WEBVTT 00:00.360 --> 00:00.840 All right. 00:00.840 --> 00:08.560 Let's go to notes and let's go and import from env the load env function. 00:08.800 --> 00:12.760 And now let's go and import the primitives from graph. 00:13.240 --> 00:16.440 So we want to import from graph. 00:16.760 --> 00:17.400 Graph. 00:17.680 --> 00:20.800 We want to import the messages state. 00:21.160 --> 00:27.760 And this is a simple object which has the dictionary of the key messages. 00:28.000 --> 00:32.000 And that messages is going to be a list of messages. 00:32.280 --> 00:37.520 So this is going to keep track on the state on all of the messages back in between our agent. 00:37.520 --> 00:40.120 So the human message and the AI message. 00:40.480 --> 00:40.920 All right. 00:40.920 --> 00:47.200 So now we want to import from graph prebuilt a prebuilt node which is called tool node. 00:47.400 --> 00:49.040 And this is a node. 00:49.040 --> 00:54.320 And I elaborate on this in the graph course which is going to execute tools. 00:54.320 --> 00:58.560 So it's going to check the last message between the agent and human. 00:58.840 --> 01:05.690 And if that last message is an I message that has a valid tool call, it's going to go and execute that 01:05.690 --> 01:10.210 tool code, assuming that tools is initiated with the tool node object. 01:11.850 --> 01:21.210 So if the agent decides to execute a search or decides to run the triple function, then those are going 01:21.210 --> 01:23.290 to run inside this tool node. 01:23.290 --> 01:26.010 And it will all be clear when we'll use it in action. 01:26.490 --> 01:26.970 All right. 01:26.970 --> 01:34.450 So we want to import from react the LM with the tool calls that we binded to and the list of tools that 01:34.450 --> 01:36.850 we defined in the react py file. 01:37.690 --> 01:38.130 All right. 01:38.130 --> 01:40.210 Let's load the environment variables. 01:40.450 --> 01:45.530 And let's implement now the reasoning engine of our agent. 01:46.250 --> 01:51.410 So basically it's going to be a node where the agent is going to receive our message. 01:51.530 --> 01:54.130 And it's going to decide whether it can answer it. 01:54.130 --> 01:56.010 Or it needs to invoke a tool. 01:56.010 --> 01:59.210 And it's going to arrange all the arguments for that too. 01:59.650 --> 02:02.050 So for that, we'll define a system message. 02:02.370 --> 02:09.330 And here we'll give the very generic system message of you are a helpful assistant with access to tools 02:09.650 --> 02:12.010 and that you can use to answer the question. 02:12.250 --> 02:17.930 And now let's go and define the first node which is going to be the agent reasoning node. 02:18.250 --> 02:23.650 So it's going to receive the state which is a message state which is a dictionary that has the key of 02:23.650 --> 02:24.330 messages. 02:24.850 --> 02:31.970 And here all that we're going to do is simply make an LM call with the user input. 02:32.170 --> 02:38.290 And the LM call is going to do all of the heavy lifting because the LM is binded with the tools. 02:38.290 --> 02:40.570 And we're going to leverage function calling here. 02:40.890 --> 02:45.090 So we're going to invoke the LM and the invocation. 02:45.090 --> 02:49.810 We're simply going to put here the role of a system. 02:49.810 --> 02:51.690 We're going to put the system message. 02:51.850 --> 02:55.210 And then we're going to put all of the messages that we have. 02:55.450 --> 02:59.090 So in the first iteration, we're going to have only the human message. 02:59.330 --> 03:05.850 In the second iteration, we're going to have also the human message and the response and so on and 03:05.850 --> 03:06.370 so on. 03:06.490 --> 03:10.170 So all the information is going to be in the messages state over here. 03:10.890 --> 03:14.690 So I'm going to access the messages key in the state. 03:14.930 --> 03:18.610 And right now I'm accessing it as a Python property. 03:18.810 --> 03:24.930 But let me go and change it soon to access a dictionary key instead just to be on the safe side. 03:25.530 --> 03:28.410 And this is going to send the human message. 03:28.410 --> 03:30.810 It's going to send it back to the LM. 03:30.810 --> 03:37.930 And then the LM decides whether we need to use a function call and execute a function, or it decides 03:37.930 --> 03:39.570 that it can output the answer. 03:40.290 --> 03:43.490 So after the node runs, we need to update our state. 03:43.690 --> 03:48.930 And if we'll return a dictionary with the key of messages and return here a list. 03:48.930 --> 03:52.130 And that list is going to have the response from the LM. 03:52.370 --> 03:59.620 So this is going to be the AI message here then is going to append this value to our state in the messages 03:59.660 --> 03:59.980 key. 04:00.340 --> 04:03.300 So the key has all the messages so far. 04:03.500 --> 04:05.260 And this node is going to update it. 04:05.260 --> 04:07.300 And it's going to add another message here. 04:07.580 --> 04:10.060 So we're done with the reasoning node. 04:10.060 --> 04:14.980 And we also need to define the node with the relevant tools. 04:15.180 --> 04:18.500 So this is going to help Landgraf execute the tools that we need. 04:19.020 --> 04:22.060 So let me define a variable called tool node. 04:22.260 --> 04:26.780 And it's going to be object of the class tool node from Landgraf prebuilt. 04:26.780 --> 04:32.900 And we're going to provide it with the tools that we defined the search and the multiply tools which 04:32.900 --> 04:35.060 are both link chain tools. 04:35.500 --> 04:40.220 So this node over here does a lot of heavy lifting. 04:40.420 --> 04:44.460 It can run tools in parallel, can run them in streaming mode. 04:44.500 --> 04:46.860 And it has a lot of cool functionality. 04:47.020 --> 04:49.820 But right now we're going to use the vanilla version. 04:49.820 --> 04:54.460 And by the way, I go deep into the implementations of this in my course. 04:54.660 --> 05:01.300 And before we had this object, it was actually a lot of boilerplate code to write just to run the tools 05:01.300 --> 05:02.140 in our graph. 05:02.740 --> 05:06.260 All right, let me go and commit now all the changes. 05:06.420 --> 05:11.700 And let me add this notepad file and I'll call this commit notes. 05:12.340 --> 05:14.100 Let me push it to the repo. 05:14.700 --> 05:17.380 And if you want you can go to the repository. 05:17.420 --> 05:18.980 Go to the notes commit. 05:19.220 --> 05:23.820 And here you have the code that we just wrote right. 05:23.860 --> 05:27.260 So we finished implementing the graph nodes. 05:27.540 --> 05:33.060 And we didn't implement a lot because we used a lot of language primitives, which saves us a lot of 05:33.100 --> 05:34.700 boilerplate code to run. 05:34.940 --> 05:39.180 And we are using the function calling features of LM methods. 05:39.180 --> 05:41.900 So the LM is going to do the reasoning for us. 05:41.900 --> 05:43.740 We don't really need to do anything here. 05:43.740 --> 05:45.380 We don't need to write a special prompt. 05:45.380 --> 05:47.300 We don't need to do any of this. 05:47.340 --> 05:51.780 And in the next video we're going to stitch everything together and implement our graph.