WEBVTT 00:00.240 --> 00:02.490 -: And in this video, we're gonna have a look at agents. 00:02.490 --> 00:05.910 Agents are a specific type of workflow 00:05.910 --> 00:08.730 where the LLM will run in a continuous loop, 00:08.730 --> 00:11.160 interacting with its environments and receiving feedback 00:11.160 --> 00:13.470 to refine its actions and decisions. 00:13.470 --> 00:17.280 Agents use tools and then they go through an agentic loop. 00:17.280 --> 00:20.700 They receive, after making an action, 00:20.700 --> 00:23.490 they'll get some feedback from the environment. 00:23.490 --> 00:26.373 After a point in time, the agent will decide to stop. 00:34.110 --> 00:35.820 Some of the use cases for agents 00:35.820 --> 00:38.310 include personal research assistants, 00:38.310 --> 00:40.200 automated code reviewing, 00:40.200 --> 00:42.030 customer service agents is a big one. 00:42.030 --> 00:43.740 And also managing social media presence 00:43.740 --> 00:47.160 by analyzing trending topics, generating relevant content 00:47.160 --> 00:48.543 and scheduling posts. 00:50.040 --> 00:53.340 Okay, let's start by importing OpenAI and Pydantic. 00:53.340 --> 00:55.560 We'll also make sure that we have the JSON, OS 00:55.560 --> 00:58.350 and OpenAI packages installed and imported. 00:58.350 --> 01:00.360 And we'll then also make sure that the client 01:00.360 --> 01:02.103 and model have been created. 01:03.540 --> 01:04.620 The first thing we're gonna do 01:04.620 --> 01:07.683 is we're gonna create a search_knowledge_base function. 01:10.591 --> 01:13.758 (instructor coughing) 01:21.080 --> 01:22.782 The first thing we're gonna do, 01:22.782 --> 01:26.532 (instructor clearing throat) 01:27.834 --> 01:30.917 (instructor gulping) 01:32.151 --> 01:33.979 the first thing we're gonna do, 01:33.979 --> 01:34.812 the first thing we're gonna do 01:34.812 --> 01:36.900 is create a search_knowledge_base function. 01:36.900 --> 01:39.030 And this is just gonna be a fake function 01:39.030 --> 01:41.400 that we're gonna take a query, which is a string. 01:41.400 --> 01:43.050 We're going to return a string 01:43.050 --> 01:47.400 and all we're gonna put in here is return summary for, 01:47.400 --> 01:48.840 and then we'll put query 01:48.840 --> 01:53.837 and then we'll put, this is a simulated summary 01:56.160 --> 01:57.453 from the knowledge base. 01:58.500 --> 02:00.180 Okay, cool, now that we have that function, 02:00.180 --> 02:02.580 the next thing we need to determine is tools. 02:02.580 --> 02:05.370 So obviously tools allow the agent to take actions. 02:05.370 --> 02:08.070 Now, you will have to write JSON schema out. 02:08.070 --> 02:13.070 So the first thing is you have to have a type of function. 02:13.080 --> 02:14.850 And then after that, then the next thing 02:14.850 --> 02:17.823 that we need to do is we need to step into that function. 02:19.200 --> 02:20.590 So we do function 02:21.960 --> 02:23.100 and then we specifically have 02:23.100 --> 02:25.170 to start listing out the individual things. 02:25.170 --> 02:27.030 So for example, what is the name of that function? 02:27.030 --> 02:28.710 Well, it's search_knowledge_base. 02:28.710 --> 02:30.750 What is the description? 02:30.750 --> 02:35.040 Query a knowledge base to retrieve relevant info on a topic. 02:35.040 --> 02:38.340 And you can see it's populated the parameters required 02:38.340 --> 02:39.360 and additional properties. 02:39.360 --> 02:41.310 And I'm gonna get rid of this one here. 02:42.761 --> 02:43.740 And there we go. 02:43.740 --> 02:47.670 So we've got our function, which basically is a schema. 02:47.670 --> 02:50.520 This is a JSON schema definition of a tool. 02:50.520 --> 02:52.110 Now we need to generate some messages. 02:52.110 --> 02:53.590 So we'll have a messages list 02:54.840 --> 02:57.510 and we'll have a role of user and content. 