WEBVTT 00:01.400 --> 00:05.420 Let's now review the react prompt engineering technique. 00:05.600 --> 00:11.750 Like most of the prompt engineering techniques, it came from a research paper that you can see right 00:11.750 --> 00:12.290 now. 00:12.320 --> 00:18.320 Now react is basically based on reasoning and acting reasoning enacting. 00:18.320 --> 00:20.810 And this is where the name comes from. 00:20.810 --> 00:24.920 One part is reasoning the re and the act is for acting. 00:25.670 --> 00:28.220 So how does it relate to us humans? 00:28.670 --> 00:35.390 It is very natural for humans when we have a complex task, is to reason and think about the steps we 00:35.390 --> 00:41.780 need to do in order to accomplish this task, perform these steps, and then move on to the next step. 00:41.810 --> 00:44.570 So we break it down to multiple steps. 00:44.570 --> 00:45.800 We reason it down. 00:45.800 --> 00:47.510 We act upon those steps. 00:47.510 --> 00:52.070 And then after we complete everything, we can complete the entire task. 00:52.820 --> 01:00.770 So what those researchers found out is that if we take the chain of thoughts and we combine it with 01:00.770 --> 01:07.810 the ability to perform those actions and get, for example, information from external sources. 01:08.170 --> 01:13.090 The LM model can generate its task that it needs to do. 01:13.120 --> 01:19.690 This is the reason part, and we saw it earlier in the chain of thought that it is capable of doing 01:19.690 --> 01:20.230 that. 01:20.230 --> 01:28.810 And if it's able to perform and execute those tasks in overall, they will be able to resolve the entire 01:28.810 --> 01:29.710 task. 01:29.950 --> 01:36.190 So here we are combining the ability of the LM to generate its own task, the reasoning to track it 01:36.190 --> 01:43.210 and to update the plan accordingly, and to also execute those steps to get additional information from 01:43.210 --> 01:44.320 external sources. 01:44.350 --> 01:48.430 So, so this is found out as a very, very powerful technique. 01:49.330 --> 01:57.790 It is pretty much the basis for LM applications that are way more factual than ChatGPT and can use external 01:57.790 --> 01:59.350 APIs and tools. 01:59.560 --> 02:05.770 So this is pretty much what pioneered is something that is called link chain, which is a very popular 02:05.770 --> 02:09.370 framework for building LLM applications. 02:10.090 --> 02:17.170 So let's review and analyze this research paper that will demonstrate us the react paradigm. 02:17.200 --> 02:20.830 So we have here a prompt with a question. 02:20.830 --> 02:29.620 So aside from the Apple remote what other device can control the program of the Apple remote was originally 02:29.680 --> 02:31.930 intended to interact with. 02:33.880 --> 02:38.920 So here we can see that for the prompt above we got iPad which is not the true. 02:38.950 --> 02:45.400 This was the simple zero shot prompt, and with the chain of thought where we asked it to describe the 02:45.400 --> 02:53.350 way it was thinking, then the LLM responded us with some other outputs iPhone, iPad, iPod, etc. 02:53.440 --> 02:56.200 but it wasn't correct as well. 02:56.890 --> 03:02.920 So the last attempt was something which is called act only, and this is available only on Llms that 03:02.920 --> 03:05.140 can interact with the outside world. 03:05.140 --> 03:07.210 So there is the syntax of search. 03:07.210 --> 03:15.900 So basically right over here, um, the the prompt says to search for Apple remote and then to list 03:15.900 --> 03:16.980 its observations. 03:17.280 --> 03:23.700 So we can see over here the final result was yes, which is simply not the answer for the question. 03:23.700 --> 03:28.140 And basically all those three examples that we saw before are wrong. 03:28.860 --> 03:37.620 And we can see that when we had models that were based on the react paradigm, and the answer that we 03:37.620 --> 03:42.330 got from them was the correct answer, which is, by the way, the keyboard function keys. 03:42.720 --> 03:49.650 So let's now analyze what's happening with the reason and the acting of the LMS when it comes to react. 