WEBVTT 00:00.080 --> 00:03.800 Alrighty, let's go now and let's run everything. 00:03.800 --> 00:10.840 And I want now to show you the logs, the outputs, and we'll have a better understanding of how everything 00:10.840 --> 00:13.080 here is working together and what we just did. 00:13.400 --> 00:16.240 Let me just go and click um continue here. 00:16.960 --> 00:17.600 All right. 00:17.640 --> 00:21.720 We can see now we're entering a new agent executor chain. 00:21.920 --> 00:29.480 And right now the prompt we wrote plugged in with all the tool descriptions of the search is all being 00:29.480 --> 00:31.440 bundled up with the prompt. 00:31.440 --> 00:33.360 And that's being sent to the LM. 00:33.400 --> 00:34.040 All right. 00:34.160 --> 00:37.680 And now the LM is going to return us an answer. 00:38.000 --> 00:39.840 And this we can see in green here. 00:40.040 --> 00:45.160 And the LM returned to find the job posting I will use the search tool. 00:45.200 --> 00:51.320 This is the tool name inputting a specific query to look for an AI engineer position involving link 00:51.320 --> 00:53.000 chain, blah blah blah blah. 00:53.200 --> 00:59.360 And then we have here action which is going to be search which is going to be the tool name. 00:59.560 --> 01:05.990 And then we have here the action input which is going to be the string of AI engineer LinkedIn job description. 01:06.030 --> 01:07.390 Bay Area LinkedIn. 01:07.430 --> 01:08.150 All right. 01:08.190 --> 01:11.670 So this right here is what the LLM returned. 01:11.910 --> 01:18.950 Now it returned us this kind of answer with this specific format of action and action input just because 01:18.950 --> 01:19.990 of the react prompt. 01:20.030 --> 01:23.190 And I will be elaborating on this in the rest of this course. 01:23.190 --> 01:25.630 But this is this kind of hack here, right. 01:25.670 --> 01:31.510 So the LLM is going to always respond this with these kinds of format where we have action, which is 01:31.510 --> 01:37.310 going to have the tool name and action input, which is going to be the input for the tool name. 01:37.670 --> 01:42.310 Now LinkedIn and this is the beautiful thing LinkedIn is going to take this for us. 01:42.670 --> 01:45.350 It's going to parse out those strings. 01:45.550 --> 01:49.790 And it's going to figure out that it has in its disposal the search. 01:49.910 --> 01:58.030 And it's going to invoke the function of the search, which we imported from the LinkedIn integration. 01:58.030 --> 02:01.670 And it's going to invoke this function with this input. 02:01.670 --> 02:05.790 And LinkedIn is going to do all this parsing and all this heavy lifting for. 02:05.790 --> 02:05.820 54. 02:06.020 --> 02:11.620 And I remind you, we had this output parser from before in the Create react agent. 02:11.620 --> 02:15.260 So that's part of the chain is going to do that for us right. 02:15.300 --> 02:21.260 So it's going to take this answer and it's going to extract from it the name of the tool that we need 02:21.300 --> 02:23.020 and the input for that tool. 02:23.300 --> 02:23.540 Right. 02:23.580 --> 02:27.860 So now link chain is going to execute the search tool. 02:27.900 --> 02:29.260 So this is what's going to happen. 02:29.260 --> 02:30.820 And it's going to get a result. 02:31.020 --> 02:32.860 And this is what we're seeing right now. 02:33.380 --> 02:37.540 And here we have the result of the tool execution. 02:37.540 --> 02:41.220 So link chain ran the tool and it got us the result. 02:41.580 --> 02:48.380 And in this result here we can see we have the query which is the same query from the action input. 02:48.660 --> 02:51.540 And then we have here the results. 02:51.580 --> 02:55.940 So here we have a list of URLs and the description of the search results. 02:55.980 --> 02:59.180 Just like we saw in the playground here. 02:59.220 --> 02:59.700 Right. 02:59.740 --> 03:02.660 So we have here now the tool result. 03:02.660 --> 03:07.140 And the tool result is often referred to as an observation. 03:07.780 --> 03:14.860 And the LinkedIn component that was responsible of running and executing this tool call was the agent 03:14.860 --> 03:15.540 executor. 03:15.980 --> 03:23.340 So now the agent executor is going to take this observation, the result of the tool, and it's going 03:23.340 --> 03:27.500 to run the same reasoning chain from before. 03:27.660 --> 03:31.620 However, now it's going to attach the observation. 03:31.620 --> 03:33.380 So the result of the tool call. 03:33.420 --> 03:37.100 So now the LLM is going to have the original question. 