02:57.510 --> 02:59.970 I'd like to know more about quantum computing. 02:59.970 --> 03:02.580 Could you give me a quick summary? 03:02.580 --> 03:05.250 Then we create a chat completions request 03:05.250 --> 03:08.340 with client.chat.completions.create, 03:08.340 --> 03:11.040 and we pass in the tools alongside the model 03:11.040 --> 03:12.150 and the messages. 03:12.150 --> 03:16.063 Now, what you'll see here is it has a .tool_calls. 03:17.040 --> 03:18.840 And basically what this is gonna mean 03:18.840 --> 03:21.630 is that ChatGPT wants to make a tool call. 03:21.630 --> 03:23.430 And the idea of that is this, 03:23.430 --> 03:26.220 and this is the function with the arguments 03:26.220 --> 03:28.380 and the name of the function that it wants to call. 03:28.380 --> 03:31.170 So we can tell ChatGPT, these are the tool_calls. 03:31.170 --> 03:33.210 So let's just do that manually to start with. 03:33.210 --> 03:38.210 So we'll say tool_calls = completion_1.choices, 03:39.028 --> 03:42.300 [0].message.tool_calls. 03:42.300 --> 03:43.620 And then we can get those tool_calls out 03:43.620 --> 03:44.460 and just have a look at them. 03:44.460 --> 03:45.900 So pasting in tool_calls here, 03:45.900 --> 03:48.870 you can see that we have a ChatCompletionMessageToolCall, 03:48.870 --> 03:50.910 which provides a tool ID 03:50.910 --> 03:54.180 and also the function with the arguments of the function 03:54.180 --> 03:55.410 and the name of the function. 03:55.410 --> 03:57.240 Now, the first thing we're gonna do is make sure 03:57.240 --> 04:01.500 that the messages also has the most recent message 04:01.500 --> 04:04.320 specifically from the chat completion_1. 04:04.320 --> 04:07.904 So we'll just do completion_1.choices, 04:07.904 --> 04:09.900 [0].message. 04:09.900 --> 04:11.760 And then that just basically means 04:11.760 --> 04:16.230 that this messages will now have this additional one 04:16.230 --> 04:19.170 in here, which is the ChatCompletionMessage. 04:19.170 --> 04:21.630 So now how can we access these tool_calls? 04:21.630 --> 04:22.463 We're gonna do a for loop. 04:22.463 --> 04:24.480 So we'll save for tool_call in tool_calls, 04:24.480 --> 04:26.400 we're gonna get the function name, 04:26.400 --> 04:31.110 which is the tool_call.function.name. 04:31.110 --> 04:32.790 And we're also gonna get the function args, 04:32.790 --> 04:34.590 and then we're gonna have something like this where we say 04:34.590 --> 04:37.410 if the function name is the search_knowledge_base, 04:37.410 --> 04:40.500 we're gonna execute that function with here. 04:40.500 --> 04:42.300 And then we're gonna add the messages 04:42.300 --> 04:45.180 and we just need to add a role of tool, 04:45.180 --> 04:47.220 the tool_call_id and the result. 04:47.220 --> 04:49.680 And if it's not this function, then we're gonna go 04:49.680 --> 04:51.210 and just raise a ValueError. 04:51.210 --> 04:53.130 So I'm just gonna raise a ValueError 04:53.130 --> 04:56.163 and then we're gonna have a look at the messages as well. 04:57.120 --> 04:59.010 And now you can see here we've got the original message. 04:59.010 --> 05:01.530 We've got this ChatCompletionMessage in the middle, 05:01.530 --> 05:04.500 and we also have the tool_call_id that specifically responds 05:04.500 --> 05:06.750 to this one here, the first tool_call. 05:06.750 --> 05:09.270 Now we can create the completion_2 request 05:09.270 --> 05:11.430 and then passing in the original messages and tools, 05:11.430 --> 05:13.580 and looking at the content of that message. 05:15.390 --> 05:18.150 And it says quantum area computing is an area of computing. 05:18.150 --> 05:20.640 So it is basically decided this is the final answer. 