WEBVTT 00:00.800 --> 00:06.160 And let's go and run this so we can see the print from the beginning. 00:06.400 --> 00:11.840 And we can see also the print inside the search function, the search tool. 00:12.000 --> 00:17.520 And we can see that the query that was actually sent is called weather in Tokyo. 00:18.480 --> 00:20.800 And now we get the response. 00:20.800 --> 00:23.920 We have tons of things in here in the response. 00:23.920 --> 00:26.600 So by the end of this course we'll know most of them. 00:26.640 --> 00:27.720 What do they mean. 00:28.040 --> 00:34.760 And let me go and let me run this in debug mode so we can see this more nicely. 00:34.760 --> 00:39.560 And we can figure out what's the actual output of the agent here. 00:39.880 --> 00:47.960 So all right so let me put a breakpoint here and if I'll check out the results variable. 00:48.320 --> 00:51.560 So it's a dictionary containing the key of messages. 00:51.560 --> 00:54.400 Here the first message is going to be the input. 00:54.400 --> 00:59.840 So you can see it's a human message with the content of what is the weather in Tokyo. 01:00.280 --> 01:06.260 And the answer is in the last message, and we can see it's labeled as an AI message, where it says 01:06.260 --> 01:09.380 that the weather in Tokyo is currently sunny. 01:09.540 --> 01:11.180 Now, what's in between? 01:11.180 --> 01:14.620 We have here an AI message and a message. 01:14.780 --> 01:20.300 And those are actually what is the agent is doing in order to get this answer. 01:20.300 --> 01:23.500 So right now it is sort of like magic. 01:23.660 --> 01:28.500 But in this video you'll have a better understanding how this is working. 01:28.660 --> 01:34.940 And by the end of this section and the next section, you'll know exactly how this is working and there 01:34.940 --> 01:36.220 will be no magic. 01:36.260 --> 01:37.940 Okay, so trust me on this. 01:37.980 --> 01:38.420 Cool. 01:38.420 --> 01:44.260 So let me now open links and let's see the trace of this agent execution. 01:44.260 --> 01:47.380 And let's dive a bit deeper of what's happening here. 01:47.620 --> 01:51.940 Now you can see now that the title here of this execution is called Land Graph. 01:51.940 --> 01:58.700 And this is because underneath the hood in the internals, it's going to be executing via the graph 01:58.700 --> 01:59.220 framework. 01:59.500 --> 02:06.360 Do not worry about this I promise we will be covering this in the course we open now the first call. 02:06.400 --> 02:08.680 So it's called here chat open AI. 02:09.000 --> 02:15.040 And by the way something interesting that the default model for link chain currently and this is was 02:15.040 --> 02:17.640 filmed on November 7th, 2025. 02:17.880 --> 02:23.480 So the current A model is GPT 3.5 and not GPT five. 02:23.760 --> 02:26.320 So this is just something interesting right. 02:26.320 --> 02:28.600 So let's analyze this LLM code. 02:28.680 --> 02:31.320 So here we can see the input to the LLM call. 02:31.320 --> 02:34.760 And we sent a message of what is the weather in Tokyo. 02:34.960 --> 02:37.000 Now when long chain made the LLM call. 02:37.000 --> 02:43.280 It also sent to the LLM information about which tools does it has access to. 02:43.320 --> 02:44.600 And it can call. 02:44.600 --> 02:49.360 And if you open this model you'll see all the information about the tool that we defined earlier. 02:49.760 --> 02:58.280 Now if you've noticed my terminology, I said that we send to the LLM the list of tools and I said LLM. 02:58.280 --> 03:06.020 And I did not say agent because right now we are in the granularity of the LM with the LM call here. 03:06.020 --> 03:11.300 So we equip it now our LM with a tool that it can choose. 03:11.660 --> 03:20.020 And with the query of what is the weather in Tokyo, it decided that it needs to invoke the search tool 03:20.180 --> 03:22.940 with the query weather in Tokyo. 03:23.100 --> 03:26.300 And we have here some kind of tool calling ID here. 03:26.620 --> 03:30.060 So notice here the LM response is not. 03:30.060 --> 03:36.020 The invocation of this tool is simply which tool to call and with which arguments here. 03:36.180 --> 03:39.100 And it's going to do so with function calling. 