WEBVTT 00:01.400 --> 00:02.800 Hey there, Eden here. 00:02.800 --> 00:06.920 And in this short video, we're going to be implementing the generation node. 00:06.920 --> 00:11.840 And the generation node is going to be the last node that is going to be executed. 00:12.320 --> 00:18.600 We execute this node after we already retrieve the information, the relevant documents, after we filtered 00:18.600 --> 00:25.520 out the documents that were not relevant to our query, and even performed a search for the question 00:25.520 --> 00:26.800 that we want to answer. 00:27.200 --> 00:31.920 So after we have all the documents, we can augment the original query. 00:32.200 --> 00:34.000 And now it's time to generate. 00:34.000 --> 00:39.480 So now it's time to simply stop everything and to send it to the LM to answer it. 00:39.880 --> 00:43.280 So this is what we're going to be implementing in this video. 00:43.320 --> 00:46.360 We're going to leverage the react prompt from the link chain hub. 00:46.560 --> 00:51.480 We're going to define a generation chain and to create a node that will call this chain. 00:51.480 --> 00:53.480 And of course we're going to be writing tests. 00:54.720 --> 00:56.480 All right let's go to the code. 00:56.680 --> 01:00.440 And we want to create a new file under our chains module. 01:00.600 --> 01:02.270 And we'll call it generation. 01:03.230 --> 01:05.550 And here we'll start with the imports. 01:05.550 --> 01:10.710 We want to import the hub from the chain because we're going to download the prompt from it. 01:10.950 --> 01:19.150 And we want also to import the str output parser, which is simply going to take our message. 01:19.150 --> 01:22.270 And it's going to get the content from it and turn it into a string. 01:23.590 --> 01:26.670 And we also want to import chat OpenAI. 01:27.150 --> 01:30.150 And let's create an LM instance from it. 01:30.590 --> 01:35.590 And now we want to pull the prompt from the LinkedIn hub. 01:36.030 --> 01:38.830 And this is a very standard racket prompt. 01:38.830 --> 01:46.670 That lens, smarting from the LinkedIn team, wrote, giving the LM the role of an assistant for question 01:46.670 --> 01:47.350 answering. 01:47.710 --> 01:53.990 Plugging in the context, which is going to be all the documents retrieved or web search that we saw 01:54.030 --> 01:57.470 from the earlier stages, and of course, the original question. 01:57.750 --> 02:00.030 So a very standard react prompt. 02:00.030 --> 02:01.630 So we're going to be using that. 02:02.220 --> 02:04.260 Our chain is going to be standard as well. 02:04.300 --> 02:11.100 We're going to create a generation chain where we pipe the prompt into the LM, and then we pipe the 02:11.100 --> 02:13.300 results into the STR output. 02:13.820 --> 02:20.740 So once we invoke this chain with the documents and the question, we're supposed to get the answer 02:20.740 --> 02:21.460 that we want. 02:22.860 --> 02:23.780 Alrighty. 02:23.780 --> 02:26.100 Let's go now and write some tests. 02:26.340 --> 02:31.340 So first I want to import the pretty print. 02:31.340 --> 02:34.660 And I want also to import the generation chain. 02:35.900 --> 02:41.820 And now let's go to the bottom of the file and let's create a test called test generation chain. 02:42.180 --> 02:47.500 And to be honest this isn't going to be an actual test that we are certain we're simply going to run 02:47.500 --> 02:51.780 this chain on a topic and we're going to see what's printed. 02:52.180 --> 02:56.860 It's just to give us a sanity check to see everything is working as expected. 02:56.900 --> 03:01.900 So we'll be querying agent memory and we'll retrieve the relevant documents. 03:02.260 --> 03:07.780 And now we want to run the generation chain with the context to be the documents we retrieved. 03:08.020 --> 03:09.380 And the question? 03:09.420 --> 03:10.580 The original question. 03:11.460 --> 03:12.540 And let's run it. 03:20.980 --> 03:22.660 And we can see it passed. 03:22.660 --> 03:26.380 And it gave us some summary about agent memory. 03:27.140 --> 03:29.140 So this looks fine. 03:29.620 --> 03:33.580 And we can even go to locksmith and to check out what happened. 03:33.580 --> 03:35.500 We can see we have here the retrieval. 03:35.860 --> 03:39.140 And we can see the documents that we retrieved from this query. 03:39.620 --> 03:43.100 And we can see even the documents content. 03:43.340 --> 03:45.300 And we can see it's relevant to memory. 03:45.700 --> 03:51.140 And if we go to the runnable sequence we can see the um the question that we asked. 03:51.140 --> 03:52.500 We can see it right over here. 03:52.540 --> 03:53.580 You're an assistant. 03:53.580 --> 03:58.420 And then we plugged in the context and we gave it the question of agent memory. 03:59.060 --> 04:03.810 And we got a response back which we output parsed into a string. 04:05.490 --> 04:10.970 All right, let's now run all the tests just to see that we didn't break anything. 04:18.890 --> 04:21.370 And we can see all of the tests pass. 04:21.370 --> 04:24.170 So I love seeing those green fee marks. 04:25.010 --> 04:27.530 Let's go and add a new file under nodes. 04:27.530 --> 04:29.890 And we want to call it generate. 04:30.330 --> 04:32.890 And here is going to be generate node. 04:33.130 --> 04:34.610 And you're guessing it. 04:34.610 --> 04:41.370 This node is going to simply take the question and take the documents from our state and simply run 04:41.370 --> 04:41.970 the chain. 04:42.530 --> 04:44.810 So it's going to be very straightforward. 04:46.170 --> 04:52.370 So all of the things we're doing here we already saw, but the last thing we're doing is updating the 04:52.370 --> 04:59.050 generation key in our graph state to be the generation the answer that the LLM responded.