05:20.640 --> 05:23.070 So it executed that search on the knowledge base 05:23.070 --> 05:25.440 and then it hasn't come up with any more tool_calls 05:25.440 --> 05:28.530 and therefore, it's given us a final answer here. 05:28.530 --> 05:29.880 You can confirm that is the case 05:29.880 --> 05:34.230 by looking at the completion_2.choices[]0.message.tool_calls 05:34.230 --> 05:35.490 and it is essentially empty. 05:35.490 --> 05:37.440 Now, there is an alternative way to write this 05:37.440 --> 05:39.330 that is a little bit better in my opinion. 05:39.330 --> 05:40.950 So you have a messages list 05:40.950 --> 05:42.885 and we're just gonna write it in a while truth. 05:42.885 --> 05:46.210 So we'll say messages.append 05:47.220 --> 05:48.750 and then we're gonna basically 05:48.750 --> 05:51.960 just have some information, "role": "user". 05:51.960 --> 05:54.660 And then instead of asking about quantum computing, 05:54.660 --> 05:57.170 we're going to ask about ChatGPT. 05:57.170 --> 06:01.350 So we'll say, can you find some information 06:01.350 --> 06:04.623 about ChatGPT and the AI knowledge base? 06:05.520 --> 06:07.710 We're also gonna create a function just above this. 06:07.710 --> 06:10.320 So we'll say def search_knowledge_base. 06:10.320 --> 06:12.750 And we're gonna say some information about ChatGPT 06:12.750 --> 06:13.740 just to mimic this. 06:13.740 --> 06:14.730 So we'll say ChatGTP 06:14.730 --> 06:17.250 is a large language model developed by OpenAI. 06:17.250 --> 06:18.540 And then what we're gonna do 06:18.540 --> 06:20.730 is we're gonna set up our while true loop. 06:20.730 --> 06:23.160 And then this is where you can basically have all 06:23.160 --> 06:25.680 of your agentic code running in a while true loop. 06:25.680 --> 06:27.750 So you start with your first completion 06:27.750 --> 06:28.583 and then what you're doing 06:28.583 --> 06:32.760 is saying client.chat.completions.create. 06:32.760 --> 06:35.340 And then after that, you paste in the model, 06:35.340 --> 06:37.293 we paste in the messages, 06:39.657 --> 06:41.550 and we're also pasting in the tools. 06:41.550 --> 06:44.130 We then always paste back in the message. 06:44.130 --> 06:45.540 if there aren't tool_calls, 06:45.540 --> 06:48.060 we can then break out of the for loop. 06:48.060 --> 06:50.850 If there is a tool_call, then we get the function name 06:50.850 --> 06:52.170 and the arguments. 06:52.170 --> 06:56.130 And then we also get the result of that tool_call. 06:56.130 --> 07:00.120 We pass in the role, the tool_call_id and the content. 07:00.120 --> 07:03.180 And then what you'll see is then basically this will run 07:03.180 --> 07:06.810 in a while true loop and it finishes after 2.7 seconds. 07:06.810 --> 07:09.090 So it actually does finish 07:09.090 --> 07:11.070 because the reason why is it decided 07:11.070 --> 07:12.690 that there weren't any tool calls 07:12.690 --> 07:15.093 and therefore, all of the agentic work was done. 07:15.960 --> 07:17.550 So you can have a look and you can see we've got 07:17.550 --> 07:18.930 that initial first message. 07:18.930 --> 07:20.586 Then we have our ChatCompletionMessage 07:20.586 --> 07:22.530 that did ask to do a tool_call. 07:22.530 --> 07:23.790 We then did that tool_call. 07:23.790 --> 07:25.440 We got some information about the fact 07:25.440 --> 07:28.140 that ChatGPT is a large language model 07:28.140 --> 07:31.440 and then also, it decided to use that information 07:31.440 --> 07:32.550 to then respond. 07:32.550 --> 07:33.720 So the key point here 07:33.720 --> 07:36.900 is that you are basically stuck in a while true loop 07:36.900 --> 07:38.970 and you also have the ability 07:38.970 --> 07:42.030 to break out if there are no more tool_calls to be made. 07:42.030 --> 07:43.890 Now, just to give you a bit of a word of warning, 07:43.890 --> 07:45.780 the problem that you can run into with this 07:45.780 --> 07:47.640 is that it endlessly makes tool_calls, 07:47.640 --> 07:50.790 trying to search or do additional steps. 