WEBVTT 00:00.440 --> 00:01.040 Hey there. 00:01.080 --> 00:01.880 Eden here. 00:01.920 --> 00:04.720 I know the previous video was a long video. 00:05.040 --> 00:08.560 However, luckily for us, this video is much shorter. 00:08.840 --> 00:12.400 In this video, we'll be implementing the Reviser agent. 00:12.640 --> 00:18.840 Now, since we have most of the infrastructure ready, then all we need to do is to simply add an instruction 00:18.840 --> 00:21.840 to the prompt and create a new class for the response. 00:21.840 --> 00:22.760 And that's it. 00:22.960 --> 00:25.880 This agent will take the last artifact written. 00:25.920 --> 00:31.920 Use the critique in order to revise the artifact and then output us a better article. 00:33.000 --> 00:36.720 All right, so I'm going back to the code to change that. 00:37.400 --> 00:40.400 And we're going to add here a new template. 00:40.560 --> 00:42.840 Now those are the revision instructions. 00:42.840 --> 00:44.560 So let's take a look at them. 00:45.000 --> 00:48.200 Revise your previous answer using the new information. 00:48.240 --> 00:52.120 You should use the previous critique to add important information to your answer. 00:52.320 --> 00:57.520 You must include numerical citations in your revised answer to ensure it can be verified. 00:57.840 --> 01:03.640 Add a reference section to the bottom of your answer, which does not count towards the word limit in 01:03.640 --> 01:04.520 the form of. 01:04.560 --> 01:06.060 And then there are URLs. 01:06.300 --> 01:12.860 You should use the previous critique to remove superfluous information from your answer, to make sure 01:12.860 --> 01:15.460 it doesn't go over 250 words. 01:16.060 --> 01:24.540 So this instruction is going to be plugged in into our original actor prompt template in line number 01:24.580 --> 01:25.540 23. 01:25.540 --> 01:32.020 So in the placeholder of First Instruction we're going to plug in the revision instructions. 01:32.380 --> 01:34.580 Let's head up to schema.py. 01:34.940 --> 01:38.900 We want now to create a new class for our revisited answer. 01:39.140 --> 01:41.220 And we'll call it Revise answer. 01:41.420 --> 01:45.020 And notice it will inherit from the answer creation class. 01:45.220 --> 01:48.260 So this means it will have all of its fields. 01:48.260 --> 01:51.260 So it will have answer reflection search queries. 01:51.460 --> 01:54.460 However it will have also something else. 01:54.660 --> 02:00.460 It will also have the references field which is going to be a list of strings. 02:00.460 --> 02:07.340 And those strings are going to be citations of URLs mostly that we'll get from the search engine. 02:07.700 --> 02:11.880 Now we haven't discussed what's going to happen with the search engine and how it's going to work. 02:11.880 --> 02:14.480 Don't worry, we're going to do it in the next video. 02:14.840 --> 02:17.320 Alrighty, let's head back to Chains pie. 02:17.560 --> 02:20.920 And now we want to write the revision chain. 02:21.160 --> 02:27.080 So after we wrote the revision instructions then we simply want to create a new chain very similar to 02:27.120 --> 02:28.160 what we did before. 02:28.560 --> 02:30.840 And we'll take the actor prompt template. 02:30.880 --> 02:37.680 We'll fill up the first instruction placeholder with the revised instructions variable, and we simply 02:37.680 --> 02:43.640 want to pipe it into our LM, GPT four turbo again using function calling. 02:43.640 --> 02:47.280 So we'll supply the tools which is the revised answer tool. 02:47.560 --> 02:51.200 And we're going to give the tool choice to be revised answer. 02:51.200 --> 02:57.840 So this would enforce the schema of the pedantic object revise answer class and would de facto make 02:57.840 --> 03:02.200 our LM adhere and grab the answer to be that kind of object. 03:03.040 --> 03:05.560 So let's also import that from schemas. 03:06.360 --> 03:10.160 And we can see in line 15 we also imported it. 03:10.400 --> 03:11.920 And that's it for the reviser. 03:11.920 --> 03:16.210 So this is the logic which is going to run in the revision node. 03:16.850 --> 03:17.650 Alrighty. 03:17.650 --> 03:22.610 So let's have a recap of what we did so far in this video with the revision node. 03:22.650 --> 03:29.930 We wrote the Revisor chain, and in the revision step or Revisor node, we're going to take the result 03:29.930 --> 03:35.970 we got from the search engine from Tavli with the relevant search queries, and that we're going to 03:36.010 --> 03:42.010 plug in into our already existing article alongside with the generated critique. 03:43.210 --> 03:49.690 And the agent is going to revise the response projected from the critique that was already written, 03:49.930 --> 03:55.970 is going to take the search results from Tavli, and it's going to add this into our response. 03:55.970 --> 04:02.050 And of course, it's going to citate in all the resources that we used from the internet. 04:03.210 --> 04:11.210 In the next video, we'll be dealing with the tool executions, and we'll handle all of the tavli web 04:11.250 --> 04:16.890 searching for relevant topics, and we're going to propagate that into our revisited answer.