WEBVTT 00:00.160 --> 00:00.760 Hey there. 00:00.800 --> 00:01.440 Eden here. 00:01.440 --> 00:06.440 And in this video, we'll be implementing the chains that will run in our graph. 00:06.720 --> 00:12.640 So we'll be implementing the generation chain which is going to be responsible to generate and revise 00:12.680 --> 00:15.200 our tweets until it's better and better again. 00:15.360 --> 00:18.040 And we're going to be implementing the reflection chain. 00:18.040 --> 00:21.040 So that chain is going to take our tweet. 00:21.040 --> 00:22.760 And it's going to give us feedback on it. 00:22.800 --> 00:24.000 It's going to critique it. 00:24.000 --> 00:28.480 And it's going to give us suggestions on how to make it better and better. 00:28.720 --> 00:35.960 And every iteration of our cycle in our graph, we're going to be feeding this critique into our generation 00:35.960 --> 00:36.360 chain. 00:36.360 --> 00:38.680 And we're going to be revising the prompt. 00:38.720 --> 00:44.120 So first, before we build the graph we need to build the components that will run inside it. 00:44.440 --> 00:46.160 So let's go and write some code. 00:47.440 --> 00:51.520 Now it's time to create a new file and we'll call it chains. 00:52.160 --> 00:58.840 So this file will hold all of our prompts and chains that we're going to be using in our graph graph. 01:00.080 --> 01:01.880 So let's start with the imports. 01:01.920 --> 01:06.560 So we'll first import chat prompt template and messages placeholder. 01:06.960 --> 01:12.900 Just a reminder the chat prompt template is going to hold our content that we either sent to the LLM 01:12.940 --> 01:19.220 as humans, or that we receive back from the LLM as an answer that is tagged as an AI. 01:21.020 --> 01:28.380 And the second class is the messages placeholder, which is going to give us flexibility to put here 01:28.380 --> 01:31.820 a placeholder for future messages that we're going to get. 01:33.740 --> 01:36.300 The second import is going to be chat open AI. 01:37.140 --> 01:37.780 Alrighty. 01:37.780 --> 01:41.700 So now it's time to write our prompts and chains. 01:42.100 --> 01:48.140 The first prompt we're going to write is going to be called the reflection prompt in our agent architecture. 01:48.420 --> 01:51.500 Then this prompt is supposed to act as our critique. 01:51.740 --> 01:53.980 So it's going to review the output. 01:53.980 --> 01:58.660 And in this case is a Twitter post and it's going to criticize it. 01:58.660 --> 02:03.460 So it's going to say how it can be better and a suggestion to improve it. 02:04.020 --> 02:04.900 And let's have a look. 02:04.900 --> 02:06.620 So we have here a system message. 02:07.020 --> 02:09.060 And you can really tell the following. 02:09.580 --> 02:11.700 You're a viral Twitter influencer. 02:11.740 --> 02:12.740 Creating a tweet. 02:12.740 --> 02:16.220 Generate critique and recommendation for the user's tweet. 02:16.380 --> 02:22.140 Always provide detailed recommendations including requests for length, virality, style, etc.. 02:22.980 --> 02:26.650 Now we want to put here a placeholder for other messages. 02:26.650 --> 02:32.170 And those are going to be the history messages that our agent is going to invoke and to criticize and 02:32.170 --> 02:34.370 get recommendations over and over again. 02:34.650 --> 02:40.210 So here you can see that we put a messages placeholder and the variable name is going to be messages. 02:40.210 --> 02:46.690 So when we initialize the reflection point we're going to plug in to this prompt in the messages. 02:46.690 --> 02:49.810 And it's going to contain a lot of messages of our history. 02:50.250 --> 02:52.730 So this was the reflection prompt. 02:52.770 --> 02:55.490 Now we're going to be writing the generation prompt. 02:55.930 --> 03:02.450 And in our agent architecture, the generation prompt is going to generate the tweets that are going 03:02.450 --> 03:08.370 to be revised over and over again after the feedback we get from the reflection prompt. 03:08.570 --> 03:12.890 So it's going to revise the tweet until it gets the perfect tweet. 03:13.610 --> 03:16.370 So let's paste this baby and review it. 03:16.890 --> 03:19.010 And this is the message is the following. 03:19.250 --> 03:24.890 You are a Twitter techie influencer assistant tasked with writing excellent Twitter posts. 03:24.930 --> 03:28.570 Generate the best Twitter posts possible for the user's request. 03:28.610 --> 03:36.230 If the user provides critique, respond with a revised version of your previous attempts And this is 03:36.230 --> 03:39.470 going to be our generation prompt or creator prompt. 03:39.470 --> 03:46.390 You can name it however you like, but it's a goal over here is to revise the tweet for it to be perfect 03:46.390 --> 03:48.230 based on the feedback it gets. 03:48.830 --> 03:55.430 And we also want to put here a placeholder for all the reflections and revisions that we had earlier. 03:55.430 --> 04:00.510 So it's going to be with the messages placeholder with the key of messages that we're going to plug 04:00.510 --> 04:00.750 in. 04:02.190 --> 04:02.830 All right. 04:02.830 --> 04:04.870 So let's go now and create our chains. 04:04.870 --> 04:07.310 So for that we need to initialize an LM. 04:07.550 --> 04:12.110 And this initialization will default to use GPT 3.5 turbo. 04:13.110 --> 04:15.830 And let's create now two simple chains. 04:15.830 --> 04:18.870 And we'll be using the link chain expression language for that. 04:19.590 --> 04:22.470 So the first chain is going to be the generation chain. 04:22.750 --> 04:26.670 And we're simply going to pipe the generation prompt into the LM. 04:27.350 --> 04:30.630 And the second chain is going to be our reflection chain. 04:30.630 --> 04:33.630 And we're going to pipe it into the LM as well. 04:34.310 --> 04:36.350 So that's pretty much it for this video. 04:36.350 --> 04:40.990 And in the next video we'll be implementing finally our land graph. 04:40.990 --> 04:41.590 Graph.