WEBVTT 00:00.080 --> 00:05.880 Let's start now by defining the list of tools we're going to give our AI agent. 00:05.960 --> 00:07.800 So it's going to be a search agent. 00:07.800 --> 00:13.880 And it's going to be a list containing only one tool, which is going to be the search tool. 00:14.280 --> 00:19.440 Now I want to discuss what is this search tool object. 00:20.000 --> 00:23.360 Now in the previous video I showed you the search API. 00:23.560 --> 00:26.320 So how to use it in the AI playground. 00:26.360 --> 00:29.200 Now they have an SDK. 00:29.360 --> 00:31.920 We can programmatically use their API. 00:32.120 --> 00:35.440 All we need to do is to generate an API key and to use their SDK. 00:36.320 --> 00:43.960 Now if we want to use an LLM as a reasoning engine and to kind of get this behavior where the LLM has 00:44.000 --> 00:52.960 access to this API, to this SDK, then we need somehow to communicate to the LLM that it has in its 00:52.960 --> 00:55.880 disposal this kind of capability. 00:56.000 --> 01:00.120 So the ability to use a API right. 01:00.160 --> 01:05.880 So we need to tell the LLM maybe what is this tool name to use it. 01:05.920 --> 01:07.840 What arguments does it received? 01:07.880 --> 01:10.240 How does the output of this tool look like. 01:10.240 --> 01:12.960 So we need to give all this information to the LM. 01:13.080 --> 01:16.760 And this is exactly why LinkedIn created the tool class. 01:17.240 --> 01:24.560 Now the LinkedIn tool is a component that will allow the LMS to interact with external utilities or 01:24.600 --> 01:29.160 APIs by generating structured calls to those tools. 01:29.280 --> 01:35.920 And they're defined with the schema that is going to specify the name, the expected arguments, and 01:35.920 --> 01:40.880 it will allow the LM to generate calls with the structured inputs. 01:41.280 --> 01:45.200 Now the model does not execute the tool itself. 01:45.600 --> 01:49.320 Instead it only produces the arguments for the tool call. 01:49.480 --> 01:57.720 And the system running length chain will execute the tool and propagate back the output to the model. 01:57.720 --> 01:59.040 And this is very important. 01:59.040 --> 02:02.320 And this is what's going to happen actually in the agent executor. 02:02.360 --> 02:03.840 This is this is a quick spoiler. 02:04.040 --> 02:11.930 Now we can convert any Python function to a tool using the tool decorator which then the model can call 02:11.930 --> 02:18.130 by generating the appropriate arguments for the tool call, and it's going to include the tool name, 02:18.410 --> 02:20.050 the dictionary of arguments. 02:20.050 --> 02:27.610 And it's going to enable structured and reliable interaction between the LM and the external API. 02:28.170 --> 02:33.290 Now how will the LM know how to use this tool and when to call it. 02:33.290 --> 02:38.410 And usually the tool description will tell the LM what this tool is doing. 02:38.410 --> 02:41.130 So it will know when to use it. 02:41.130 --> 02:46.250 And in that LM call will also provide what arguments that this tool receive. 02:46.450 --> 02:47.890 So what are the names. 02:47.890 --> 02:48.930 What are the types. 02:49.050 --> 02:54.090 So eventually the LM has all the information of what it has in its disposal. 02:54.130 --> 02:59.650 What are the tools when to invoke them and what are they expecting and what are they outputting. 02:59.970 --> 03:00.490 All right. 03:00.530 --> 03:02.170 So back now to the code. 03:02.450 --> 03:07.090 The search object is from the link chain to build integration. 03:07.250 --> 03:16.250 And this is where the team, they took their SDK and they wrapped it as a link chain tool and basically 03:16.250 --> 03:20.730 created an object that is going to hold the description of this function. 03:20.730 --> 03:26.370 It's going to hold the arguments that it needs and the return value of this function. 03:26.690 --> 03:29.490 And that's what's going to be communicated to the LM. 03:29.930 --> 03:30.290 All right. 03:30.290 --> 03:32.610 So let me now run this in debug mode. 03:32.610 --> 03:36.570 And I want to examine the tools variable here. 03:36.570 --> 03:39.130 And I want to show you this tool object. 03:39.730 --> 03:43.130 So let us wait until we hit that breakpoint. 03:43.650 --> 03:49.490 And when we hit it let's open now the local variables here sorry the global variables. 