WEBVTT 00:00.160 --> 00:03.960 So now it's time to cast this function into a long chain tool. 00:04.200 --> 00:09.560 And we're going to later equipped our agent to have the ability of using this tool. 00:09.880 --> 00:15.280 So in the previous video, we defined our tool manually by using the tool class of chain. 00:15.560 --> 00:23.160 But we can also use a tool decorator that will take a function and convert it into a long chain tool. 00:23.400 --> 00:27.360 So for that I'm going to import the tool function. 00:28.360 --> 00:33.240 And I'm simply going to use this decorator above this function. 00:35.680 --> 00:41.960 The tool decorator is a long chain utility function that will take this function and create a chain 00:41.960 --> 00:42.800 tool from it. 00:42.840 --> 00:48.360 It's going to plug in the name of the function, what it receives as arguments, what it returns, its 00:48.360 --> 00:51.840 description, and populate it in the chain tool class. 00:52.960 --> 01:01.880 So if I'm going to debug this code and put a breakpoint right over here and let's now inspect this function, 01:01.880 --> 01:02.760 get text length. 01:02.760 --> 01:12.720 We can see now it's a structured tool and we have the name and we can see and have its description which 01:12.760 --> 01:17.800 includes the function name, its arguments, what it returns and the description that we wrote. 01:17.840 --> 01:23.680 This will be used by the LM reasoning engine to decide whether to use this tool, and we'll soon see 01:23.680 --> 01:24.640 how it works. 01:25.880 --> 01:29.880 Just one minor change we need to make to our getText function. 01:29.880 --> 01:37.400 So sometimes the LM may pass as our text some unnecessary characters which are not alphabetical. 01:37.400 --> 01:41.440 So for example backslash n or some single quote. 01:41.640 --> 01:44.840 So here I'm simply removing those and cleaning it up. 01:45.000 --> 01:50.680 And all the code for each video in this section is available in the videos resources. 01:51.120 --> 01:56.320 Now, I highly suggest that you use this implementation of the function in the video, even though I 01:56.320 --> 01:57.320 forgot to add it. 01:59.520 --> 02:00.040 Cool. 02:00.040 --> 02:07.590 So now I want now to stop it and let's create a new variable and let's go and call it tools. 02:08.030 --> 02:13.790 It's going to be a list of long chain tools, which now is going to be populated only with get text 02:13.830 --> 02:14.430 length. 02:14.710 --> 02:18.950 Now this list of tools we're going to supply to our react agent. 02:20.030 --> 02:25.390 So remember when previously the agent selected the correct tool to use from our prompt. 02:25.630 --> 02:27.830 And it was sort of like magic. 02:28.190 --> 02:34.510 So now we're going to dive deep and to see how it's actually implemented and what's happening under 02:34.510 --> 02:34.990 the hood. 02:36.350 --> 02:36.870 Okay. 02:36.910 --> 02:39.670 Now it's time to go and see this in action. 02:40.030 --> 02:44.470 And all of this tool selection basically is going to happen from a very special prompt. 02:44.510 --> 02:45.670 We're going to be writing. 02:45.870 --> 02:48.270 So I'm going to write here the template variable. 02:48.270 --> 02:54.190 And now I'm going to attach here a very smart prompt that will make our tool selections. 02:54.910 --> 02:57.990 So you can find this prompt in the source code. 02:57.990 --> 02:59.870 And we'll do it later in this video. 03:00.350 --> 03:05.310 However right now I want to go to the Lang Chain Hub, which is a part of Lang Smith, and the Lang 03:05.310 --> 03:08.110 Chain Hub is a marketplace for prompts. 03:08.150 --> 03:13.990 It's all free, and people share their very well crafted prompts, including the prompts for this react 03:13.990 --> 03:14.630 agent. 03:14.790 --> 03:20.830 Now, this prompt was created by Harrison Chase, the creator of Link Chain, but everyone can submit 03:20.830 --> 03:23.350 prompts and share them on the chain hub. 03:23.870 --> 03:29.070 Anyways, I'm going to open my browser and I'm going to search for Lang Chain Hub. 03:29.350 --> 03:32.830 I'm going to go to the second link, which is a part of Lang Smith. 03:33.190 --> 03:35.030 And right over here we have the list of prompts. 03:35.030 --> 03:38.670 And I'm going to filter by agents and I'm going to write here react. 03:39.110 --> 03:43.350 So right here we have a prompt that was crafted by Harrison Chase. 03:43.350 --> 03:44.470 You can see it right here. 03:44.790 --> 03:47.350 And we can see the prompt right now. 03:47.590 --> 03:50.790 So on the right side you can see when it was last committed. 03:50.790 --> 03:52.070 When was the last change. 