WEBVTT 00:01.120 --> 00:04.480 In today's video, let's clarify certain concepts. 00:04.680 --> 00:08.480 We will begin with embedding model followed by vector embedding. 00:08.760 --> 00:11.040 And then we will learn about vector search. 00:11.360 --> 00:16.040 These are the concepts that would help you understand the move framework in detail. 00:17.080 --> 00:18.720 Let's go through the background. 00:19.040 --> 00:23.640 In case of the text data military and army, they have similar meaning. 00:23.800 --> 00:27.400 Even though the words military and army are very different. 00:27.880 --> 00:35.520 For semantic search to work effectively, representation of military and army must significantly capture 00:35.520 --> 00:37.440 their semantic similarity. 00:38.240 --> 00:44.160 This is where vector representations are used and why the derivation is so important. 00:44.680 --> 00:51.840 If we want to compare two sentences, we don't want to compare the words they contain, but rather whether 00:51.840 --> 00:54.000 or not they mean the same thing. 00:54.720 --> 01:01.480 To preserve the data's meaning, we need to understand how to produce vectors where the relationship 01:01.480 --> 01:03.640 between the vectors makes sense. 01:04.240 --> 01:07.440 This is where the next obvious question comes up. 01:07.800 --> 01:09.600 Vector embedding. 01:10.160 --> 01:16.540 Embedding models are built by passing large amount of labeled data to a neural network. 01:16.860 --> 01:21.860 We then train this neural network to perform all sorts of different tasks. 01:22.380 --> 01:30.500 Embedding model processes input as natural language query and returns a multidimensional vector that 01:30.540 --> 01:33.180 represents the contextual information. 01:34.140 --> 01:40.980 Here on the left hand side, shop for table is a natural language query that is passed to an embedding 01:40.980 --> 01:48.340 model, and we get a multidimensional vector representation that represents the contextual information. 01:48.700 --> 01:52.620 In this case, it shop for data and refund fees. 01:52.940 --> 01:59.300 Now how we can use this multidimensional vector is where we cover it in the next topic, which is vector 01:59.300 --> 02:00.060 embedding. 02:00.300 --> 02:07.180 Once trained, an embedding model can transform raw data into vector embeddings that we saw in the. 02:07.220 --> 02:08.460 I'll give you the slide. 02:08.860 --> 02:14.540 A vector embedding is an array of numbers that is used to describe an object. 02:15.300 --> 02:19.220 For example, refund could have a vector here. 02:19.380 --> 02:22.220 More likely the array would be a. 02:22.260 --> 02:28.810 It's not longer than that, but this is a numeric representation of the natural language query. 02:29.290 --> 02:37.050 These vectors of numbers are then used by machine learning model to understand the meaning of the text. 02:37.450 --> 02:39.530 How does this all come together? 02:39.970 --> 02:47.090 On the left hand side, we have best places to visit as the natural language query, which is injected 02:47.090 --> 02:48.690 to the embedding model. 02:49.130 --> 02:55.690 Embedding model processes the and creates a multidimensional vectors here. 02:56.050 --> 03:03.090 What this means is all the similars matching vectors are all put together here, and a cluster and their 03:03.090 --> 03:05.130 distance would be smaller. 03:05.450 --> 03:10.810 Whereas here in this case a refund fees avoid debit card fees. 03:11.210 --> 03:18.170 They would have vector representations which are there close to each other, but they are far from the 03:18.170 --> 03:19.410 travel vectors. 03:19.650 --> 03:22.730 This is what vector embeddings means. 03:23.690 --> 03:28.210 Same applies for the dinner or cooking topic that we have here. 03:29.210 --> 03:36.750 Now the obvious next question is how does this work with the search Similarity measures are a function 03:36.750 --> 03:43.470 distance that takes two vectors as input and calculates the distance value between them. 03:43.790 --> 03:45.750 The distance can take shapes. 03:46.110 --> 03:50.030 It can be a geometric distance between two points. 03:50.430 --> 03:53.070 It could be an angle between the vectors. 03:53.470 --> 03:58.550 There are different algorithms that calculates the distance between the two vectors. 03:59.030 --> 04:07.950 Ultimately, we use the calculated distance to judge how close or far apart two vector embeddings are. 04:08.870 --> 04:15.030 In this situation, once the vectors were calculated and decided where they go on the two dimensional 04:15.030 --> 04:17.390 vector or multidimensional vector. 04:17.510 --> 04:25.510 When the user asks a query what are the best places to visit that would journey to Europe and those 04:25.510 --> 04:27.670 multidimensional vector here? 04:28.390 --> 04:33.430 When a user would ask how to make waffle, that would be closer to cook a dinner. 04:33.670 --> 04:34.910 How to make a cake. 04:35.310 --> 04:39.110 But it would be far from traveling or refunding fees.