WEBVTT 1 00:00:00.260 --> 00:00:02.580 Let's learn how embeddings work. 2 00:00:03.400 --> 00:00:08.760 So embeddings are what is called semantic meaning 3 00:00:08.760 --> 00:00:10.560 and similarity searches. 4 00:00:11.080 --> 00:00:14.340 So it's not like it's doing the old 5 00:00:14.340 --> 00:00:18.780 style search of if the two words are 6 00:00:18.780 --> 00:00:21.460 the same or fuzzy search and so on. 7 00:00:21.560 --> 00:00:24.120 It's more a semantic search where we have 8 00:00:24.120 --> 00:00:25.120 AI involved. 9 00:00:26.420 --> 00:00:29.620 So in a normal search, a kitten would 10 00:00:29.620 --> 00:00:31.820 not be close to a cat because it 11 00:00:31.820 --> 00:00:36.300 has totally different length, totally different letters and 12 00:00:36.300 --> 00:00:36.840 so on. 13 00:00:37.440 --> 00:00:40.700 But in a vector store where you had 14 00:00:40.700 --> 00:00:46.260 words like wolf, dog, cat, banana, apple, the 15 00:00:46.260 --> 00:00:48.860 word kitten would go close to a cat. 16 00:00:50.780 --> 00:00:54.920 And behind the scenes, such vectors have 1536 17 00:00:54.920 --> 00:00:56.320 different dimensions. 18 00:00:57.580 --> 00:00:59.760 In our world, we of course have the 19 00:00:59.760 --> 00:01:01.720 three or the four if you take time. 20 00:01:02.360 --> 00:01:05.700 But in mathematics, you can have infinite amount 21 00:01:05.700 --> 00:01:08.220 of vectors and dimensions. 22 00:01:09.380 --> 00:01:12.160 So that is how this works. 23 00:01:12.720 --> 00:01:15.320 This is a bit abstract, so we are 24 00:01:15.320 --> 00:01:17.160 going to go into some code and actually 25 00:01:17.160 --> 00:01:19.660 show a little on how we make these 26 00:01:19.660 --> 00:01:22.160 vectors, how they look like, and also how 27 00:01:22.160 --> 00:01:25.740 we check that something fits together. 28 00:01:28.540 --> 00:01:31.780 So in here, we have an embedding data 29 00:01:31.780 --> 00:01:33.060 sample. 30 00:01:34.080 --> 00:01:37.240 And we're using the normal client here, but 31 00:01:37.240 --> 00:01:39.720 now we're not going to new up a 32 00:01:39.720 --> 00:01:40.940 chat client agent. 33 00:01:41.060 --> 00:01:43.160 We don't use it at all in this 34 00:01:43.160 --> 00:01:43.760 scenario. 35 00:01:43.760 --> 00:01:47.120 Instead, we're taking what is called an embedding 36 00:01:47.120 --> 00:01:47.520 client. 37 00:01:49.000 --> 00:01:51.780 And in OpenAI's terms, one of the best 38 00:01:51.780 --> 00:01:54.100 they have is text embedding tree small. 39 00:01:54.760 --> 00:01:57.220 You also have a larger one, but not 40 00:01:57.220 --> 00:02:02.500 all vector stores can support such big embeddings, 41 00:02:02.660 --> 00:02:04.200 so I tend to use this one. 42 00:02:07.699 --> 00:02:10.740 So if we make one of these embedding 43 00:02:10.740 --> 00:02:16.400 generators, we can actually embed this sentence. 44 00:02:16.600 --> 00:02:18.140 In this case, I've just made a little 45 00:02:18.140 --> 00:02:20.820 Q&A part of our Wi-Fi data. 46 00:02:20.940 --> 00:02:22.340 What is the Wi-Fi password at the 47 00:02:22.340 --> 00:02:22.700 office? 48 00:02:23.280 --> 00:02:25.720 And the office password, it's guest 42. 49 00:02:27.160 --> 00:02:29.840 And if we use this embedding generator, just 50 00:02:29.840 --> 00:02:34.340 say generate this text and turn it into 51 00:02:34.340 --> 00:02:34.820 a vector. 52 00:02:36.620 --> 00:02:39.920 So this actually sent up a signal to 53 00:02:39.920 --> 00:02:42.260 OpenAI and got back. 54 00:02:43.040 --> 00:02:47.020 So embeddings are rather fast and also rather 55 00:02:47.020 --> 00:02:51.120 cheap, way cheaper than normal LLM tokens in 56 00:02:51.120 --> 00:02:53.320 terms of cost and speed. 