WEBVTT 00:00.280 --> 00:00.880 Hey there. 00:00.920 --> 00:01.760 Eden here. 00:01.760 --> 00:06.640 And in the next two videos, we're going to be implementing our tool executor node. 00:07.080 --> 00:13.560 And this node is going to take as an input the I message, which is going to have the information of 00:13.560 --> 00:15.040 the wanted search queries. 00:15.440 --> 00:22.960 And it's going to run Tavileh to get us real time results and real time information from the web. 00:24.320 --> 00:29.040 Through this video, we'll have all the components for our graph ready and we'll build our graph. 00:29.360 --> 00:31.160 So let's go to the code. 00:31.840 --> 00:32.560 Alrighty. 00:32.560 --> 00:34.720 So let's go and create a new file. 00:35.040 --> 00:38.040 And I want to call it tool executor. 00:39.080 --> 00:43.320 Let's start by importing low dotenv and load the environment variables. 00:43.800 --> 00:47.720 And now I want to import from link chain Tavileh. 00:47.840 --> 00:54.040 I want to import the Tavileh search tool and make sure you have Tavileh installed. 00:54.040 --> 00:56.400 So you can do that by writing poetry. 00:56.440 --> 00:58.000 Add link chain Tavileh. 00:58.120 --> 01:01.720 And you want to make sure that you have the API key. 01:01.720 --> 01:05.920 So let me go and take that real quick and I'll log in. 01:11.760 --> 01:16.760 And let me go and copy one API key and let me put it in my env file. 01:16.760 --> 01:20.000 And by the way, don't worry about me exposing my API keys. 01:20.040 --> 01:22.640 I revoked them after I finished filming this video. 01:22.960 --> 01:29.320 And we want now to import the structured tool class from Link Chain, which is a component that's going 01:29.320 --> 01:34.240 to allow us to convert a Python function into a tool that can be used by Llms. 01:34.480 --> 01:40.480 So it's going to take that function and provide to the LLM a structured schema for the function, which 01:40.480 --> 01:43.080 will help the LLM understand how to use this tool. 01:43.760 --> 01:48.760 And then we're going to import the tool node class from Landgraaf. 01:49.240 --> 01:53.560 And this is a very cool class that saves us tons of work. 01:53.600 --> 01:56.760 And it's a node in the graph that we can invoke. 01:57.160 --> 02:01.360 And it's going to look in the state for the messages key. 02:01.600 --> 02:07.930 It's going to check the last message, and it's then going to see if there's any tool calls that were 02:07.930 --> 02:09.450 decided by the alarm. 02:09.570 --> 02:13.650 And if there are, it's going to execute those tools for us. 02:13.650 --> 02:15.770 And it can do this even in parallel. 02:16.290 --> 02:18.250 So this saves us tons of work. 02:18.250 --> 02:21.410 And trust me, before that we would need to do everything ourselves. 02:21.410 --> 02:25.010 And I actually did that in the original version of the course. 02:25.050 --> 02:30.090 And I'll even leave this implementation as optional in case you want to check out all the heavy lifting 02:30.090 --> 02:32.530 that tool node performs for us anyways. 02:32.530 --> 02:37.690 So this is going to be the node that is going to execute the tools, and we're going to instantiate 02:37.690 --> 02:39.970 it with the tools that it needs to run it. 02:39.970 --> 02:41.530 We're going to see it in this video. 02:41.530 --> 02:43.170 So it's going to be very clear. 02:44.050 --> 02:44.610 Cool. 02:44.650 --> 02:49.730 So let me just import the answer question and the revised answer classes. 02:49.970 --> 02:54.210 Let's go and initialize our search tool. 02:54.770 --> 02:58.610 So we'll start by creating an object of search. 02:58.970 --> 03:03.570 So we'll do that with max results equals to five simply to get five results. 03:03.930 --> 03:10.170 And it's going to give us back a link chain tool with the function of the search engine. 03:10.570 --> 03:14.970 And we don't want to use it as is, like we usually do. 03:15.010 --> 03:16.730 We want to do a cool trick here. 03:17.130 --> 03:23.930 So we want to take the original tool and its functionality, and we want to create from it two different 03:23.970 --> 03:24.490 tools. 03:24.530 --> 03:29.890 It's going to be two different tools with the same functionality of the search, but they're going to 03:29.890 --> 03:35.210 have different names because they serve different purposes in the application workflow. 03:35.610 --> 03:41.530 So we're going to have an answer question tool, which is going to be used during the initial research 03:41.530 --> 03:44.410 phase when the agent is first answering the question. 