WEBVTT 00:00.440 --> 00:01.240 Alrighty. 00:01.280 --> 00:04.800 In this video, we'll implement the web searching node. 00:05.440 --> 00:12.160 We'll be leveraging the search engine, which is highly optimized to downstream search results, into 00:12.200 --> 00:13.920 an LLM based application. 00:14.520 --> 00:20.560 And to use Tavileh, we need to make sure we have in our dot env file the Tavileh API key. 00:21.800 --> 00:25.880 Let's go to the code and let's add a new file under node. 00:25.880 --> 00:28.240 And we want to call it web search. 00:29.520 --> 00:33.800 And here we want to start with the imports to import typing. 00:34.080 --> 00:41.680 We want to import the Lang document class because we want to convert the search results into a Lang 00:41.960 --> 00:42.760 document. 00:43.080 --> 00:51.400 And finally we want to import the search class, which is a link chain tool that runs the search engine 00:51.400 --> 00:52.120 on the queries. 00:52.120 --> 00:53.080 We will provide it. 00:54.000 --> 00:54.760 All right. 00:54.760 --> 00:58.240 So let's now also import the graph state. 00:58.760 --> 01:05.720 And finally, we want to create an object of the Search results tool. 01:06.280 --> 01:09.400 Let's initialize it with max results equals to three. 01:09.400 --> 01:12.200 So we'll get at most three results. 01:12.360 --> 01:14.320 And let's define our function. 01:14.320 --> 01:15.640 We'll call it web search. 01:15.640 --> 01:22.520 And it will receive the state and return a dictionary will print something just for debugging. 01:23.080 --> 01:27.680 And we'll extract the question and documents from the graph state. 01:27.720 --> 01:33.320 Now notice because we execute the web search node only after we grade the documents. 01:33.600 --> 01:35.840 So it's after we filter them out. 01:35.840 --> 01:39.480 So we're not supposed to have any non-relevant documents. 01:39.640 --> 01:45.400 So all of the documents we're going to have in the documents list are going to be relevant for our query. 01:45.960 --> 01:46.400 All right. 01:46.400 --> 01:48.440 We're going to debug this file soon. 01:48.440 --> 01:54.960 So I'm going to import the environment variables and make sure that I have my API key in my env file. 01:55.680 --> 01:58.360 Let's add some if name equals main. 01:59.160 --> 02:05.120 And we're going to plug in the question to the agent memory and the documents to be none. 02:05.120 --> 02:07.760 So we don't want to put anything at the moment. 02:07.760 --> 02:12.200 This is the scenario where we didn't find any relevant documents at all. 02:12.960 --> 02:13.720 All right. 02:13.720 --> 02:21.720 So now we want to invoke the search tool, and we want to invoke it with the query to be our question. 02:23.280 --> 02:24.600 So let's debug it. 02:24.600 --> 02:27.680 And we want to examine the results we get from Tavelli. 02:28.240 --> 02:34.800 So because we limited the max results to be three, we're supposed to have here only three search results. 02:34.800 --> 02:38.080 And we can see it right here we have a list of three elements. 02:38.280 --> 02:42.760 And every element is a dictionary with the content key and the URL key. 02:43.120 --> 02:49.240 And what we want to do is to take all the content from all the elements of this list, and to combine 02:49.240 --> 02:52.480 them into one document of length chain. 02:52.720 --> 02:54.640 So that's what we're going to do next. 02:54.880 --> 02:55.680 Alrighty. 02:56.040 --> 02:58.040 So let's go back to the code. 03:01.560 --> 03:05.120 And let's stop debugging and let's add it. 03:05.120 --> 03:08.320 So we want to iterate through every element on the list. 03:08.320 --> 03:17.000 And we want to use the join function with backslash n to take the elements in the key of content and 03:17.000 --> 03:19.080 to simply join everything together. 03:19.680 --> 03:22.080 So we get only one huge string. 03:22.320 --> 03:22.920 All right. 03:22.960 --> 03:24.640 Now let's run this in debug. 03:24.640 --> 03:27.160 And let's see that we get indeed this string. 03:30.560 --> 03:33.520 And I'm going now to evaluate this expression. 03:35.920 --> 03:38.480 And we can see we have one big string. 03:39.040 --> 03:40.920 Now we want to take this string. 03:40.920 --> 03:47.800 And from it we want to create a long chain document where its content is going to be this big string. 03:49.440 --> 03:52.680 So I want to create a new variable called web results. 03:52.680 --> 03:54.640 And it's going to be a document. 03:54.640 --> 03:58.760 And in the page content I'm going to give the joint result. 03:58.760 --> 04:04.680 And if we currently have documents in our state then those are relevant documents. 04:04.680 --> 04:10.400 So we want to append to this document, list the document that is containing the web search. 04:10.720 --> 04:16.600 And if not, then this means that we didn't find any document that's relevant, and we simply want to 04:16.640 --> 04:21.280 put in the documents a list with one element, which is the web results. 04:21.520 --> 04:26.240 And finally, we want to return and update the state of our graph execution. 04:26.480 --> 04:29.520 So in the documents key we're going to put the documents. 04:29.520 --> 04:32.600 And in the question we're simply going to keep the original question. 04:33.200 --> 04:35.520 And that's it for the web searching node. 04:35.760 --> 04:37.040 Very straightforward. 04:37.680 --> 04:43.000 And the last thing I want to do in this video is to simply rename the file name. 04:43.360 --> 04:46.040 So I will go and rename it. 04:46.040 --> 04:48.560 And I simply want to add here an underscore. 04:48.560 --> 04:50.880 So I'll call it web underscore search.