WEBVTT 00:00.760 --> 00:05.760 And now we want to import the create agent function from LinkedIn. 00:08.280 --> 00:13.840 Now, in order to create an agent, we need to give the agent two minimal things. 00:14.120 --> 00:19.520 The first one is going to be tools, and the second is going to be an LM, which is going to be the 00:19.520 --> 00:20.440 reasoning engine. 00:20.600 --> 00:26.760 So let's go and import the tool decorator from LinkedIn. 00:27.960 --> 00:31.840 Let's go and import from LinkedIn core. 00:31.880 --> 00:34.600 I want to import human message. 00:34.840 --> 00:39.480 And we're going to be using human message to invoke the agent. 00:39.480 --> 00:42.760 So this is going to be the input for the agent execution. 00:43.280 --> 00:48.200 And lastly let's go and import from OpenAI chat OpenAI. 00:49.040 --> 00:53.960 And this is going to supply for our agent our LM that we're going to be using. 00:54.240 --> 01:03.710 All right so now is a good point to introduce blockchain tools and tools for the agent and tools for 01:03.710 --> 01:04.430 an LLM. 01:04.630 --> 01:06.830 And let's go and clarify this topic. 01:07.030 --> 01:07.710 All right. 01:07.710 --> 01:14.950 So a tool is a function that an agent can execute. 01:15.150 --> 01:17.710 It can be any function that we want. 01:17.750 --> 01:21.830 So we can write the internal implementation of that function. 01:22.070 --> 01:28.990 So this means we can give our agent endless possibilities of actions it can perform. 01:29.350 --> 01:35.270 So we can make it call an API to search in a database to run and execute code. 01:35.470 --> 01:37.270 We have total flexibility. 01:37.310 --> 01:38.790 What this tool can do. 01:39.110 --> 01:44.550 And once we write this tool, we can plug it into the agent and boom! 01:44.550 --> 01:49.830 Like magic, the agent has the ability to go and execute this tool. 01:50.310 --> 01:53.630 And creating a tool is actually very, very simple. 01:53.830 --> 02:01.630 All we need to do is to implement a function, a regular Python function with type hints and with some 02:01.670 --> 02:02.430 docstrings. 02:02.430 --> 02:08.670 So we'll have some description of what this function does, what arguments does it receive and what 02:08.710 --> 02:10.150 output does it output. 02:10.430 --> 02:14.030 We then use the tool decorator by link chain. 02:14.270 --> 02:21.110 This transforms this regular Python function into a tool which can be plugged in into an agent. 02:21.430 --> 02:23.670 Now I want to emphasize a couple of things. 02:23.710 --> 02:33.230 Now the type hinting the arguments, names and the docstrings of the function are very, very important 02:33.230 --> 02:35.430 because underneath the hood. 02:35.430 --> 02:40.830 And trust me, we'll be learning exactly how this works in later in this course. 02:41.030 --> 02:47.670 But underneath the hood, the large language model is going to use this description and the arguments 02:47.990 --> 02:48.790 type hints. 02:48.790 --> 02:54.670 And it's going to be using that in order to choose whether to call this tool or not, and with which 02:54.710 --> 03:01.550 arguments and state of the art models these days support this functionality using something which is 03:01.660 --> 03:06.220 called function calling, which we will be elaborating a lot on this course. 03:06.540 --> 03:14.300 Now, the gist of function calling is that instead of LMS just returning text or pictures or videos, 03:14.580 --> 03:21.980 they can also return a special response, which is a function that needs to be invoked in case it decides 03:21.980 --> 03:22.260 to. 03:22.460 --> 03:26.620 And this is an extra capability that an LM can provide us. 03:26.740 --> 03:30.460 And we are going to be treating this like a black box right now. 03:30.580 --> 03:33.060 So we're not going to be diving deep into this. 03:33.100 --> 03:35.340 We have plenty of time to do this in the course. 03:35.380 --> 03:35.900 All right. 03:35.900 --> 03:37.140 So back to the code. 03:37.180 --> 03:39.620 We are going to implement a search agent. 03:39.620 --> 03:42.020 So it's going to have a search tool. 03:42.220 --> 03:45.300 And the search tool is going to receive a query. 03:45.340 --> 03:47.780 It's going to search it in the internet. 03:47.780 --> 03:49.940 And then it's going to output the response. 03:49.940 --> 03:50.540 Right. 03:50.580 --> 03:54.900 So this is the Python signature of the tool we're going to be writing. 03:54.900 --> 03:57.820 So this is a custom tool that we're going to be writing here. 03:58.140 --> 04:02.370 And let me write here a very simple implementation. 04:02.370 --> 04:07.170 Let's simply here print the query and let's return the string. 04:07.210 --> 04:09.490 Tokyo weather is sunny right now. 04:09.490 --> 04:11.330 This is a static function. 