WEBVTT 00:00.240 --> 00:00.760 All right. 00:00.760 --> 00:03.120 So now we want to invoke the graph. 00:03.440 --> 00:05.640 So let's write app dot invoke. 00:05.680 --> 00:07.360 And let's give it the input. 00:07.800 --> 00:11.960 The input is going to be a dictionary with the messages key. 00:12.200 --> 00:17.960 And this is going to be a list of human messages in the content here is going to be for example what 00:17.960 --> 00:19.600 is the weather in Tokyo. 00:19.920 --> 00:22.200 List it and then triple it. 00:23.200 --> 00:29.680 And because we need here some real time information here, we're expecting that the agent will first 00:29.680 --> 00:35.280 deduce that it needs to call the search tool to search for the weather in Tokyo. 00:35.520 --> 00:40.280 And then after we get this result, it needs to invoke another tool, the triple tool. 00:40.440 --> 00:44.600 So after it invokes both tools, it will be able to answer the question. 00:44.920 --> 00:46.880 So this is what we're expecting here. 00:47.560 --> 00:52.840 So let's go now and let's print the result and let's go. 00:52.960 --> 00:56.400 And we want to take the last message and to take the content. 00:57.760 --> 01:00.680 Let me change the index here to last the const. 01:01.000 --> 01:02.920 And this should run. 01:02.920 --> 01:04.400 And let's see what happens here. 01:13.920 --> 01:15.800 We're still waiting for our result. 01:17.240 --> 01:20.000 So far it looks like it's running and boom! 01:20.000 --> 01:21.920 Let's take a look at the output here. 01:22.160 --> 01:24.400 So this is the current weather in Tokyo. 01:24.640 --> 01:29.560 It listed a bunch of metrics and then it took the humidity which is 69. 01:29.560 --> 01:33.240 And it multiplied it by three which is 207. 01:33.560 --> 01:35.480 So I wasn't that explicit. 01:35.480 --> 01:38.840 So let me write what is the temperature in Tokyo listed. 01:38.840 --> 01:39.760 And then triple it. 01:39.760 --> 01:41.080 Let's see what we get there. 01:53.520 --> 01:54.760 So it's still running. 01:54.800 --> 01:59.080 Notice it runs a bit long because it makes a bunch of tool calls. 01:59.080 --> 02:01.720 So it first needs to reason which is an LM call. 02:01.760 --> 02:03.120 Then it needs to call the tool. 02:03.120 --> 02:04.480 Then it needs to reason again. 02:04.920 --> 02:05.680 And we got here. 02:05.680 --> 02:10.840 The current temperature in Tokyo is 15 Celsius and triple it giving 45. 02:11.200 --> 02:14.910 So the line graph agent is working as expected now. 02:15.110 --> 02:22.230 And I think now in order to understand a bit more about what's happening with the flow, let's go and 02:22.230 --> 02:29.830 open Langschmidt and let's check out the traces of the graph execution, and let's see the nodes execution. 02:29.830 --> 02:33.590 And the tool calls to really see it with our eyes. 02:34.030 --> 02:37.030 So I'll go here to the react function calling project. 02:37.030 --> 02:39.110 And you see we have here two traces. 02:40.190 --> 02:42.110 Let me go and open the second one. 02:43.510 --> 02:46.030 So here we see something actually very interesting. 02:46.030 --> 02:51.070 Right from the get go we see that the search tool was called twice. 02:51.230 --> 02:54.550 And maybe we should expect it to be called once. 02:54.550 --> 03:00.110 But for my experience, this kind of behavior is actually quite regular because because this is a search 03:00.110 --> 03:03.950 result, then sometimes we don't get the right result. 03:03.950 --> 03:07.670 And sometimes the information that we find online is not enough. 03:07.950 --> 03:14.350 So in that case, the agent is smart enough to detect it and to run the tool again until it gets the 03:14.350 --> 03:14.870 answer. 03:14.910 --> 03:20.590 And by the way, from my experience, a technique if you want to save those tool calls For example, 03:20.590 --> 03:24.590 in the cavity search tool we can add an argument of max results. 03:24.590 --> 03:27.510 So this is the number of results that will return us. 03:27.670 --> 03:33.590 So we can configure it to not be only one result but to be five results for example. 03:33.590 --> 03:37.430 And then one of those results for sure is going to have the answer that we want. 03:37.470 --> 03:38.830 I mean that's the heuristic. 03:38.830 --> 03:40.830 It's not for sure, but it's a heuristic. 03:42.350 --> 03:46.710 So let's go and check out the trace one by one from the beginning. 03:47.030 --> 03:49.630 Let's go and check out the agent node here. 03:49.630 --> 03:50.670 The agent reason. 