03:49.830 --> 03:57.840 Now with this particular problem, the LMS first comes up with a thought, a breakdown of what it needs 03:57.840 --> 03:58.500 to do. 03:58.530 --> 04:00.720 Like we simply know with the chain of thoughts. 04:00.720 --> 04:06.690 So the LMS says, okay, we need to search for the Apple remote and find what program it was originally 04:06.690 --> 04:08.160 designed to interact with. 04:08.880 --> 04:15.750 So it derived an act, something, something that it needs to do to access an external resource, which 04:15.780 --> 04:16.200 he received. 04:16.200 --> 04:18.480 The syntax of search for the Apple remote. 04:18.510 --> 04:24.300 Once he does that and searches for the Apple remote and gets information from it from outside sources, 04:24.300 --> 04:25.740 it makes an observation. 04:25.740 --> 04:32.070 And the observation was that according to the external source, then the Apple remote was originally 04:32.070 --> 04:34.830 designed to control the front row media center. 04:34.860 --> 04:39.660 Now this creates another thought to the LM model. 04:39.780 --> 04:44.790 This is similar to the chain of thoughts, but we can act upon those thoughts. 04:44.910 --> 04:51.090 The second one is that okay, now the LM knows that there is something to do with the front row media 04:51.120 --> 04:56.400 center, and that's why it's determining that the next act should be to search for something which is 04:56.400 --> 04:57.930 called the front row. 04:58.020 --> 04:59.070 For some reason. 04:59.070 --> 05:01.440 Now it observed that it didn't find anything. 05:01.440 --> 05:04.800 It did find something similar to front row software. 05:04.800 --> 05:09.450 So now it creates a new thought that it should examine the front row software. 05:09.450 --> 05:13.610 So the act deriving from it would be to search for the front row software. 05:13.640 --> 05:19.190 It makes one observation after upon the act and it not so it's not so relevant. 05:19.190 --> 05:25.820 But the final thought is what is bringing us closer to the answer, because the final thought is that 05:25.820 --> 05:32.060 the front row software is controlled by the Apple remote or the Apple key function keys. 05:32.060 --> 05:36.350 So this is from the external sources that we got on the previous step. 05:36.440 --> 05:41.450 So from this it understands that the answer is the keyboard function keys. 05:41.540 --> 05:47.180 So this is mind blowing because it is simulating the human thought. 05:47.180 --> 05:53.000 And if I needed to answer this question this is exactly what I would have done. 05:53.000 --> 05:55.910 So this is crazy crazy crazy. 05:55.910 --> 05:57.680 And this is super powerful. 05:58.850 --> 06:02.990 Now you might be thinking, whoa, this is some sort of magic. 06:02.990 --> 06:04.130 How does this happen? 06:04.130 --> 06:07.430 I mean, I haven't really told you what's what's happening behind the scenes. 06:07.460 --> 06:09.620 I mean, how how does this even work? 06:09.650 --> 06:17.280 So basically, what's happening is simply integrating some kind of code with the chain of thoughts we 06:17.280 --> 06:24.150 saw earlier, that the chain of thoughts can create thoughts and can create steps, and the LM is totally 06:24.150 --> 06:26.220 capable of generating those steps. 06:26.220 --> 06:30.300 The next thing to do is to simply act upon those steps. 06:30.300 --> 06:32.430 So this can be done by code. 06:32.520 --> 06:39.210 And it's not even that hard as you may think, because doing so only involves taking the output step, 06:39.480 --> 06:45.000 which describes what we need to do, and looking up for keywords like search and finding the topic. 06:45.030 --> 06:47.100 So this can be easily done by code. 06:47.130 --> 06:55.710 And then to rerun the prompt with the observations that we got and do it over and over until we get 06:55.710 --> 06:56.670 the solution. 06:56.670 --> 06:58.890 So this is pretty much mind blowing. 06:58.890 --> 07:07.020 And this is exactly the basis for something which is called the link chain, which is the most popular 07:07.020 --> 07:14.270 framework right now to implement LM applications which can handle and do complex stuff like interacting 07:14.270 --> 07:18.410 with external sources, persisting data, and a lot more.