WEBVTT 00:00.760 --> 00:04.040 Let's understand the data flow steps for evaluation. 00:04.800 --> 00:09.720 We learned this in our previous section in Knowledge base where there is a local data source. 00:10.040 --> 00:11.480 We use the PDF here. 00:12.000 --> 00:15.280 Load the PDF into AWS account that is S3 bucket. 00:15.880 --> 00:19.800 Then we perform chunk operations on the PDF or the data source. 00:20.280 --> 00:26.720 And then the embedding model would embed those chunks and store it in the vector database. 00:28.120 --> 00:34.800 When we perform a search, that search query is then embedded using embedding model and looks for the 00:34.800 --> 00:41.800 relevant content in the vector database, and then gives us the top 5 or 10 chunks that matches the 00:41.800 --> 00:42.760 search query. 00:43.800 --> 00:45.840 Depending on the embedding model. 00:46.960 --> 00:48.760 So that's the retrieval part. 00:49.560 --> 00:55.960 Let's move on to the generation part, where we provide a prompt and the search content that were retrieved 00:55.960 --> 00:57.760 to a large language model. 00:58.440 --> 01:03.240 Using the search content and the prompt it generates the response. 01:03.240 --> 01:04.800 This is the Rag pipeline. 01:07.400 --> 01:10.580 For evaluation, we need more than just the responses. 01:10.580 --> 01:12.860 What we need is evaluation data set. 01:13.900 --> 01:17.020 This is also the ground truth that we have to provide. 01:17.020 --> 01:21.580 Evaluation also needs the context that we got from this search context. 01:24.540 --> 01:26.060 We also need the prompt. 01:26.060 --> 01:29.420 So this evaluation data set would also have the prompt. 01:30.700 --> 01:34.980 And then all of this content is then passed to the prompt template. 01:37.580 --> 01:43.980 And then this prompt template here would be used and be passed to the large language model two. 01:44.700 --> 01:52.300 This model may or may not be the same as LM one, but the content that goes into those large language 01:52.300 --> 01:54.620 models are definitely different. 01:55.500 --> 02:02.860 So in the step two here we pass the context ground truth data prompt and the response from the current 02:02.860 --> 02:04.620 state into another LM. 02:05.340 --> 02:11.780 And then that large language model would generate the evaluation, which is what we will cover with 02:11.780 --> 02:14.420 a hands on activity in our next section. 02:15.180 --> 02:15.660 All right. 02:15.660 --> 02:16.660 Thank you so much. 02:16.660 --> 02:18.420 I'll see you in the next video.