WEBVTT 00:00.200 --> 00:00.720 All right. 00:00.720 --> 00:02.280 So we have the tool. 00:02.280 --> 00:04.720 And now we need to configure an LM. 00:04.920 --> 00:06.520 So let's go and do that. 00:06.520 --> 00:09.040 So let me initialize a chat open AI. 00:09.240 --> 00:11.200 And I'm going to give it GPT four. 00:11.600 --> 00:15.200 There's a reason why I'm using GPT four and not GPT five. 00:15.200 --> 00:18.680 And I'm going to reveal this reason in the end of this section. 00:18.960 --> 00:25.920 And now let's go and use the link chain hub to pull something, which is called a react prompt. 00:25.960 --> 00:29.160 Now before we run this, just a quick note here. 00:29.160 --> 00:36.400 In case you're using the EU endpoint for meat, you might encounter this error where it says that it 00:36.400 --> 00:37.920 cannot find this prompt. 00:37.960 --> 00:43.640 And the reason you're getting this, because the public prompt or the public prompts are hosted in the 00:43.640 --> 00:48.120 US instance of Lenox-smith, which is the default public hub instance. 00:48.280 --> 00:54.840 And if you're using the EU instance, then it does not automatically mirror all the public prompts from 00:54.840 --> 00:55.720 the US hub. 00:55.960 --> 00:58.880 So in case you're getting this, we have two options. 00:58.880 --> 01:06.430 Either you remove the environment variable of length underscore endpoint, or you can simply go and 01:06.430 --> 01:10.630 copy paste this prompt template and create a prompt template object. 01:10.790 --> 01:12.110 And let me run it in debug. 01:12.110 --> 01:13.910 And let me show you this react prompt. 01:14.270 --> 01:18.150 So this react prompt over here is going to be a prompt template. 01:18.310 --> 01:21.150 But it's going to be a very special prompt template. 01:21.350 --> 01:23.910 Those are going to be its input variables. 01:23.950 --> 01:27.910 It's going to receive an input which is going to be the user's question. 01:28.270 --> 01:30.990 It's going to receive tools and tool names. 01:30.990 --> 01:36.350 And those are all things that length chain is going to extract from the tool object that we built earlier. 01:36.790 --> 01:41.270 And it's going to have also the agent scratchpad, but ignore it for now. 01:41.270 --> 01:45.270 In the next section we'll be elaborating on this agent scratchpad here okay. 01:45.470 --> 01:48.830 So this is the variables that it's going to receive. 01:49.190 --> 01:52.750 Now let me show you the template that it's using. 01:52.750 --> 01:53.110 Right. 01:53.110 --> 01:56.590 So we can have an idea what's going to be sent to the LM. 01:56.870 --> 01:58.230 So the template here is. 01:58.230 --> 02:05.350 Answer the following question as best as you can, and let me open it in the prompt tab so we can review 02:05.390 --> 02:07.270 it in a much nicer UI. 02:07.310 --> 02:09.910 Okay so this is the actual prompt template. 02:10.270 --> 02:12.990 Now you can see here it's going to receive tools. 02:13.430 --> 02:14.550 The tool names. 02:14.750 --> 02:17.550 And then we have here the question is going to be the input. 02:17.550 --> 02:20.110 And we have here thought Agent Scratchpad. 02:20.150 --> 02:20.590 Okay. 02:20.910 --> 02:25.110 So in this react prompt we have this middle part over here. 02:25.350 --> 02:28.430 And this is actually what's going to be the magic. 02:28.430 --> 02:31.470 And it's going to be the secret sauce of our reasoning. 02:31.470 --> 02:33.110 And this is how it all started. 02:33.150 --> 02:33.350 Okay. 02:33.390 --> 02:35.710 It's simply everything is a prompt with LLM. 02:35.710 --> 02:38.670 And it really amazes me to really think about that. 02:38.670 --> 02:44.190 This is the basis for all of this modern technology that we have today, like cursor, Cloud Code, 02:44.430 --> 02:46.870 Devin, and all of those deep agents. 02:47.030 --> 02:50.750 And I'm not going to elaborate on this prompt for that. 02:50.750 --> 02:55.630 I have a dedicated video to discuss why did it came from, why it's useful. 02:55.630 --> 02:57.660 And this is called the react prompt. 02:57.900 --> 03:00.260 It originated from react paper. 03:00.260 --> 03:02.460 And this is really the basis for everything. 03:02.860 --> 03:06.980 But for this section all we need to know is that this prompt works. 03:06.980 --> 03:10.060 And for some reason everything would work with this prompt. 03:10.100 --> 03:12.420 Okay, so this is what we need to have in mind. 03:12.460 --> 03:14.100 We're not going to elaborate on this video. 03:14.100 --> 03:18.100 I highly suggest you check this theoretical part in case you want to dive deeper into this. 03:18.620 --> 03:19.420 All right. 03:19.420 --> 03:27.140 So now we want to go and we want to create our reasoning engine. 03:27.300 --> 03:27.980 All right. 03:27.980 --> 03:34.660 So for that we'll use the create react agent function which is going to receive the LM. 03:35.020 --> 03:39.820 It's going to receive the tools and it's going to receive this react prompt. 03:39.980 --> 03:43.180 Now this is going to create for us a chain. 03:43.180 --> 03:45.780 And this is going to be our reasoning chain. 03:45.780 --> 03:48.100 You can see that it's going to return a runnable. 03:48.100 --> 03:50.100 So this is an equivalent for a chain. 03:50.460 --> 03:56.570 And it's going to plug all of the values of of the tools, and the input that the user will provide 03:56.570 --> 03:58.930 is going to provide everything to that prompt. 