07:50.790 --> 07:52.950 So a nice quality of life feature 07:52.950 --> 07:55.080 is to add a maximum number of iterations. 07:55.080 --> 07:58.890 So you can say maximum number of iterations, 07:58.890 --> 08:00.510 and you can just put maybe 10. 08:00.510 --> 08:03.060 And then we can say we can adjust the while true loop. 08:03.060 --> 08:04.440 So I'm just gonna use cursor to do this. 08:04.440 --> 08:06.940 So we'll say adjust the while true loop 08:07.800 --> 08:12.030 to work only up to 10 iterations. 08:12.030 --> 08:14.010 And so that's just gonna say every single time 08:14.010 --> 08:16.800 we do an iteration, we're basically just gonna add on one 08:16.800 --> 08:18.180 to the iteration variable. 08:18.180 --> 08:20.790 And then after that, then we make sure that we don't hit 08:20.790 --> 08:22.140 that number of iterations, 08:22.140 --> 08:24.240 and then we can track how many iterations we make 08:24.240 --> 08:25.080 in real time. 08:25.080 --> 08:27.750 So this can be like another good way of figuring out, 08:27.750 --> 08:30.360 so we made two iterations, the zero and then the one. 08:30.360 --> 08:31.890 And this can be a very good way 08:31.890 --> 08:33.690 of avoiding endless while loops. 08:33.690 --> 08:35.520 Something else that I also want you to consider 08:35.520 --> 08:36.990 is you might also not want 08:36.990 --> 08:39.360 to break when there are no tool_calls. 08:39.360 --> 08:41.460 You might want to break on a specific goal. 08:41.460 --> 08:44.220 So imagine you're trying to get a certain number of leads 08:44.220 --> 08:46.320 or a certain number of email addresses scraped. 08:46.320 --> 08:49.950 You could have that as your goal inside of the while loop 08:49.950 --> 08:53.490 and you could endlessly scan new areas of Google 08:53.490 --> 08:55.980 to specifically go and look for that information. 08:55.980 --> 08:57.600 I actually created a demo version 08:57.600 --> 08:59.880 of this in an AI-powered outreach assistant. 08:59.880 --> 09:02.310 So I'll put a link to the public repository 09:02.310 --> 09:03.540 for you to go and have a look at. 09:03.540 --> 09:05.070 But if we scroll down and have a look 09:05.070 --> 09:06.690 at the agentic while loop, 09:06.690 --> 09:08.670 you can see this is where we start to do this. 09:08.670 --> 09:10.560 So we have our while true loop. 09:10.560 --> 09:13.470 We also then decide okay, if we've gone past the time, 09:13.470 --> 09:14.760 we break out of the loop 09:14.760 --> 09:16.170 and if we've hit our objective, 09:16.170 --> 09:18.030 we also break out of the loop. 09:18.030 --> 09:20.100 So those are some easy ways to make sure 09:20.100 --> 09:21.960 that we always break out of the loop. 09:21.960 --> 09:24.150 And also, the other thing you'll want to be aware of 09:24.150 --> 09:27.240 is when you're running agents for a large amount of time, 09:27.240 --> 09:29.220 the amount of chat message histories 09:29.220 --> 09:32.250 that you get will become exponentially more expensive. 09:32.250 --> 09:34.500 So you might want to prune the messages, 09:34.500 --> 09:37.860 but you must keep the tool pairs, the tool_call 09:37.860 --> 09:41.340 and the tool_call result paired messages next to each other 09:41.340 --> 09:44.100 otherwise OpenAI will throw an error at you. 09:44.100 --> 09:45.540 So make sure that if you are going 09:45.540 --> 09:47.010 to prune the message history, 09:47.010 --> 09:50.460 have a look at the helpers.py inside of here 09:50.460 --> 09:52.800 where you can see there is some advanced pruning 09:52.800 --> 09:55.391 so that we can specifically always keep the system message 09:55.391 --> 10:00.060 and have these remainder pairs of the tool_call_id 10:00.060 --> 10:03.240 with the tool_call result chat message history. 10:03.240 --> 10:05.490 Cool, hopefully this is a good intro into agents 10:05.490 --> 10:08.153 and feel free to let me know if you've got any questions.