03:39.260 --> 03:41.500 And how does this work under the hood? 03:41.540 --> 03:43.620 We are soon in this section. 03:43.620 --> 03:46.460 And in the section later we're going to figure out. 03:46.500 --> 03:55.260 Alright, so after the LM decided that it needs to call the search tool link chain then went and invoked 03:55.260 --> 03:58.900 the search tool with the query Weather in Tokyo. 03:58.900 --> 04:00.660 And this we can see right over here. 04:00.860 --> 04:07.800 And we can see that the response of these tools execution is Tokyo, weather is sunny and this is the 04:07.800 --> 04:11.360 static string that we defined in our Python function earlier. 04:11.640 --> 04:17.480 So we went and it ran this function with the argument of weather in Tokyo. 04:17.520 --> 04:20.480 So let's recap what happened up until now. 04:20.880 --> 04:26.520 So the LM decided which tool to call with which arguments. 04:27.200 --> 04:33.040 Then link chain went in ran this function with those arguments. 04:33.280 --> 04:36.000 So something very interesting is happening right now. 04:36.000 --> 04:38.680 And let me give you a quick preview of what's to come. 04:38.960 --> 04:45.840 So we have here a reasoning engine that decides which tool to call and with which arguments. 04:46.080 --> 04:53.560 And we have here now an agent execution runtime which actually goes in execute and run those tools and 04:53.560 --> 04:54.720 gets the output. 04:55.400 --> 05:02.840 Now link chain is going to be making another LM call, but this time it's going to contain the input. 05:02.880 --> 05:09.340 It's going to contain the LM decision to call the search tool, and it's going to contain the answer 05:09.340 --> 05:11.180 after executing the search tool. 05:11.340 --> 05:18.340 So the LLM should have a very easy time now generating the answer, because it has all the information 05:18.340 --> 05:19.060 it needs. 05:19.060 --> 05:25.020 And we can see here that the answer is that the weather in Tokyo is currently sunny right now. 05:25.060 --> 05:30.740 Remember that I told you that now all of those messages are going to be making more sense. 05:30.980 --> 05:33.100 The first message was the input. 05:33.420 --> 05:40.300 The second message was the AI's decision to call the search tool, and with which arguments link chain 05:40.300 --> 05:42.260 went and executed that tool. 05:42.260 --> 05:48.500 And then the result of this tool is structured here in this tool message object. 05:48.500 --> 05:53.860 And the tool message is simply a data structure to represent a result of a tool execution. 05:53.860 --> 05:56.820 And in this case the result of executing the search tool. 05:56.980 --> 06:05.140 And finally link chain went and made another LLM call with the input with the decision of the LLM to 06:05.140 --> 06:09.320 execute the search tool with the search tool a result. 06:09.320 --> 06:10.960 So the tool execution results. 06:10.960 --> 06:14.960 And then the final answer is that the weather in Tokyo is sunny. 06:15.080 --> 06:22.960 And also notice that the final LM call did not choose to execute a tool, but it chose to return an 06:22.960 --> 06:25.880 answer because it had all the information it needed. 06:26.320 --> 06:35.080 All right, let me go back to the code now and let me change the model to be GPT five and let me run 06:35.080 --> 06:35.720 it again. 06:36.280 --> 06:44.040 Now notice this time it's going to run a bit slower, because GPT five is a slower model than GPT 3.5 06:44.080 --> 06:44.560 turbo. 06:48.640 --> 06:50.080 And we got a result. 06:50.080 --> 06:54.400 Let me go now and show you the trace now of the latest run. 06:54.800 --> 06:56.440 And it's very similar. 06:56.440 --> 07:00.920 Just notice here there is a difference because the model name now is going to be GPT five. 07:01.240 --> 07:03.480 And let's go now back to the code. 07:03.480 --> 07:06.920 And let's go and add more meat here. 07:06.960 --> 07:08.760 Now to this search agent.