03:37.260 --> 03:42.620 And it's going to have in its history that it's invoked the search tool with the search result. 03:42.620 --> 03:45.380 And now it's going to come up with the answer. 03:45.540 --> 03:46.100 Right. 03:46.140 --> 03:48.260 So this is what's going to happen here. 03:48.260 --> 03:49.900 And this is why we need the scratchpad. 03:49.940 --> 03:52.020 The scratchpad from the prompt template before. 03:52.060 --> 03:54.340 This is going to hold all of this history right. 03:54.380 --> 03:57.380 All of the tools that were selected and the tool result. 03:57.780 --> 04:01.660 And now the LLM has all the information it needs. 04:01.660 --> 04:05.460 And in this iteration it can actually return us the answer. 04:06.140 --> 04:09.900 And now we can see here in the green the answer of the LLM. 04:10.220 --> 04:12.330 And here we can see it has. 04:12.370 --> 04:13.450 Final answer. 04:13.650 --> 04:15.890 And now the final answer of the LM. 04:16.170 --> 04:18.730 And the agent finished running. 04:18.730 --> 04:25.250 And it provided us with three job posting for an AI engineer using LinkedIn. 04:25.770 --> 04:31.930 And at the end, LinkedIn is going to return us an object, which is a dictionary which has the key 04:31.970 --> 04:34.730 of input and has the key of output. 04:34.730 --> 04:40.930 So the input is going to be our original query and the output is going to be the final LLM response. 04:40.970 --> 04:41.530 All right. 04:41.570 --> 04:44.330 Let's now go and review the trace. 04:44.570 --> 04:48.890 So let me open up my browser and let me go to the Linux platform. 04:49.330 --> 04:54.770 And here I have the project link Smith project which I defined which is called search agent. 04:55.170 --> 04:56.850 And let's check out what we have here. 04:56.970 --> 04:59.490 So we can see here an agent executor run. 04:59.690 --> 05:01.330 Here we can see our input. 05:01.370 --> 05:02.890 Here we can see the output. 05:03.290 --> 05:06.210 And we can see we got here an answer. 05:06.250 --> 05:09.210 Let me go and click this URL here. 05:09.370 --> 05:12.570 Let me show you that this is a LinkedIn job posting. 05:12.570 --> 05:14.400 So we can see Um. 05:14.600 --> 05:16.240 Yeah, this is a job posting. 05:16.240 --> 05:17.520 Let's check out this one. 05:22.160 --> 05:22.440 Yeah. 05:22.480 --> 05:25.240 This is also a job posting for an AI engineer. 05:26.200 --> 05:28.560 And let me go and click on this one. 05:29.640 --> 05:30.000 Yeah. 05:30.040 --> 05:31.080 And this one also. 05:31.080 --> 05:33.200 And here you can see even LinkedIn in the title. 05:34.160 --> 05:36.120 All right let's go back to the traces. 05:36.120 --> 05:38.480 And let's try to break it down step by step. 05:38.720 --> 05:40.920 And this is the prompt template. 05:40.920 --> 05:43.400 So we can see this is the input for the prompt template. 05:43.400 --> 05:46.120 And the output is this text right over here. 05:46.480 --> 05:54.080 Now you can see that this is the prompt populated with all the details about our tools which are the 05:54.400 --> 05:54.840 search. 05:54.960 --> 05:58.240 So we have here the tool names and the tool descriptions. 05:58.240 --> 06:01.760 And this is what eventually is going to be passed to the LLM. 06:01.920 --> 06:08.320 And we can see here instead of the placeholder, we have the question that we gave as an input when 06:08.320 --> 06:09.480 we ran the agent. 06:10.120 --> 06:10.640 All right. 06:10.680 --> 06:12.840 And this is the LLM call. 06:12.840 --> 06:16.870 So we can see now here we simply sent that to the LM. 06:16.870 --> 06:19.710 And here we can see now the LM response. 06:20.310 --> 06:22.390 And after we get the LM response. 06:22.390 --> 06:27.230 And this is the part I didn't dive into, the engine is going to parse it. 06:27.510 --> 06:30.230 So it's going to take from this raw text here. 06:30.230 --> 06:35.230 And it's going to extract the tool name and the tool input that we need. 06:35.430 --> 06:38.350 And we can see here that the type is agent action. 06:38.350 --> 06:41.110 So this is the action that the agent needs to do. 06:41.150 --> 06:42.270 We have here log. 06:42.270 --> 06:47.030 So this is sort of the justification of why the agent did what it did. 06:47.350 --> 06:49.990 And everything here is very organized. 06:49.990 --> 06:51.430 And don't worry about this. 06:51.470 --> 06:55.270 In the next section we'll be diving deep into all of this and how this is working. 06:55.550 --> 06:55.990 All right. 