WEBVTT 00:00.040 --> 00:03.640 So in this video we're going to implement the react py file. 00:03.640 --> 00:08.360 And this is going to hold all the reasoning logic which we're going to use in our graph. 00:08.520 --> 00:14.080 So the TLDR for this is that we're going to use an average function calling that is going to be the 00:14.080 --> 00:17.000 reasoning engine which decides which tool to call. 00:17.920 --> 00:26.440 Let's start with the imports and lets me import from env the load function so I can load the environment 00:26.440 --> 00:31.960 variables and all the API keys I want to import from link chain core. 00:32.360 --> 00:35.800 I want to import the tool decorator. 00:35.960 --> 00:40.680 So this will help us to create a tool from our functions that we define. 00:41.160 --> 00:45.240 I want to also import from link chain OpenAI the chat OpenAI. 00:45.480 --> 00:48.000 So we can make LLM calls to GPT. 00:48.600 --> 00:57.680 We want to import from link chain Tavileh the search object which is a pre-made search tool. 00:57.680 --> 01:02.000 So we can use and plug it in to our llms to our agent. 01:02.680 --> 01:06.900 And now let's go and load our environment variables. 01:07.300 --> 01:09.500 And this is time to begin the coding. 01:09.740 --> 01:12.220 And we want now to implement the tools. 01:12.500 --> 01:20.220 So for this example we'll be implementing the triple tool which is a function which takes in as an input 01:20.260 --> 01:22.100 an integer a float. 01:22.220 --> 01:25.580 And it's going to return us this float triple. 01:25.860 --> 01:30.220 Now notice here all that I'm doing is writing the prefix here of the function. 01:30.220 --> 01:32.620 And I'm going to be auto completed by cursor. 01:32.780 --> 01:36.620 This is simply, by the way, mind blowing how coding has changed over the years. 01:36.620 --> 01:42.500 And with LMS how easy it is, um, and how convenient it is to to implement those kind of things. 01:42.820 --> 01:45.220 So we can see here an example implementation. 01:45.220 --> 01:47.660 Let me remove this unnecessary code here. 01:47.940 --> 01:54.380 And we have here the triple function we have here the description of this function which is going to 01:54.380 --> 01:56.700 be propagated eventually to the LM. 01:56.700 --> 02:00.500 And the LM is going to decide whether to use this function or not. 02:01.220 --> 02:03.780 Anyways so this is the triple function. 02:03.820 --> 02:11.030 Now if we want to turn it into a link chain tool, All we need to do is to use the decorator of tool 02:11.070 --> 02:16.110 here that we imported, and we get a link chain tool that we can plug in. 02:16.870 --> 02:17.670 Alrighty. 02:17.710 --> 02:26.070 So now we want to equip our agent with not just the triple tool, but we also want to give it the search 02:26.070 --> 02:28.310 tool as a tool that it can use. 02:28.710 --> 02:32.110 So in order to do that let's define a new variable. 02:32.110 --> 02:33.510 It's called tools. 02:33.510 --> 02:40.870 And tools is going to be a list which is going to be an object of the search prebuilt tool and the triple 02:40.910 --> 02:42.670 tool that we just implemented. 02:43.630 --> 02:49.750 So the first element of this list is this object of this class that we search initialized with the max 02:49.750 --> 02:51.190 results equals to one. 02:51.470 --> 02:56.870 So in this example we're simply limiting the search to fetches only one search result. 02:57.190 --> 03:03.750 And by the way this search tool has already the description prebuilt in case you were wondering. 03:03.910 --> 03:10.590 So the package maintainers implemented the search description. 03:10.590 --> 03:13.890 And this is going to be propagated eventually to the LM. 03:16.650 --> 03:20.090 And the second element of that list is the triple tool. 03:20.610 --> 03:21.210 All right. 03:21.210 --> 03:27.690 So now we want to tell the LM somehow that it has access to this kind of tools. 03:29.290 --> 03:29.650 All right. 03:29.650 --> 03:34.010 So now I want to discuss the reasoning capabilities of Dlrm. 03:34.210 --> 03:38.090 So how does the LM decide which tool to call. 03:38.530 --> 03:44.930 So by now it should be familiar with the algorithm and the react prompt. 03:45.170 --> 03:53.770 And this is a very fancy prompt that is a derivative from the react paper which is helping to leverage 03:53.770 --> 03:56.250 the reasoning capabilities of the LM. 03:56.530 --> 03:58.530 So this is how everything started. 03:58.530 --> 04:02.290 And in the early days of agents this is how it was used. 04:02.290 --> 04:07.730 People used this special prompt in order to leverage Llms to decide which tool to call. 04:08.330 --> 04:14.380 And by now things have evolved and we have better ways to do things which, by the way, all derive 04:14.380 --> 04:16.300 from this very special prompt. 