03:49.930 --> 03:53.810 And we can see now we have the tools list here. 03:53.810 --> 03:58.290 It has only one element and it's called search here. 03:58.770 --> 04:05.610 Now this object here if we'll check out its attributes we can see we have an attribute called description. 04:05.650 --> 04:11.770 And here we can see description is a search engine optimized for comprehensive, accurate and trusted. 04:11.930 --> 04:14.010 And this text goes on and on. 04:14.010 --> 04:16.770 And this is the description of the search tool. 04:16.970 --> 04:20.860 So this is one thing that's going to be propagated to the LM column. 04:20.860 --> 04:23.740 So the LM would have this description of this video. 04:24.020 --> 04:31.260 Now we need to tell LM what kind of arguments these two receives so it can create the tool call. 04:31.300 --> 04:37.780 So here we can see we have here a dictionary where the keys are going to be the parameters that the 04:39.460 --> 04:40.860 search is going to receive. 04:40.860 --> 04:47.300 So we can see we have here query include domains exclude domains and other business logic that really 04:47.300 --> 04:48.180 may require. 04:48.580 --> 04:54.980 Now it's super important to note that the description and the arguments description, it's super important 04:54.980 --> 04:56.580 for us to be explicit. 04:56.580 --> 05:02.660 So if this is going to be ambiguous, the LM is going to have a hard time to crafting with the tool. 05:03.220 --> 05:09.180 And this is, by the way, one of the reasons why I used the link chain integration instead of simply 05:09.180 --> 05:15.940 importing the search client with the SDK and wrapping it as a selection tool, simply because the team 05:15.940 --> 05:22.180 knows their product the best and they can write the description for their arguments and for their function. 05:22.420 --> 05:25.140 A much better than I can write. 05:25.140 --> 05:29.500 So here you can see a bunch of other fields and that are the arguments. 05:29.700 --> 05:32.180 And, uh, yeah, this is pretty much it. 05:32.220 --> 05:33.140 Let me go show you. 05:33.140 --> 05:36.580 Now, um, this is the rest of the tool description here. 05:36.820 --> 05:38.540 Let me go show you the tool name. 05:38.540 --> 05:41.860 The tool name is going to be right over here. 05:41.860 --> 05:45.580 So we can see tool name is underscore search. 05:45.940 --> 05:55.540 So just to conclude we can think about a tool as a wrapper of a Python function that's going to have 05:55.540 --> 06:01.540 the fields of name description arguments and maybe some extra metadata. 06:01.580 --> 06:03.300 But that's pretty much it right. 06:03.340 --> 06:07.020 So and this is what's going to be propagated eventually to the LLM. 06:07.420 --> 06:09.420 And we're going to see this in action very soon. 06:10.100 --> 06:10.540 All right. 06:10.540 --> 06:17.220 So let me now go back to the code and let's go and examine the actual implementation of the search class 06:17.220 --> 06:17.540 here. 06:17.900 --> 06:21.260 And we can see that it inherits from base tool. 06:21.260 --> 06:23.020 And this is a link chain object. 06:23.020 --> 06:28.870 And this is going to have the interface of a having the attributes of name description arguments that 06:28.870 --> 06:30.030 we talked about earlier. 06:30.910 --> 06:37.470 And here in this description we can see some examples of how we can instantiate this object here, which 06:37.470 --> 06:38.870 is going to be a link chain tool. 06:38.990 --> 06:42.110 So we can also fixate some of the arguments. 06:42.110 --> 06:45.870 So there will be fixed when calling a search API. 06:47.190 --> 06:49.510 And here we can see for example the name. 06:49.510 --> 06:53.030 We can see the description and we can see the schema. 06:53.030 --> 06:56.350 So it's going to describe what kind of arguments it's going to receive. 06:56.630 --> 07:01.630 And some other descriptions of the other arguments that it's going to receive. 07:02.070 --> 07:09.590 Now when it's going to be executed then it's going to execute this underscore run function. 07:09.990 --> 07:15.470 And it's going to use the SDK to make the API call to the service. 07:15.710 --> 07:19.990 I remind you, the LM is not going to make this call to the API. 07:20.350 --> 07:23.790 It's going to only reason and decide that we need to call it. 07:23.790 --> 07:29.950 So the actual call and execution of this method is going to happen in our link chain application. 07:29.950 --> 07:33.550 So that's going to be in the agent executor which we're going to check out later.