WEBVTT 00:00.200 --> 00:06.040 Remember in the previous videos when I wanted to call the agent executor, I wanted to call it the chain 00:06.040 --> 00:08.160 and I will say it would be clearer later. 00:08.160 --> 00:11.160 So now is the time we're going to shed some light on this. 00:11.520 --> 00:18.880 So theoretically we can simply take the result and we'll extract the output key in that result. 00:18.880 --> 00:20.400 Dictionary reminder. 00:20.400 --> 00:21.840 It also has an input key. 00:22.080 --> 00:24.040 And we can take that string. 00:24.200 --> 00:30.360 And we can use the output parser parse method in order to turn it into a Pydantic object. 00:30.360 --> 00:33.120 And this is pretty much what we did in debugging. 00:33.320 --> 00:40.560 However, that is not really pythonic and we actually can use the LCL, the link chain expression language 00:40.560 --> 00:44.080 to get the same result, but more elegantly. 00:44.400 --> 00:47.480 And for that we'll be using the runnable lambda. 00:47.840 --> 00:56.240 Now, the runnable lambda in the Chain expression language is a lightweight wrapper that turns any Python 00:56.240 --> 01:04.110 function, usually a lambda function or simply callable function, into a runnable, an object that 01:04.110 --> 01:05.630 implements the link chains. 01:05.630 --> 01:06.910 Runnable interface. 01:07.030 --> 01:12.870 So this interface standard arises how components like chains, like models, are invoked in batched 01:12.870 --> 01:14.950 and streamed in link chain. 01:15.190 --> 01:22.350 So the runnable lambda is going to allow us to easily integrate custom logic or simple transformations. 01:22.590 --> 01:25.870 And basically it's going to give us the invoke interface. 01:25.870 --> 01:28.910 So it's going to turn everything to be invocable. 01:29.550 --> 01:32.950 And let me bring everything down to earth with the first example. 01:33.270 --> 01:41.710 So I want to implement the extract output variable, which is going to be a runnable because the runnable 01:41.710 --> 01:44.790 lambda class implements the runnable interface. 01:45.230 --> 01:49.070 And here we have the transformation here it's a very simple transformation. 01:49.070 --> 01:54.510 So it's going to be this lambda function over here where it's going to take the input and it's going 01:54.510 --> 01:57.910 to extract from that input which is referred to as x. 01:57.910 --> 02:02.590 It's going to take x at the output key okay. 02:02.660 --> 02:07.940 so if the input is going to be a dictionary, it's going to look up for the output key. 02:08.060 --> 02:13.220 And then this is going to be what the invoke method is going to return us here. 02:13.540 --> 02:19.860 Now this is going to be piped right after we execute the agent executor with our input. 02:19.980 --> 02:23.260 So this is going to take the agent executors output. 02:23.260 --> 02:26.740 And it's going to extract the output key of it. 02:26.780 --> 02:31.980 Yeah I know the the wording here um was double wording but this is what it's going to do. 02:32.300 --> 02:34.860 And now we want to parse it with the output parser. 02:35.020 --> 02:39.700 So I'll create another runnable lambda which is called parse output. 02:39.980 --> 02:46.700 And the implementation of it is going to be a function, a lambda function that is going to use the 02:46.740 --> 02:49.820 output parser.parse function on the input it gets. 02:50.140 --> 02:58.140 And the input this runnable is going to get is going to be the output of the extract output runnable 02:58.220 --> 02:58.500 okay. 02:58.540 --> 03:05.080 So if we're going to pipe everything together and compose everything together with the agent executor. 03:05.080 --> 03:06.400 With the pipe operator. 03:06.800 --> 03:12.360 Then first we are going to give the input is going to be passed to the agent executor. 03:12.560 --> 03:16.080 And when we invoke it, it's going to run the react agent. 03:16.080 --> 03:21.960 It's going to use the given tools and it's going to output a the response of the agent. 03:22.120 --> 03:24.320 So the output is going to be a dictionary. 03:24.480 --> 03:31.840 Now this dictionary is going to be passed to the extract output runnable which is going to extract the 03:31.840 --> 03:32.800 output field. 03:33.120 --> 03:40.200 And then the extracted string is going to be passed to the parse output which is going to parse it into 03:40.200 --> 03:42.040 the structured Pydantic model. 03:42.080 --> 03:49.400 So at the end when we're going to do chain dot invoke the, it's going to be running the compose chain 03:49.600 --> 03:55.120 on the input query, and it's going to return us a parsed structured response. 03:55.160 --> 03:57.800 Now I know it's a lot to digest. 03:57.800 --> 04:03.880 And trust me, I swear for me to learn how to use an expression language took me forever. 