03:59.210 --> 04:05.610 Now, hopefully when we're going to run this chain, then the model, the LM is going to reason on the 04:05.610 --> 04:10.690 input and it's going to say, hey, you need now to run the search with the following query. 04:10.810 --> 04:11.410 All right. 04:11.530 --> 04:15.290 So this is hopefully what this chain is going to return us. 04:15.450 --> 04:16.050 All right. 04:16.130 --> 04:18.730 So this is the reasoning engine right now. 04:18.810 --> 04:23.050 So we get that part really set up here we have the reasoning engine. 04:23.050 --> 04:26.010 And now we need the execution of the tools here. 04:26.530 --> 04:27.210 All right. 04:27.210 --> 04:31.890 So it's going to be here this agent executor that we imported from before. 04:32.210 --> 04:35.250 Now it's going to receive the chain from before. 04:35.250 --> 04:36.450 By the way I called it agent. 04:36.450 --> 04:37.650 This is not really an agent. 04:37.650 --> 04:39.090 It's going to be a simple chain. 04:39.450 --> 04:41.090 It's going to receive the tools. 04:41.290 --> 04:44.130 And I also added here verbose equals to true. 04:44.130 --> 04:46.010 So we'll be able to see when we run it. 04:46.050 --> 04:47.730 We'll be able to see the reasoning process. 04:47.730 --> 04:49.730 And we'll be able to see more logs here. 04:50.090 --> 04:50.490 All right. 04:50.490 --> 04:52.640 So this is going to be our agent. 04:52.640 --> 04:53.440 Runtime. 04:53.840 --> 04:56.320 I remind you, don't worry about the internal implementation. 04:56.320 --> 05:00.240 We're going to see this and we're going to implement everything from zero in the next section okay. 05:00.280 --> 05:02.920 So you know exactly how this thing is working right now. 05:03.120 --> 05:06.400 But what this is going to do, it's going to run in a while loop. 05:06.400 --> 05:10.760 And it's going to run this react chain over and over again. 05:10.760 --> 05:15.520 And every time there needs to be a tool call, it's going to call that tool, and it's going to feed 05:15.520 --> 05:20.400 the output of the tool back to the LM, and it's going to do this until it finishes the task here. 05:20.440 --> 05:21.080 All right. 05:21.080 --> 05:23.080 So this is basically what it's going to do. 05:23.600 --> 05:26.840 And don't worry about the arguments that it will receive. 05:26.840 --> 05:31.600 We're going to cover it in the next section when we'll implement everything here from zero. 05:31.800 --> 05:34.600 So I promise you you'll know exactly how this is working. 05:34.840 --> 05:38.760 Let me just show you a quick peek of the implementation of link chain. 05:38.760 --> 05:41.080 So this is the Create React agent here. 05:41.240 --> 05:42.880 And you can see what it's going to return. 05:42.880 --> 05:48.240 Here is a runnable object where it's going to take the prompt the react prompt. 05:48.240 --> 05:50.750 And it's simply going to send it to them. 05:50.950 --> 05:54.430 It has this output parsing which I'm going to discuss later. 05:54.430 --> 05:57.390 So please, please, please don't worry about this output parsing. 05:57.430 --> 06:02.350 Okay, I remind you what we're interested here is how to use the link chain interface. 06:02.350 --> 06:06.470 In the next section we're going to see exactly how it's implemented because we're going to be implementing 06:06.470 --> 06:07.270 it from zero. 06:07.550 --> 06:08.150 All right. 06:08.150 --> 06:13.630 So just to summarize what we saw so far, we defined our reasoning engine, which is going to come from 06:13.630 --> 06:19.350 the create react agent function which is going to return us with a chain which is simply going to receive 06:19.350 --> 06:23.950 the tools, or it's going to receive the user's query, and it's going to send everything to the LLM. 06:24.310 --> 06:31.430 Once we have that, the LLM is going to return us a response and the agent executor is going to be responsible 06:31.430 --> 06:34.830 to maybe call a tool or to run another LLM call. 06:34.830 --> 06:38.430 And it's going to be the orchestrator of this prompt over here. 06:38.870 --> 06:39.430 All right. 06:39.430 --> 06:41.830 So we are about to finish now the code. 06:42.070 --> 06:45.670 And let's now go and let's invoke our agent. 06:45.990 --> 06:48.030 So let me just create a new variable. 06:48.030 --> 06:54.390 And I'm going to call it chain, and I'm going to make it equal to the agent executor here. 06:54.430 --> 06:58.630 Now, I know this looks funny here, but this is actually preparation for something that we're going 06:58.630 --> 06:59.630 to be doing next. 07:00.110 --> 07:03.470 And now let's go and invoke this chain. 07:03.510 --> 07:04.070 All right. 07:04.070 --> 07:07.150 So I'm going to paste here this snippet. 07:07.190 --> 07:12.790 And here we have a variable called result which is going to be our chain when we invoke it. 07:12.870 --> 07:18.390 And here we're going to provide the invoke method the keyword argument of input. 07:18.390 --> 07:20.110 And it's going to be a dictionary. 07:20.270 --> 07:23.630 Now this dictionary is going to have the key of input. 07:23.830 --> 07:30.590 And this input right over here is going to be what's actually going to be populated in the prompt template 07:30.590 --> 07:31.830 that we defined above. 07:32.030 --> 07:37.550 And the query is going to be search for three job postings for an AI engineer using link chain in the 07:37.550 --> 07:40.430 Bay area on LinkedIn and list their details. 07:40.950 --> 07:45.510 And in the next video we're going to run, debug and examine the output. 07:45.510 --> 07:47.390 So hopefully it will be clearer then.