictionary and is accessing the input key of that dictionary. 03:06.910 --> 03:10.450 So this is how we're going to be tailoring everything together. 03:10.450 --> 03:16.720 Eventually we're going to have in our agent an LLM, which is going to receive our prompt filled with 03:16.720 --> 03:17.890 all the arguments. 03:17.890 --> 03:24.400 And when we'll use the invoke function, we'll populate the input keys with the value that we're going 03:24.400 --> 03:25.900 to be giving in the invoke. 03:25.900 --> 03:28.630 And this will propagate until the LLM call. 03:28.660 --> 03:33.550 So let's now use the invoke method of this agent. 03:33.550 --> 03:39.040 And we want to pass it a dictionary where the input is going to be what we're going to be asking for 03:39.040 --> 03:39.760 the agent. 03:39.760 --> 03:44.110 And here we're asking what is the length of dog in characters? 03:44.110 --> 03:46.810 Now notice the way that we write here. 03:46.810 --> 03:54.070 The prompt is super important because our agent is going to receive it, and it's going from that determined 03:54.070 --> 03:57.310 which tool to use and what input will it receive. 03:57.310 --> 04:02.770 And even it will affect the output parsing later when we parse the results of the LLM. 04:02.920 --> 04:05.500 So let's run it and see what we get. 04:05.500 --> 04:07.870 And we can see that the LM responded. 04:07.900 --> 04:13.060 I can use the Get text Links tool to determine the length of the text dog. 04:13.060 --> 04:19.570 I will parse the text dog as an input to the get text length function, and after that we can see we 04:19.570 --> 04:21.820 have a backslash n action colon. 04:21.820 --> 04:24.670 It's going to be the name of the function get text length. 04:24.670 --> 04:28.540 Then we're going to have a backslash n action input which is going to be dog. 04:28.750 --> 04:30.940 So this is what the LM responded. 04:30.940 --> 04:33.640 And this is actually the reasoning engine. 04:33.640 --> 04:39.820 The LM got our prompt and it responded us with exactly what needs to be run, what tool needs to be 04:39.820 --> 04:40.570 selected. 04:40.570 --> 04:46.240 And this covers in the react algorithm, the query part which we plug into the agent. 04:46.240 --> 04:51.340 And how does the agent comprise an elaborate prompt which is sent to the LM. 04:51.340 --> 04:57.610 Then the LM functions as a reasoning agent to select the correct tool, and returns us a response containing 04:57.610 --> 05:01.330 all the information about which tool is selected and need to be run. 05:01.330 --> 05:05.230 And now we'll need to go and parse this output that we just saw. 05:05.410 --> 05:08.650 And now we're going to see how we can take this response. 05:08.650 --> 05:11.320 And we can actually invoke this function. 05:11.950 --> 05:19.150 So to do that we'll need to parse this text and to retrieve from it the action and action input and 05:19.150 --> 05:23.830 translate it to it, to the function that we need to call and to the arguments that we need to supply 05:23.830 --> 05:24.100 it. 05:24.130 --> 05:26.530 So we're going to be implementing that. 05:26.530 --> 05:29.350 And we're going to do that by using regular expressions. 05:29.350 --> 05:32.050 So I hope you're ready to writing some regex. 05:32.050 --> 05:33.100 Just kidding. 05:33.100 --> 05:38.500 Linkchain is going to do all the work for us because they already implemented the output parser, which 05:38.500 --> 05:45.850 take the output of the LLM in the react agent and simply parse it into the components that we need. 05:45.850 --> 05:52.900 So this output we want to pass into something which is called the react single input output parser. 05:53.200 --> 05:55.840 And we're going to import it from link chain. 05:55.840 --> 05:57.580 And we want to invoke it. 05:57.670 --> 05:59.320 So let's run it. 06:01.810 --> 06:05.560 And we can see that when we run it we'll get an error. 06:05.560 --> 06:09.100 And we have the error of could not pass the LM output. 06:09.100 --> 06:12.430 And this is probably the most common error when it comes to agent. 06:12.430 --> 06:14.170 It's an output parsing error. 06:14.170 --> 06:20.980 So remember I told you that we have a lot of weight and a lot of impact on the text that we input our 06:20.980 --> 06:21.580 agent. 06:21.580 --> 06:23.230 So this is what I meant. 06:23.230 --> 06:29.560 And this error basically means that the LM responded to us, not something in the format that the output 06:29.560 --> 06:30.580 parser expects. 06:30.580 --> 06:36.040 So the output parser expects something to be in the format of backslash n action backslash n action 06:36.040 --> 06:36.700 input. 06:36.790 --> 06:38.830 So let's go to our input. 06:38.830 --> 06:42.460 And let's go to refine it a bit and remove text from it. 06:42.460 --> 06:46.060 So hopefully this would now work correctly. 