04:04.000 --> 04:09.430 So I want you to take a deep breath, and hopefully it will be clearer when we're going to review the 04:09.430 --> 04:12.390 traces in Lamb-smith very, very soon. 04:12.710 --> 04:12.910 Okay. 04:12.950 --> 04:15.950 So let's first run all this and let's see what we get. 04:17.390 --> 04:20.150 Let's go now and fast forward everything. 04:27.190 --> 04:29.030 And we can see we hit the breakpoint. 04:29.030 --> 04:33.870 And here we can see now the result is of type agent response here. 04:34.510 --> 04:36.110 And this is our pedantic object. 04:36.110 --> 04:39.470 And we can see the answer attribute with the answer. 04:39.670 --> 04:47.430 And we also have now the sources attribute which is going to be a list of pedantic source objects. 04:47.630 --> 04:49.590 So this is the structured output. 04:49.590 --> 04:51.950 And from here you know we can serialize it. 04:51.990 --> 04:53.550 We can do a lot of cool things here. 04:54.270 --> 04:54.630 All right. 04:54.630 --> 04:56.230 Let me show you now the trace. 04:57.070 --> 05:01.790 And in the trace you can see now the two runnable lambdas we saw from before. 05:03.030 --> 05:05.700 And let's see an example of the first one. 05:05.700 --> 05:09.740 And this is going to be the extract output runnable lambda. 05:10.300 --> 05:17.420 And here we can see that the input when we invoke this runnable is going to be what the agent executor 05:17.420 --> 05:18.100 outputted us. 05:18.100 --> 05:21.820 So it's going to be a dictionary with the input key and the output key. 05:22.260 --> 05:29.020 And the output of this runnable is going to be the content of the output key here. 05:29.020 --> 05:34.420 So it's going to be the JSON string of our object that we want to parse now. 05:34.620 --> 05:41.060 And this JSON string this is what we're going to be propagating to the output parsing lambda. 05:41.060 --> 05:42.100 So let's go here. 05:42.580 --> 05:48.380 And here we can see now that the input for this runnable is going to be the JSON string. 05:48.820 --> 05:52.300 And the output is going to be the Pydantic class. 05:52.820 --> 06:00.380 So it's going to have the answer and it's going to have the sources attributes and the overall chain. 06:00.620 --> 06:07.770 Let me now show you the overall chain is going to get the inputs and it's going to output us this pedantic 06:07.770 --> 06:08.490 object. 06:08.770 --> 06:12.010 So this was done with the link chain expression language. 06:12.290 --> 06:12.770 All right. 06:12.810 --> 06:18.130 Now I remind you that I still owe you an explanation why I did not use GPT five. 06:18.170 --> 06:24.450 Here, you see, I'm using GPT four here and not GPT five, even though it's the latest and greatest 06:24.450 --> 06:24.810 here. 06:25.210 --> 06:27.130 Now, the reason for it. 06:27.130 --> 06:31.250 Let me just show you what's going to happen if I'm going to be using GPT five. 06:31.650 --> 06:35.050 So let's go and let's go and switch it. 06:35.490 --> 06:37.370 Let me now run everything. 06:39.610 --> 06:41.770 And you can see now I get an error here. 06:42.290 --> 06:47.330 Now this is a 400 code error with unsupported parameter. 06:47.370 --> 06:49.610 Stop is not supported for this model. 06:49.930 --> 06:50.490 All right. 06:50.490 --> 06:54.090 And it's not that GPT five is a bad model. 06:54.210 --> 06:56.730 It's actually the most advanced and most capable. 06:56.930 --> 07:00.570 The reason is that this model does not support the stop argument. 07:00.690 --> 07:06.210 And in the next section, we'll be elaborating on this stop argument in the react algorithm. 07:06.210 --> 07:07.870 And why do we need it? 07:08.190 --> 07:13.990 But in fact, if we're going to be using the tool calling agent, which is going to be leveraging function 07:13.990 --> 07:20.390 calling, it should work just as fine and will be elaborating all of that in the course about function 07:20.390 --> 07:22.070 calling, about tool calling. 07:22.190 --> 07:25.190 And this is the next evolution of the agents. 07:25.230 --> 07:30.150 But what we just witnessed here, what we just saw here, this is the basic for everything. 07:30.350 --> 07:33.310 If you understand that, you will understand everything. 07:33.870 --> 07:35.310 All right let's finish it up. 07:35.310 --> 07:40.790 Let me go in format everything and let me add and commit everything here. 07:41.070 --> 07:43.550 So those are all the files that changed. 07:43.950 --> 07:46.190 And let me go and commit it. 07:46.310 --> 07:51.510 So I'll call the message here added output parsing. 07:52.150 --> 07:53.790 And let me push that. 07:54.270 --> 08:00.750 Now if you want to get the code you can simply go to the branch project slash search agent. 08:00.950 --> 08:03.590 And you can check out the commits there. 08:03.830 --> 08:08.150 And yeah, you can compare the code and even clone it and use it.