04:11.330 --> 04:15.050 It doesn't really search over the internet, it simply returns the string. 04:15.330 --> 04:17.370 Tokyo weather is sunny. 04:17.490 --> 04:19.410 This is what this function is doing. 04:19.570 --> 04:22.970 Of course, we'll be writing the real implementation in this video. 04:23.370 --> 04:27.210 So let's go now and add to this function the doc strings. 04:27.370 --> 04:32.290 And here I have a description that this is a tool that searches over the internet. 04:32.330 --> 04:36.570 The argument is query and it returns the search result. 04:36.650 --> 04:42.490 So this is not the best description right now but it will suffice for this example. 04:42.490 --> 04:50.050 In general, the rule of thumb is that we want to make this description as explicit and unambiguous 04:50.050 --> 04:51.210 as possible. 04:51.210 --> 04:55.690 So the LLM would really have an easy time to decide whether to call this tool or not. 04:55.730 --> 04:56.290 All right. 04:56.290 --> 05:02.010 So this is now a regular Python function if we want to convert it into a tool. 05:02.080 --> 05:05.960 Let's go and decorate it with the tool decorator of chain. 05:06.240 --> 05:11.360 And this is going to convert this Python function into a link chain tool. 05:11.560 --> 05:15.200 Now what is exactly a link chain tool. 05:15.240 --> 05:18.120 This is still not clear at this point of time. 05:18.200 --> 05:20.400 Trust me, we'll get into that later. 05:20.600 --> 05:28.080 But what chain is going to do is going to take all this metadata of the tool, name the arguments that 05:28.080 --> 05:35.560 they take the description and it's going to format it nicely, and it's going to help us plug this metadata 05:35.560 --> 05:38.480 and this information to an LLM call. 05:38.480 --> 05:40.320 This is what is going to be doing under the hood. 05:40.320 --> 05:40.800 Don't worry. 05:40.840 --> 05:42.240 We're going to soon see this. 05:42.240 --> 05:42.720 All right. 05:42.720 --> 05:47.320 So assume now we have now an object which we can inject to our agent. 05:47.320 --> 05:50.200 And it can suddenly has a certain capability. 05:50.640 --> 05:52.160 So let's go and create that agent. 05:52.160 --> 05:56.160 And for that let's go and create an instance of the LLM. 05:56.440 --> 05:59.400 And I'll initialize a chat OpenAI object. 05:59.640 --> 06:02.590 And let me now create a list of tools. 06:02.590 --> 06:05.790 So it's going to be a list containing the search tool here. 06:06.190 --> 06:09.670 And let's go now and create the agent. 06:09.950 --> 06:12.310 So let's define a variable. 06:12.310 --> 06:14.430 We'll call this variable agent. 06:15.310 --> 06:19.430 We're going to give it as an input our model. 06:19.430 --> 06:27.670 So our LM and we're going to provide it with the list of tools which is a list with a single item which 06:27.670 --> 06:29.070 has the search tool. 06:29.390 --> 06:29.870 Boom. 06:29.910 --> 06:30.470 That's it. 06:30.510 --> 06:36.270 We have now an agent ready to go, ready to run with a search capability. 06:36.470 --> 06:36.830 Right. 06:36.870 --> 06:38.070 So it has an LM. 06:38.230 --> 06:45.550 It has a list of tools which contains only one tool, which is the search tool that is allegedly supposed 06:45.550 --> 06:49.350 to do some online search, but currently it's only going to return us. 06:49.390 --> 06:56.950 Tokyo weather is sunny here, so let's go and try to run a certain query with the agent right now. 06:57.150 --> 06:57.630 All right. 06:57.630 --> 06:59.870 So this agent is actually a runnable. 06:59.870 --> 07:02.950 So let me write a result variable and let's go. 07:02.950 --> 07:04.550 Right agent dot invoke. 07:05.070 --> 07:12.350 And in order to run the agent, we need to give it a dictionary with the messages field. 07:12.670 --> 07:20.110 And the messages field should be a list of messages for the agent which is going to contain the input. 07:20.110 --> 07:21.950 So this is the user's query. 07:22.230 --> 07:29.910 So here under messages I'm going to be sending the human message alongside with the content of what's 07:29.910 --> 07:31.750 the weather in Tokyo. 07:31.910 --> 07:35.190 Now notice here that the key here is messages. 07:35.190 --> 07:38.270 But I did not provide here a list of messages. 07:38.270 --> 07:43.670 I just provided one message, one human message, not inside the list. 07:43.870 --> 07:50.550 And this is okay because link chain under the hood is going to cast this into a list containing a single 07:50.550 --> 07:52.590 element of a human message. 07:52.590 --> 07:55.270 So this is why it should still be working. 07:55.270 --> 07:58.870 If you want, you can put here this inside of a list. 07:59.270 --> 08:02.510 So let me go now and print here the result.