03:51.030 --> 03:55.790 So we can see now that we have here um our system message. 03:55.790 --> 03:57.790 And we supplied both tools here. 03:57.790 --> 04:02.230 And we see we got the answer that we should invoke the search tool here. 04:02.230 --> 04:06.670 And we should invoke the search tool with the current temperature in Tokyo. 04:07.230 --> 04:12.390 So let's go and check out now the second node we should continue. 04:12.630 --> 04:14.750 So the should continue is actually not a node. 04:14.790 --> 04:16.790 It's the conditional edge function. 04:16.950 --> 04:19.710 And right over here we can see that this is the input it gets. 04:19.950 --> 04:24.110 It checks the last message which is going to be an AI tool call. 04:24.110 --> 04:26.470 And besides it should go to the Act node. 04:26.950 --> 04:36.230 So then the graph goes to execute the node where it invokes then executes the search function. 04:36.550 --> 04:38.390 So we can see it right over here. 04:38.630 --> 04:41.150 And the input is the current temperature in Tokyo. 04:41.430 --> 04:44.590 And here we got the result here. 04:44.790 --> 04:50.510 And notice that in this search query the content that we get doesn't really contain the temperature 04:50.510 --> 04:51.150 in Tokyo. 04:51.590 --> 04:57.070 So and this is really a non-helpful search result. 04:57.510 --> 05:02.710 So what happens then because we have an edge from the node to the agent region. 05:02.750 --> 05:06.430 We're then going to give the agent this result of this tool. 05:06.430 --> 05:11.430 And it's going to figure out whether it can return the answer or it needs to do something else. 05:11.790 --> 05:19.190 So we can see here we are giving also the tool result to the reasoning engine to the reasoning node. 05:19.390 --> 05:25.550 And then it decides that it needs to run another search with the current temperature in Tokyo. 05:25.550 --> 05:29.550 And this time we can see we have here an included domain here. 05:30.230 --> 05:37.590 So we can see here that after it goes to, the should continue conditional edge, and it decides that 05:37.590 --> 05:40.750 it needs now to go and do the Act node. 05:40.750 --> 05:43.910 Let me go and open that. 05:44.110 --> 05:45.430 You can see it's right over here. 05:45.470 --> 05:49.870 Act and let's go now to the second search result. 05:50.270 --> 05:51.510 Let's go for that. 05:56.230 --> 05:56.910 All right. 05:56.910 --> 06:05.910 So here we can see that in the content of the search query we have that we have 15°C which is much better 06:05.910 --> 06:06.670 than before. 06:07.270 --> 06:11.270 So now we need to go to the node. 06:11.270 --> 06:14.950 And now the agent reason is going to have the valid search here. 06:15.430 --> 06:22.750 And we can see now that it decides that it should call the triple function with the input num to be 06:22.790 --> 06:25.950 15, which is extracted from the previous tool call. 06:26.230 --> 06:29.550 So this is exactly what we wanted to happen. 06:29.870 --> 06:34.630 It then goes and executes the triple tool in the tool node. 06:34.630 --> 06:37.150 So it's going to receive the input of 15. 06:37.190 --> 06:39.420 Then the result is going to be 45. 06:39.700 --> 06:41.660 And I'm not going to go over it. 06:41.660 --> 06:42.660 It's pretty trivial. 06:42.660 --> 06:47.580 And I'm attaching to this video's resources, this trace, which I'm going to make public. 06:47.700 --> 06:48.300 Right. 06:48.340 --> 06:54.420 So at the end, the final node that is going to be execute is going to be the agent reason. 06:54.420 --> 06:57.540 And then it decides that it doesn't need to do anything right now. 06:57.540 --> 06:58.620 So it should finish. 06:59.820 --> 07:00.260 All right. 07:00.260 --> 07:04.300 So let me go and let me go and commit this code here. 07:04.740 --> 07:12.060 So let me add the files and let me commit now and let me call this commit graph. 07:12.820 --> 07:14.260 Going to push it to the repo. 07:16.140 --> 07:19.020 And you can find the code for this section. 07:19.060 --> 07:20.140 Let me go here. 07:21.820 --> 07:25.820 So this is the graph commit with all the code that we just wrote. 07:28.820 --> 07:30.180 I hope you enjoyed this section. 07:30.180 --> 07:36.300 And the entire goal of this section was to show you how easy it is to implement the react agent with 07:36.540 --> 07:37.100 graph. 07:37.100 --> 07:40.980 So not only are you using graph here, but we leverage function calling. 07:41.020 --> 07:45.820 And this actually gives us a lot of stability and better performance for our react agent.