WEBVTT 00:01.040 --> 00:05.720 So I want to add here some real world searching over the internet. 00:05.960 --> 00:09.640 So I'm going to import from Tavileh, the Tavileh client. 00:11.240 --> 00:14.920 So let me now go and initialize this Tavileh client. 00:15.000 --> 00:17.960 So I'm simply going to create a variable called Tavileh. 00:18.160 --> 00:21.200 And I'm going to initialize the client. 00:21.400 --> 00:27.360 And this client when it's going to initialize and it's going to be looking for an environment variable 00:27.360 --> 00:28.920 called Tavileh API key. 00:29.080 --> 00:33.000 So this is where we put in our dot env file the API key. 00:33.040 --> 00:35.200 So everything should be working right now. 00:35.720 --> 00:42.360 So let me now instead of returning the Tokyo weather is sunny, let's go and run the actual search query 00:42.360 --> 00:43.960 and let's return the results. 00:43.960 --> 00:45.600 We'll be doing it with Tavileh. 00:45.840 --> 00:50.480 Search and search need to receive an argument called query and query. 00:50.520 --> 00:55.440 Here is going to be equal to our query, which is the input to our search function. 00:57.720 --> 00:58.400 Amazing. 00:58.560 --> 01:00.080 So let's go and run it now. 01:00.080 --> 01:03.540 And it's going to have now a real searching capability. 01:04.700 --> 01:06.660 So we'll just wait for it to finish. 01:07.860 --> 01:10.460 We can see now it's searching over the web. 01:11.660 --> 01:17.540 And once we have an answer, we'll head up to Lenox-smith to review again the traces. 01:18.460 --> 01:21.820 So everything now should be very similar. 01:21.820 --> 01:30.100 But if we go now to the search execution here, we can see that instead of returning a static string 01:30.100 --> 01:36.260 here, we're actually going to output some real information of the weather in Tokyo. 01:36.260 --> 01:39.820 And this is an example of a search made. 01:40.060 --> 01:43.020 And it got the information from the weather API. 01:43.820 --> 01:46.660 And here you can see all the information of the weather. 01:46.660 --> 01:51.180 So it also provided us with other results and other answers. 01:51.180 --> 01:54.460 So it provided us with the top results here. 01:54.460 --> 01:59.460 And if we want we can control this parameter of how many results you get. 01:59.740 --> 02:06.600 And if I go to the last LM response here, we can see now we get a different answer, which is a much 02:06.600 --> 02:13.600 more elaborate answer here, which is grounded on real time information, which we got through research. 02:14.040 --> 02:14.680 All right. 02:14.680 --> 02:21.440 So now let's go and let me change now the query to be what I showed you in the demo. 02:21.680 --> 02:28.720 So now let's have a look and let's see if we can get some job openings for link chain professionals. 02:29.080 --> 02:31.040 So let me go back to the code here. 02:31.240 --> 02:38.160 And here in the message I'm going to tell the agent that I want to search for three job postings for 02:38.160 --> 02:46.120 an AI engineer in chain in the Bay area in LinkedIn, and to list all of their details here. 02:46.520 --> 02:51.280 And don't worry about this prompt if you'll check out the repo in the videos resources, you can simply 02:51.280 --> 02:52.280 copy and paste it. 02:52.680 --> 02:56.080 Anyways, let me go now and let's go and run this. 03:01.760 --> 03:06.440 And we can learn from this log over here that we are running now. 03:06.480 --> 03:09.020 LinkedIn is running the search tool here. 03:09.020 --> 03:12.980 It's running it five times with five different queries. 03:12.980 --> 03:21.820 So the first query is to look inside LinkedIn.com slash jobs looking for LinkedIn, Bay area, San Francisco 03:21.860 --> 03:23.540 San Jose AI engineer. 03:23.540 --> 03:25.700 So we have multiple search queries here. 03:25.900 --> 03:26.420 All right. 03:26.420 --> 03:28.340 We got now a result. 03:28.380 --> 03:30.980 We can see now here in that it finished. 03:30.980 --> 03:33.940 We can see here a it's all printed out. 03:33.980 --> 03:34.860 Not very nicely. 03:34.860 --> 03:37.100 So let's go and open it. 03:37.100 --> 03:38.540 And let's check out the trace. 03:39.420 --> 03:43.060 And now we can see our input our query. 03:43.340 --> 03:51.340 And notice that in the output of the LLM in the AI response here we get multiple tool calls. 03:51.340 --> 03:54.540 So the AI decides to call multiple tool calls. 03:54.540 --> 03:57.260 And it's going to execute them in parallel. 03:57.260 --> 04:03.220 So this is because function calling in GPT five supports multiple function calling and will be elaborating 04:03.220 --> 04:04.340 it later in the course. 04:04.620 --> 04:07.460 This is an example of one tool call. 04:07.780 --> 04:11.440 This is an example of another tool call. 04:12.000 --> 04:14.600 And we have here a five tool course. 04:14.840 --> 04:21.480 And the final request to the MLM contains all the tool calls and all their answers. 04:21.680 --> 04:24.680 And let me go here and down. 04:24.680 --> 04:27.640 And let's see the final answer of the LM. 04:27.840 --> 04:29.320 And here we got the answer. 