03:52.310 --> 03:58.390 You can see what kind of type this prompt is and the number of downloads it has, how many followers 03:58.390 --> 03:58.830 it has. 03:58.830 --> 04:05.230 Because prompts is something that is dynamic and people always refine and refactor their prompts. 04:05.510 --> 04:10.830 And right now we're simply going to copy and paste it into our IDE. 04:12.310 --> 04:15.910 So I'm going to click the copy button and going to paste it. 04:16.590 --> 04:18.830 Let's just format it a bit more nicely. 04:18.830 --> 04:24.430 I'm going to take all of those lines over here and going to indent it in the react algorithm. 04:24.430 --> 04:26.870 This prompt is going to be sent to the LM. 04:26.870 --> 04:32.310 And this is going to generate the thought of the LM which is going to help us select the correct tool. 04:32.830 --> 04:34.990 So let's analyze this prompt. 04:35.030 --> 04:40.830 This prompt is in my opinion, the most important prompt in LM application development. 04:41.350 --> 04:45.190 And it's an implementation of the react paper. 04:45.510 --> 04:51.070 And if you want to find more about the theory behind this prompt, I highly suggest you going over the 04:51.070 --> 04:57.070 theory section where I analyze prompt engineering techniques from the theoretical point of view and 04:57.070 --> 04:58.670 analyze their research papers. 04:59.070 --> 05:01.630 And you can find a video about react. 05:01.670 --> 05:03.750 Anyways, let's analyze this prompt. 05:03.910 --> 05:08.550 Answer the following questions as best as you can, and you have access to the following tools where 05:08.550 --> 05:10.750 we have a placeholder for tools. 05:11.020 --> 05:15.300 And here we're going to plug in our list of tools and the tool descriptions. 05:15.900 --> 05:17.300 And it continues. 05:17.300 --> 05:18.700 Use the following format. 05:18.700 --> 05:20.340 Question the input question. 05:20.340 --> 05:21.980 You must answer thought. 05:21.980 --> 05:24.100 You should always think about what to do. 05:24.500 --> 05:27.340 And then we have action and action input. 05:27.340 --> 05:32.020 And the action is going to be the action to take and should be one of the tool names. 05:32.020 --> 05:34.540 And here we're going to plug all of our tool names. 05:34.860 --> 05:38.540 And the action input is going to be the input for that action. 05:39.020 --> 05:42.820 So in our case our action is going to be get text length. 05:42.820 --> 05:46.820 And our input for it is going to be the text that we want to count the length of. 05:47.260 --> 05:52.420 And after that we have an observation which is going to be the result of the action. 05:52.660 --> 05:57.340 So this is going to be the result of the tool we ran with the tools input. 05:57.820 --> 06:03.300 Now we can see here that this thought action action input observation can repeat n times. 06:03.540 --> 06:08.980 And this is referencing the iterations of the agent when it's deciding which tool to use. 06:09.140 --> 06:13.060 Uses that tool then keeps going and doing this over and over again. 06:14.300 --> 06:17.900 So the final answer would be derived from the following format. 06:17.940 --> 06:21.980 Thought I know the final answer and final answer column. 06:22.020 --> 06:24.540 The final answer to the original input question. 06:24.820 --> 06:28.220 So here we have the output indicator begin. 06:28.580 --> 06:33.100 And our question which is an input placeholder. 06:33.340 --> 06:35.660 And then thought agent scratchpad. 06:35.940 --> 06:40.340 So in the thought agent scratchpad we're going to discuss later in this video. 06:40.660 --> 06:45.780 And from a quick analysis of this prompt we can classify this prompt as a chain of thought prompt. 06:45.980 --> 06:51.860 Because we are asking the LM to tell us how it's thinking and how it's coming up with its answer. 06:52.140 --> 06:58.460 It's also a few shot prompt because we're supplying examples and telling them, how do we want our format 06:58.500 --> 06:59.700 output to be with. 07:00.260 --> 07:04.660 And this is an implementation of the react paper Reasoning in Acting. 07:04.660 --> 07:09.420 And again, if you were interested in the theory of all of this, you can check out the theory section 07:09.420 --> 07:10.980 where I dive deep into this. 07:11.900 --> 07:12.380 Cool. 07:12.860 --> 07:17.690 So since we'll be dealing with the Agent Scratchpad only later in this video. 07:17.730 --> 07:19.210 I'm going to remove it. 07:21.810 --> 07:25.290 And this is our prompt now that we're going to use. 07:25.290 --> 07:29.170 And this is how our agent is going to select the correct tool to use. 07:30.570 --> 07:34.650 So from our template we want to create a blank chain prompt template. 