06:49.390 --> 06:51.760 So we can see that we get a response back. 06:51.760 --> 06:57.760 And we can see now that the output parser gave us something in the format of tool, each get length 06:57.760 --> 06:59.620 and the tool input is doc. 06:59.800 --> 07:02.290 So this is exactly what we need. 07:02.290 --> 07:07.930 We can also see that in the log property we have a what the LLM responded. 07:07.930 --> 07:13.990 And we can indeed see that it's in indeed in the correct backslash in action and backslash in action 07:13.990 --> 07:16.780 input format that the output parser expects. 07:16.780 --> 07:19.670 So I've been mumbling about this output parser. 07:19.670 --> 07:22.460 So let's check out how does it look and what does it do. 07:22.490 --> 07:28.670 So according to the description it parses the react style LLM call that have a single tool input. 07:28.670 --> 07:33.380 So it expects the output of the LLM to be in one of two formats. 07:33.380 --> 07:36.260 And we'll be focusing on this video in the first format. 07:36.260 --> 07:39.770 And the second format will be focused on in the next video. 07:39.860 --> 07:45.060 So the first format is the format of action and action input. 07:45.060 --> 07:52.530 And this basically the case where the LLM tells us that we need to perform some kind of action, aka 07:52.530 --> 07:58.350 select a tool and then invoke it so it will come in the format of action, which is going to be the 07:58.350 --> 08:03.720 tool name and the action input, which is going to be the input that we're going to supply our tool. 08:05.010 --> 08:10.620 So in our example the action was get text length which is the name of the function. 08:10.620 --> 08:12.780 And the action input was dog. 08:13.410 --> 08:18.930 So now let's go and show you the logic of what length chain is doing with all that output parsing. 08:18.930 --> 08:24.270 It has a function called parse which is going to be executed, and it's going to return us something 08:24.270 --> 08:25.950 which is called an agent action. 08:25.950 --> 08:32.130 And this is simply a link chain object, which is going to hold the tool name and the tool input that 08:32.130 --> 08:34.380 we want to supply it after we passed it. 08:34.380 --> 08:40.050 And that is going to help us invoke the tool that the LM selected from the reasoning engine. 08:40.050 --> 08:44.220 And let's have a look at the parsing using regular expressions. 08:44.220 --> 08:51.060 So we can see we have a regular expressions that is supposed to extract the action and the action input. 08:51.060 --> 08:55.680 And you can see in the bottom lines over here that we take what we matched. 08:55.680 --> 09:01.380 If it was in that format and we place it in the action and action input variables. 09:01.380 --> 09:08.130 And the last thing I want to show you is that we return an agent action object, which is built from 09:08.130 --> 09:13.260 the action and the tool input and the text, which was the original answer from the LM. 09:13.260 --> 09:18.030 So by taking this we can select the tool that we want to run and run it. 09:18.720 --> 09:21.630 So right now we finished with the parsing part. 09:21.630 --> 09:24.690 And now we want to go to the tool execution. 09:24.690 --> 09:29.760 So we have now all the information about which tool we want to execute. 09:29.760 --> 09:32.640 So now we just need to execute this tool. 09:33.090 --> 09:37.560 Now let's change the name of the variable res to agent step. 09:37.890 --> 09:44.370 And it's going to be of the type of agent action or agent finish. 09:44.730 --> 09:47.730 So let's import those two long chain classes. 09:47.730 --> 09:51.030 We discussed a bit about Agent Action and Agent Finish. 09:51.030 --> 09:53.250 We're going to discuss soon. 09:53.250 --> 09:54.690 Let's add a print. 09:54.690 --> 10:01.380 And now if the agent step is the type of agent action, it will hold all the information of the tool 10:01.380 --> 10:06.900 we want to run and execute and get our observation which is the result of this tool. 10:06.900 --> 10:09.450 So we want to write this logic. 10:09.450 --> 10:15.330 So if we get back from the output parser, the agent step, which is going to be an instance of the 10:15.330 --> 10:19.770 class agent action, we want to extrapolate the tool to use. 10:19.770 --> 10:21.780 So we'll get the tool name. 10:21.780 --> 10:26.010 And from the tool name we're going to find from the list of tools the tools to run. 10:26.010 --> 10:28.020 So let's extrapolate the tool name. 10:28.020 --> 10:31.620 And it's going to be the agent stepped dot tool. 10:31.620 --> 10:35.070 We'll define a variable called tool to use. 10:35.070 --> 10:37.560 And this is going to be the function that will run. 10:37.560 --> 10:39.540 And now we need to take the tool name. 10:39.540 --> 10:41.790 And we need to find the tool from it. 10:41.790 --> 10:46.260 So we'll need to define a new function which is called find tool by name. 10:46.260 --> 10:50.190 It's going to receive the tools list that we defined earlier. 