06:55.990 --> 06:58.950 So then the agent executor ran the tool. 06:59.110 --> 07:02.510 So here we can see that this is the tool input. 07:02.510 --> 07:04.590 And here we get the tool output. 07:04.990 --> 07:06.350 So this is what we have here. 07:06.510 --> 07:10.870 And then we have another iteration of the reasoning. 07:11.070 --> 07:17.070 So this is now the second prompt which is sent to the LM which is kind of like the original one. 07:17.070 --> 07:23.020 But it also has the information about the tool call that we made and what is the tool result. 07:23.020 --> 07:29.940 And then the agent is going to sorry, the LM is going to respond with an answer that we can see right 07:29.940 --> 07:31.860 over here at the bottom here. 07:32.140 --> 07:34.540 And yeah, that's pretty much it. 07:34.580 --> 07:39.260 Link chain is going to go take it going to parse it going to do it very very nicely. 07:39.260 --> 07:41.900 And then it's going to output us the response here. 07:42.500 --> 07:46.620 Let me just go and commit this code now to the repository. 07:47.060 --> 07:49.100 So let's create a new commit. 07:50.420 --> 07:50.980 Sorry. 07:51.020 --> 07:52.180 Let's create the commit. 07:52.180 --> 07:56.740 And let me call it react search agent. 07:57.380 --> 07:59.260 And let me go and push it. 08:00.740 --> 08:01.140 Sorry. 08:01.140 --> 08:02.740 Let's go and fix that. 08:02.740 --> 08:04.740 I have a bit of configuration issue. 08:04.780 --> 08:05.980 Don't worry about it. 08:07.460 --> 08:14.860 And if you want to check out the exact code I was using, go to the branch project slash react agents 08:15.180 --> 08:16.940 go here to the commit list. 08:17.100 --> 08:19.500 And here is the commit I just pushed. 08:19.980 --> 08:21.740 Let me go and minimize this. 08:21.740 --> 08:28.020 And here you can see the exact code that I was using with everything I just showed you in the previous 08:28.020 --> 08:28.580 videos. 08:29.820 --> 08:31.740 Hey there Eden here popping up. 08:31.740 --> 08:34.060 And that was a lot. 08:34.300 --> 08:39.180 So we created and ran our first link chain agent. 08:39.700 --> 08:44.980 Now it was comprised of the reasoning part and the execution part. 08:45.180 --> 08:49.500 And we saw it all boils down to a very special react prompt. 08:49.580 --> 08:54.460 The entire goal of this section is simply to get you familiar with the interface. 08:54.500 --> 08:58.940 It's okay that you're not understanding every moving part here under the hood. 08:59.100 --> 09:01.340 This is what we're going to do in the next section. 09:01.340 --> 09:07.220 We're going to break this down one layer deeper, and we're going to understand all the chain components 09:07.220 --> 09:09.020 that are running together here. 09:09.020 --> 09:11.780 And then we'll truly appreciate all the heavy lifting. 09:12.140 --> 09:13.620 Link chain is doing for us. 09:13.860 --> 09:19.260 And the cool thing that we can see here is that the reasoning engine, it all boils down into a prompt 09:19.260 --> 09:20.380 we send to the LM. 09:20.420 --> 09:21.700 I mean, how cool is that? 09:21.740 --> 09:27.970 And this is the basis for all the genetic technology we're seeing right now, this is the core primitive. 09:28.010 --> 09:29.130 This is what was started. 09:29.490 --> 09:30.090 All right. 09:30.090 --> 09:33.730 So the answer we got from our agent is text. 09:33.770 --> 09:37.250 While text is nice, we can't really use it in software. 09:37.290 --> 09:41.330 What we really want to do is to get structured output. 09:41.370 --> 09:45.810 We want to get this kind of output into a Python object. 09:45.810 --> 09:52.650 We can go and we can maybe serialize, we can pass downstream to another application and something we 09:52.650 --> 09:54.410 can work with programmatically. 09:54.530 --> 10:00.970 So in the next couple videos we're going to introduce the idea of output parsers which are going to 10:01.010 --> 10:02.410 take the LLM response. 10:02.410 --> 10:07.050 And they're going to structure it as JSON objects as Pydantic objects. 10:07.050 --> 10:11.090 And they're going to help us use programmatically the output of agents. 10:11.290 --> 10:14.810 And this is very important when building robust software. 10:15.130 --> 10:17.330 Now I'm going to give you a quick hint. 10:17.890 --> 10:20.450 The answer to that is going to be a prompt as well. 10:20.770 --> 10:22.730 So everything boils down to a prompt. 10:22.730 --> 10:23.930 This is beautiful here. 10:24.050 --> 10:25.810 Anyways, let's go to the code.