04:29.320 --> 04:35.600 So and we got here all of the job postings on LinkedIn and the URL to LinkedIn. 04:37.080 --> 04:40.280 Notice that some of them are no longer accepting candidates. 04:40.280 --> 04:43.680 But this is the final response here of the LM. 04:44.320 --> 04:44.680 All right. 04:44.680 --> 04:46.360 I want to go back to the code now. 04:46.360 --> 04:51.360 And I want to discuss now a best practice when using LinkedIn. 04:51.840 --> 04:59.800 So in our example we wrote a custom tool and used the SDK inside that tool. 05:00.480 --> 05:07.920 However, this will require us as the developers to know every bit and byte and to know how the SDK 05:07.960 --> 05:08.680 is working. 05:09.160 --> 05:19.490 And in fact I do not know exactly all the internals of the SDK so heavily because it has a LinkedIn 05:19.490 --> 05:20.290 integration. 05:20.450 --> 05:30.050 The team, they wrote a package link chain which takes the SDK and wraps it as a chain tool. 05:30.370 --> 05:40.050 And you can trust you should trust the vendor, in this case the team, to write a much better description, 05:40.050 --> 05:48.090 much better arguments, and they can do a far better job writing the link chain tools than we do here. 05:48.330 --> 05:55.010 So what I'm going to do right now, I'm going to remove the custom tool, and I'm going to use the built 05:55.010 --> 05:59.450 in Tavileh search tool that the Tavileh team wrote. 06:00.250 --> 06:04.650 So let me add link chain Tavileh just to make sure I have this package installed. 06:05.530 --> 06:08.690 And let me now remove all of these search tool. 06:08.730 --> 06:09.930 We do not need it. 06:10.330 --> 06:11.890 Also the Tavileh client. 06:12.170 --> 06:20.310 And let's instead of this import let's right from link chain underscore tavileh let's go and import. 06:20.710 --> 06:25.470 It's called Tavileh search and Tavileh search. 06:25.510 --> 06:29.310 Let's go and check out the implementation is a base tool here. 06:29.550 --> 06:30.950 So it's a link chain tool. 06:31.270 --> 06:33.430 So let me go back now. 06:33.430 --> 06:38.630 And because this is already a tool all I need to do is to initialize this object. 06:38.790 --> 06:41.590 I won't be sending it any custom arguments. 06:42.550 --> 06:45.190 So let's go now and run it again. 06:45.470 --> 06:49.830 And I'm going to fast forward this because it may take some time. 06:50.430 --> 06:55.350 And let me go back now to link and let's check out this trace. 06:56.630 --> 07:02.950 So notice here that the name of the tool here is called underscore search. 07:02.950 --> 07:05.030 And this is how Tavileh named their tool. 07:05.190 --> 07:07.030 And our name was search. 07:07.030 --> 07:09.230 And you can compare it with the previous trace. 07:09.830 --> 07:13.750 And let's go and check out an example of an execution here. 07:14.070 --> 07:21.130 And we can see here we have the query like we had before, but we have here some extra arguments that 07:21.130 --> 07:24.130 the LM decided to make with the tool call. 07:24.170 --> 07:27.850 So those are arguments that I didn't know about. 07:27.890 --> 07:31.010 So they are include domains linkedin.com. 07:31.010 --> 07:35.970 So this is probably going to say to to really to only get results from LinkedIn.com. 07:36.290 --> 07:42.090 And it has an extra argument which is called search depth which is advanced. 07:42.210 --> 07:46.970 And I will be explaining this further in the course because we'll be seeing much more examples with 07:46.970 --> 07:47.490 David. 07:47.530 --> 07:49.650 What is this argument here? 07:49.850 --> 07:55.890 But the point I want to show you here is that we are getting a much more elaborate tool calls here. 07:56.090 --> 08:03.570 And while the answer is very similar to what we executed before, however, it's a bit more accurate 08:03.610 --> 08:10.570 now because the agent was able to make much more specific tool calls, which should ground the answer 08:10.570 --> 08:12.370 in better sources. 08:12.410 --> 08:19.150 All right, let me go now and let me bring back the implementation with our custom tool here, and let 08:19.150 --> 08:24.790 me push everything to the repository and I will be linking everything in the videos resources. 08:25.510 --> 08:26.070 All right. 08:26.070 --> 08:29.510 So let me go now and add everything into git. 08:29.510 --> 08:33.070 And let me now see all the files here. 08:33.470 --> 08:35.350 So let me now commit everything. 08:35.390 --> 08:41.590 Let me right in the message intro to search agents and let me push it. 08:42.230 --> 08:42.550 Sorry. 08:42.550 --> 08:44.870 Let me set the upstream and let me push it again. 08:45.790 --> 08:49.990 All right let's go back to the repo and you can find everything. 08:49.990 --> 08:55.950 If you go to the branches it should be called project slash search agent. 08:56.510 --> 08:58.390 You can see it right over here. 08:59.430 --> 09:02.750 And let me now show you here the commit. 09:09.430 --> 09:12.750 And let's go check out the main file. 09:15.510 --> 09:18.550 And everything you can see it's right over here.