03:44.610 --> 03:50.450 And then we want to have this revised answer tool which is used during the revision phase when the agent 03:50.450 --> 03:53.250 is improving its answer based on the reflection. 03:53.490 --> 03:56.890 So both tools are going to run the search. 03:57.810 --> 04:02.610 And that's why we need those objects the answer question object and the revised answer object. 04:02.610 --> 04:06.410 Because we want to get their names in order to label those tools. 04:07.170 --> 04:14.210 But theoretically we can use one tool, but having separate names in two separate tools allows the system 04:14.210 --> 04:19.770 to clearly track which stage of the research process triggered the search initial research versus the 04:19.770 --> 04:20.930 revision research. 04:21.090 --> 04:24.610 So it's going to help us in debugging and evaluating the response. 04:24.890 --> 04:25.370 All right. 04:25.370 --> 04:28.130 So let me create a new function. 04:28.130 --> 04:34.770 And I'm going to call this function run queries which is going to receive as an input the search queries 04:34.770 --> 04:36.130 which is a list of strings. 04:36.530 --> 04:40.450 And in its description it's going to have run the generated queries. 04:40.730 --> 04:45.450 And let's also add the quarks here in case the LLM is going to fail some other values. 04:45.450 --> 04:47.490 So we won't get an error in that case. 04:47.890 --> 04:51.330 And the implementation here is going to be very straightforward. 04:51.330 --> 04:55.610 We're going to run the Tavileh tool with those queries. 04:55.610 --> 04:57.410 So we're going to iterate over them. 04:57.410 --> 05:02.610 And we're going to run the query with the batch function which is going to run them concurrently. 05:03.090 --> 05:03.530 All right. 05:03.530 --> 05:08.930 So now we want to create from this function we want to create two different tools with different names. 05:09.170 --> 05:14.090 So one tool is going to be answer a question and the other tool is going to be revised answer. 05:14.370 --> 05:19.090 And they're both going to run the search the run queries function. 05:19.090 --> 05:22.450 And the structured tool class is going to help us achieve that. 05:23.250 --> 05:26.290 And they're all going to run in the tool mode. 05:26.290 --> 05:29.690 So the tool mode is going to run two different tools. 05:29.930 --> 05:36.090 And one tool is going to be the search tool that originated from the first creation of the research. 05:36.250 --> 05:41.130 And the other tool is going to be a search query that originated from the reflection revision. 05:41.130 --> 05:42.970 And that's the cool trick over here. 05:43.090 --> 05:44.250 So let's go and do that. 05:44.250 --> 05:46.610 Let's go create an object of tool node. 05:46.810 --> 05:51.450 And in order to do that we'll need to supply the list of tools that it's going to run. 05:51.610 --> 05:53.050 So let's create the first tool. 05:53.050 --> 05:55.410 So we'll take the structured tool class. 05:55.410 --> 06:01.450 And it's going to have a method from function which is going to receive a function and convert it into 06:01.450 --> 06:05.450 a tool with schema and description and all of that that we saw before. 06:05.650 --> 06:08.460 And another argument is going to be the name of the tool. 06:08.500 --> 06:10.820 So here the name is going to be the name of the class. 06:11.380 --> 06:16.260 And we want to do another tool which is going to run the same function. 06:16.260 --> 06:20.620 But it's going to have a different name of the revised answer class. 06:20.780 --> 06:21.780 And that's it. 06:22.020 --> 06:22.900 We are done. 06:22.940 --> 06:26.140 Here we have all the moving parts in order to create the graph. 06:26.340 --> 06:30.300 Now I remind you this tool is going to examine the state. 06:30.300 --> 06:32.380 It's going to check the last message. 06:32.380 --> 06:35.860 And if there is a tool call it's going to execute the relevant tool call. 06:37.460 --> 06:44.580 All right let me run some formatting over here and let me go and add everything we did right now. 06:44.580 --> 06:45.620 And let's commit it. 06:46.220 --> 06:52.540 So I'm going to commit this to the branch of slash projects slash reflection agent. 06:54.900 --> 07:01.260 And if you want to see the code you can select in the repo a branch of project slash reflection agent. 07:01.700 --> 07:06.780 And you can check out the commit here, which has all the code that we did in this video. 07:06.780 --> 07:09.620 And it's going to be linked in the videos resources.