04:16.340 --> 04:23.020 What I am referring to right now is function calling, and function calling is a feature in most modern 04:23.020 --> 04:30.740 llms today that allow us to give the LLM when we initialize it, the definition and the instructions 04:30.740 --> 04:33.180 and the details of the tools that we want to give. 04:33.220 --> 04:40.300 The LLM and the LLM is going to return us in the response, whether we need to call those functions 04:40.300 --> 04:43.020 with which arguments if it needs to. 04:43.060 --> 04:49.220 And I elaborate on function calling in this course, in this video and in this video, and we are actually 04:49.220 --> 04:53.700 not exposed to the implementation of function calling in LLM vendors. 04:53.700 --> 04:57.260 So each LLM vendor has its own way of doing things. 04:57.300 --> 05:05.700 The implementation is probably a special system prompt, which is similar to the react prompt and helps 05:05.700 --> 05:11.620 the LLM to decide whether it needs to choose a tool or not, and returns the answer in the correct format. 05:11.900 --> 05:20.240 However, now the LLM vendor is responsible to parse the the LLM response and to arrange everything 05:20.240 --> 05:27.240 nicely in the response we're getting and putting the function call in the function call key in the response 05:27.240 --> 05:28.200 from the LLM. 05:28.240 --> 05:31.040 And again, don't worry if you don't understand this in full. 05:31.360 --> 05:36.920 All you should know and all you should take right over here that this is the modern way to choose tools. 05:36.920 --> 05:42.720 And if we're going to use function calling, we're basically offloading the task of choosing the correct 05:42.720 --> 05:44.840 tool to the LLM vendor. 05:44.840 --> 05:51.720 And we don't need to handle this parsing like we did from before, because the LLM does it for us. 05:51.720 --> 05:56.560 So for us, it's a very big win because it's less code that we need to write. 05:56.560 --> 06:02.640 And actually the results and the performance of this is only improving over time because all the LLM 06:02.640 --> 06:07.920 vendors, they have dedicated engineers that are defining this and are making sure that we get quality 06:07.920 --> 06:11.760 results and the correct tools to be called and the correct arguments. 06:12.120 --> 06:12.720 All right. 06:12.720 --> 06:18.800 So right now we want to implement the reasoning capabilities of our agent. 06:19.200 --> 06:23.740 And unlike we did before, we're not going to use the react prompt. 06:23.780 --> 06:27.060 We have something which is better and this is function calling. 06:27.180 --> 06:31.020 So we're going to initialize an LM that supports function calling. 06:31.060 --> 06:34.500 And we're going to give that LM the tools that we defined. 06:34.540 --> 06:41.700 Now the search tool and the triple tool both are link chain tools which have the descriptions of which 06:41.740 --> 06:43.060 arguments do they receive. 06:43.660 --> 06:47.260 So I'm going to initialize a chat open AI. 06:47.620 --> 06:54.340 And I'm going to use the bind tools method and supply the tools that we already wrote. 06:54.740 --> 07:02.900 And link chain is going to take the tool descriptions and it's going to send that to the LM. 07:02.980 --> 07:05.580 Every request that we're going to make. 07:05.940 --> 07:12.660 So the LM can now return us this function call or tool calling field with the correct function to be 07:12.660 --> 07:13.140 called. 07:13.580 --> 07:19.180 And we don't need to handle any parsing because the LM vendor handles the parsing for us. 07:19.180 --> 07:22.900 And if there is a function call we get it in a special key in the response. 07:23.420 --> 07:28.430 And for more information about the implementations of function calling, please check out those videos 07:28.430 --> 07:29.070 over here. 07:29.910 --> 07:30.350 Cool. 07:30.350 --> 07:32.150 So this is it for this video. 07:32.270 --> 07:36.910 And I want to run this, um, script right over here. 07:37.110 --> 07:41.670 Let's see that everything runs even though we're not running anything right now, just to see we don't 07:41.670 --> 07:44.990 get an error and we can see it's working. 07:44.990 --> 07:48.270 And let me go and commit the code where you can find it. 07:50.510 --> 07:52.750 So let me write git add. 07:54.230 --> 07:57.990 And we have here this file that was added right now. 07:58.550 --> 08:00.150 So let's go and commit it. 08:01.310 --> 08:05.790 And I'll call the commit function calling reasoning. 08:06.950 --> 08:08.270 And let me push it. 08:09.790 --> 08:11.390 And let me show you the repo. 08:11.910 --> 08:17.870 And if you'll go to the commits you can see you have this commit right over here where you can see the 08:17.870 --> 08:18.430 code. 08:18.630 --> 08:22.390 And in the next video we're going to implement our lane graph nodes. 08:22.390 --> 08:27.270 And those are the executables which are going to run during our execution.