07:34.890 --> 07:36.850 So I'm going to write a prompt variable. 07:36.850 --> 07:39.770 And it's going to be an instance of prompt template. 07:39.770 --> 07:42.250 And I'm going to import this link chain class. 07:42.970 --> 07:47.010 And this prompt template is going to be initialized from template function. 07:47.010 --> 07:50.490 And here I'm going to supply it and plug in the template. 07:51.170 --> 07:53.490 Now I'm going to use the partial method. 07:53.690 --> 08:02.210 And this method will populate and plug in the values of the placeholders that we already have, because 08:02.210 --> 08:04.370 we know we have the tools and the tools name. 08:04.370 --> 08:06.730 So we can plug it in into our prompt right now. 08:07.050 --> 08:12.610 And the input variable placeholder will come dynamically when we run this chain. 08:12.610 --> 08:14.610 So it's going to be from the user. 08:14.610 --> 08:19.810 So because we already know what we want to plug in in the tool names and list of tools. 08:20.130 --> 08:23.530 Then let's do it and initialize it with this partial method. 08:24.130 --> 08:27.450 So we want to initialize the tools placeholder. 08:27.730 --> 08:31.610 And here we're going to put the list of tools that we had already created. 08:31.930 --> 08:39.330 And we want also to plug in the value of tool names and tool names is going to be a comma separated 08:39.410 --> 08:40.650 tool names list. 08:40.650 --> 08:44.770 So I'm going to use the join method with the comma separator. 08:45.050 --> 08:48.370 And I'm simply going to iterate over all the tools. 08:48.370 --> 08:49.890 So for t in tools. 08:49.890 --> 08:52.170 And I'm going to access the tool name. 08:52.410 --> 08:59.450 So this eventually will print a nice string of the tool names concatenated together separated by commas. 08:59.770 --> 09:02.010 Now we're almost done with the prompt template. 09:02.010 --> 09:08.530 But in the tools keyword I can just plug in the list of tools that we gave, because this is a list 09:08.530 --> 09:16.250 of Lang tool objects, and we want here to supply a string which will represent the tools description 09:16.250 --> 09:19.530 because Lmns receive only text as input. 09:20.650 --> 09:27.970 So luckily for us, Lang Chain already implemented this logic that converts a tool list of tool objects 09:27.970 --> 09:31.050 into a string that's representing their description. 09:31.050 --> 09:36.930 So instead of passing tools, we'll call the function render text description. 09:37.770 --> 09:39.690 And let's now import it from Lang chain. 09:40.090 --> 09:43.290 And we're going to plug here the list of tools. 09:43.490 --> 09:48.970 So if we'll examine this Lang chain utility function we can see all it's doing is iterating through 09:49.010 --> 09:50.210 the tools we gave it. 09:50.410 --> 09:55.650 It's formatting nicely into a string the tool names and tool descriptions. 09:55.650 --> 09:56.370 That's it. 09:56.610 --> 10:01.610 So we're going to eventually supply a string into our LLM. 10:02.130 --> 10:06.410 So let's run it for a sanity check to see that it also runs smoothly. 10:06.850 --> 10:09.050 And we can see it worked with no errors. 10:09.370 --> 10:14.450 And now we want to define our LLM which is going to be our reasoning engine. 10:14.690 --> 10:18.930 So I'm going to use chat OpenAI the chat version of GPT 3.5. 10:19.170 --> 10:24.600 We're going to be using temperature equals to zero because we don't want any creative answers. 10:25.080 --> 10:28.720 And now we're going to plug in this top argument. 10:28.720 --> 10:32.840 And here the values are going to be backslash n observation. 10:33.080 --> 10:40.600 And so this will tell the LM to stop generating words and to finish working once it's outputted the 10:40.600 --> 10:42.640 backslash n observation token. 10:43.440 --> 10:44.680 And why do we need it. 10:44.680 --> 10:50.240 Because if we want put this top token then the LM would continue to generate text. 10:50.240 --> 10:54.720 And it's going to guess one word after another observation. 10:54.720 --> 10:56.720 And the observation is the result of the tool. 10:56.720 --> 10:59.960 And this is something that will come from running our tool. 11:00.280 --> 11:04.640 And if it comes from the LM then this is simply an hallucination. 11:04.640 --> 11:08.160 And it's not really the real time result from running the tool. 11:08.800 --> 11:14.800 Now, I know that for some of you it may sound kind of abstract, but I promise you, in the series 11:14.800 --> 11:20.360 of videos we'll be seeing this with our own eyes and you'll understand the importance of using this 11:20.360 --> 11:22.640 stop token when we use the LM.