WEBVTT 00:00.000 --> 00:01.350 -: Okay, so in this video we're gonna have a look 00:01.350 --> 00:03.120 at the chat completions endpoint 00:03.120 --> 00:05.880 versus the responses API endpoint. 00:05.880 --> 00:07.440 What I want you to do is open 00:07.440 --> 00:10.440 the chat_completions_vs_ responses_api notebook. 00:10.440 --> 00:11.880 And we're just gonna run this 00:11.880 --> 00:13.620 and have a look at the differences. 00:13.620 --> 00:17.100 So the main thing is that you've got this .create method, 00:17.100 --> 00:18.420 and it's happening on both 00:18.420 --> 00:21.630 either the chat.completions part of OpenAI, 00:21.630 --> 00:24.930 or it's happening on the responses part of OpenAI. 00:24.930 --> 00:27.780 Now, if you go and have a look at the differences, 00:27.780 --> 00:30.330 the state management for chat completions 00:30.330 --> 00:33.240 is you have a messages array or a Python list. 00:33.240 --> 00:35.730 And in every call it's completely stateless. 00:35.730 --> 00:38.910 You have to manually add the previous messages. 00:38.910 --> 00:41.790 In the responses API, by default, 00:41.790 --> 00:44.040 it is a stateful back service. 00:44.040 --> 00:47.280 So OpenAI will hold your messages 00:47.280 --> 00:50.340 and you just have to use the previous response ID 00:50.340 --> 00:51.360 inside of that. 00:51.360 --> 00:53.190 But remember, we also overcame that 00:53.190 --> 00:55.620 by setting the stories equal to false 00:55.620 --> 00:58.620 to make the response's API endpoint stateless. 00:58.620 --> 01:01.530 And then we manually optimize the message history 01:01.530 --> 01:03.930 so we could do custom things like token counting. 01:03.930 --> 01:07.200 The chat completions endpoint provides a messages array 01:07.200 --> 01:10.290 and you have to extract that with this choices list, 01:10.290 --> 01:12.240 with the responses API, 01:12.240 --> 01:15.120 you don't have an input for the messages, 01:15.120 --> 01:17.010 it's just the input parameter. 01:17.010 --> 01:20.430 So the responses API is designed for building agents 01:20.430 --> 01:23.010 and will replace the current assistance API, 01:23.010 --> 01:25.440 and it offers a smoother development experience 01:25.440 --> 01:27.960 by holding the chat context between turns. 01:27.960 --> 01:29.940 If we're looking at a minimal example, 01:29.940 --> 01:32.280 and I'm gonna basically just explain through this code 01:32.280 --> 01:33.240 so you can see what's happening. 01:33.240 --> 01:37.860 So you've got OpenAI, we've got that gpt-4.1-mini, 01:37.860 --> 01:39.420 and then we're setting up a client, 01:39.420 --> 01:41.940 in the client.chat.completions.create, 01:41.940 --> 01:44.010 everything is saying from the model's perspective, 01:44.010 --> 01:46.800 but we have this messages argument 01:46.800 --> 01:49.170 and we paste our messages in here, 01:49.170 --> 01:52.410 and then, you know, when we want to then add those 01:52.410 --> 01:56.040 to extract the messages before we do responses.choices 01:56.040 --> 01:59.610 and get this first one out, .message.content. 01:59.610 --> 02:01.380 Then we can create a follow-up request 02:01.380 --> 02:03.750 where we pass in the original conversation 02:03.750 --> 02:06.030 plus our additional message as well. 02:06.030 --> 02:07.590 Tell me another joke. 02:07.590 --> 02:10.530 And then we can also get the messages from here 02:10.530 --> 02:15.530 via that .choices[0].message.content. 02:15.540 --> 02:18.510 If we look at the responses API like we have before, 02:18.510 --> 02:20.970 you've got rid of the messages argument here 02:20.970 --> 02:23.850 for the .create, and we've replaced it with input, 02:23.850 --> 02:25.710 but the item is basically the same. 02:25.710 --> 02:29.730 It can either be a textual string, like as a prompt, 02:29.730 --> 02:32.130 or it can be a list of messages. 02:32.130 --> 02:34.830 We also have the previous response ID 02:34.830 --> 02:37.890 where basically we don't have to manually save the state 02:37.890 --> 02:39.030 of the message history, 02:39.030 --> 02:41.880 we're just using that previous response ID. 02:41.880 --> 02:45.180 So just to summarize with the chat completions API endpoint, 02:45.180 --> 02:47.730 you manually have to handle the conversation state, 02:47.730 --> 02:50.070 and it uses the messages parameter 02:50.070 --> 02:54.210 and returns an output under the choices messages.content. 02:54.210 --> 02:55.890 With the responses API, 02:55.890 --> 02:57.720 you don't have to manage message history 02:57.720 --> 02:59.130 unless you want to, 02:59.130 --> 03:01.260 and uses the input parameter 03:01.260 --> 03:03.270 and returns the aggregated output 03:03.270 --> 03:05.310 as the .output_text property. 03:05.310 --> 03:07.710 Cool. All right, I'll see you in the next video.