10:50.190 --> 10:53.280 And it's going to receive the tool name. 10:53.550 --> 10:55.470 So let's go and copy that. 10:55.470 --> 10:57.450 And let's define this function. 10:57.570 --> 11:04.500 And this function arguments is going to be tools which is going to be a list of tools. 11:07.500 --> 11:13.020 And the second argument is going to be a tool name that we're going to find which is going to be a string. 11:13.020 --> 11:16.080 This function is going to return us a tool. 11:17.070 --> 11:20.220 And the implementation is going to be very straightforward. 11:20.220 --> 11:22.290 We're going to be iterating through the tools. 11:22.290 --> 11:30.210 And if the tool dot name is equal to the tool name that we're searching for, then we want to return 11:30.210 --> 11:30.960 the tool. 11:32.310 --> 11:39.090 And if we finished iterating and we didn't find the tool, let's raise a value error and write that 11:39.090 --> 11:41.640 we couldn't find the tool with the tool name. 11:42.720 --> 11:49.650 So now that we finished implementing the fan tool by name function, we want now to run the tool. 11:49.650 --> 11:51.840 And for that we'll be needing the tool input. 11:51.840 --> 11:55.380 So we're going to get it from the agent step dot tool input. 11:55.380 --> 11:59.760 And now let's define a new variable and we'll call it observation. 12:00.060 --> 12:02.970 And it's going to be the results of us running the tool. 12:02.970 --> 12:05.520 So it's going to be the tool to use. 12:05.520 --> 12:08.760 And the tool has the attribute of function. 12:08.760 --> 12:12.000 And this is the function that we defined the tool on. 12:12.000 --> 12:15.630 And here we're simply going to plug in the tool input. 12:16.980 --> 12:18.180 Just to be safe. 12:18.180 --> 12:23.550 I converted it into a string because I know that the function is expecting to get a string. 12:23.670 --> 12:29.280 However, of course the length chain implementation is much more robust and can handle much more cases 12:29.280 --> 12:32.880 where we need special attributes and special types for our inputs. 12:32.910 --> 12:38.430 And now let's go and print our observation which is the result of running the tool. 12:38.910 --> 12:41.730 Now let's run it and let's see what we get. 12:44.820 --> 12:48.630 We can see we got a value error that it couldn't find the tool. 12:48.630 --> 12:56.460 And this is because I passed the wrong argument to this function I passed instead of tool name the list 12:56.460 --> 12:57.050 of tools. 12:57.050 --> 13:01.190 So I'm simply going to change the second argument to be the tool name. 13:02.120 --> 13:03.830 So let's rerun it. 13:05.780 --> 13:12.620 And we can see now that we get observation equal to five because we passed to our tool the argument 13:12.620 --> 13:14.330 of dog with quotes. 13:14.330 --> 13:16.340 And it also counted the quotes. 13:17.150 --> 13:22.820 And just to make it a bit more explicit, I'm going to add to our tool a print where we're printing 13:22.820 --> 13:27.950 the tool name, and we're also mentioning the argument that we entered with. 13:30.590 --> 13:32.570 So now let's rerun it. 13:33.620 --> 13:40.760 And indeed we can see we entered the function gettext length with the dog word and wrapped with the 13:40.760 --> 13:41.510 quotes. 13:42.650 --> 13:47.660 So if we just want to make it a bit more nicer, let's remove the quotes from the word dog. 13:47.840 --> 13:51.320 And I'm just want to refine a bit more the prompt. 13:51.320 --> 13:58.040 So I'm going to simply write the final prompt, which is going to be what is the length in characters 13:58.040 --> 13:59.450 of the text dog. 14:00.590 --> 14:06.770 And when we run the agent, we can see that we indeed chose the correct tool, ran it, and we got that 14:06.770 --> 14:12.860 the observation was three, So it counted the number of letters in the word dog. 14:12.950 --> 14:18.230 So in this video we reviewed a massive part of the react implementation of Lang Chain. 14:18.230 --> 14:24.470 We saw how we can start from a user query to count the number of letters in the word dog, how we plugged 14:24.470 --> 14:25.460 it into the agent. 14:25.460 --> 14:30.500 Then the agent wrapped it up in a special call of the react algorithm. 14:30.500 --> 14:33.590 Sent this LM call to GPT 3.5. 14:33.620 --> 14:40.850 We got back a response from the LM, which had the thought and all the information of the tool selection, 14:40.850 --> 14:44.450 and this was actually the reasoning engine of the LM. 14:44.480 --> 14:51.560 We then saw how link chain parses all of this information and all of this response, and we saw then 14:51.560 --> 14:57.020 how link chain transforms it into a tool to run and to execute that tool.