57 00:02:54.200 --> 00:02:57.720 So what we get back is a vector 58 00:02:57.720 --> 00:03:01.040 and we get how when it was created, 59 00:03:01.300 --> 00:03:03.320 what model did it, and then we get 60 00:03:03.320 --> 00:03:04.100 the vector back. 61 00:03:04.200 --> 00:03:07.000 And the vector is just a long series 62 00:03:07.000 --> 00:03:07.900 of numbers. 63 00:03:08.960 --> 00:03:11.320 And I will not claim to understand this 64 00:03:11.320 --> 00:03:15.580 fully because I'm not a mathematician, but we 65 00:03:15.580 --> 00:03:18.480 can technically write it out if we want 66 00:03:18.480 --> 00:03:20.480 to and see how it looks. 67 00:03:21.800 --> 00:03:24.780 And it doesn't give too much meaning for 68 00:03:24.780 --> 00:03:24.960 us. 69 00:03:25.060 --> 00:03:28.640 It's just 1536 of these numbers back and 70 00:03:28.640 --> 00:03:31.420 forward in various different dimensions. 71 00:03:31.740 --> 00:03:33.460 How will it fit that dimension? 72 00:03:36.940 --> 00:03:39.820 This is a rather small embedding. 73 00:03:40.460 --> 00:03:42.600 We can also do a larger embedding here. 74 00:03:42.700 --> 00:03:44.540 I have the first chapter of write and 75 00:03:44.540 --> 00:03:48.020 produce that I can turn into a vector. 76 00:03:49.140 --> 00:03:58.220 And all vectors become these 1536 dimensions. 77 00:03:58.980 --> 00:04:03.020 So it's not like this vector I create 78 00:04:03.020 --> 00:04:07.220 here is any bigger or any smaller than 79 00:04:07.220 --> 00:04:08.360 the one we just made. 80 00:04:08.440 --> 00:04:11.340 It has the same things here. 81 00:04:11.820 --> 00:04:15.180 But what happens is the smaller we make 82 00:04:15.180 --> 00:04:18.860 it, the less defined it will be to 83 00:04:18.860 --> 00:04:21.980 know what area this is in, what area 84 00:04:21.980 --> 00:04:22.680 this is in. 85 00:04:23.120 --> 00:04:26.040 And the bigger, the more it needs to 86 00:04:26.040 --> 00:04:29.340 compress the data and lose some of its 87 00:04:29.340 --> 00:04:29.700 meaning. 88 00:04:30.340 --> 00:04:34.520 So I would say a good vector is 89 00:04:34.520 --> 00:04:37.060 roughly in the middle of such things like 90 00:04:37.060 --> 00:04:37.320 this. 91 00:04:37.400 --> 00:04:40.980 This might be okay in terms of size. 92 00:04:41.780 --> 00:04:43.460 But the more you put into a vector, 93 00:04:43.640 --> 00:04:46.180 the more it will lose its value. 94 00:04:47.760 --> 00:04:49.720 But you can also, if you just put 95 00:04:49.720 --> 00:04:51.720 in every single word into a vector, it 96 00:04:51.720 --> 00:04:54.200 wouldn't know what to do with that because 97 00:04:54.200 --> 00:04:56.520 it will just be too generic for anything. 98 00:04:58.120 --> 00:05:00.560 Vectors can technically be very long. 99 00:05:00.940 --> 00:05:02.840 You can put in up to 6000 words 100 00:05:02.840 --> 00:05:05.220 roughly because they're counted in tokens. 101 00:05:05.340 --> 00:05:09.740 We cannot really know exactly precisely, but that's 102 00:05:09.740 --> 00:05:12.420 roughly 10 to 15 pages if you wish 103 00:05:12.420 --> 00:05:12.620 to. 104 00:05:13.640 --> 00:05:16.780 But figuring out how big a vector should 105 00:05:16.780 --> 00:05:19.060 be, should it be one sentence at a 106 00:05:19.060 --> 00:05:21.000 time, should it be one paragraph at a 107 00:05:21.000 --> 00:05:23.540 time, should it be one chapter at a 108 00:05:23.540 --> 00:05:25.980 time, a page, and so on, is a 109 00:05:25.980 --> 00:05:31.320 whole science that we can talk, we could 110 00:05:31.320 --> 00:05:34.800 make a course just about that by itself. 111 00:05:35.120 --> 00:05:38.160 And I'm no expert, but I'm slowly learning 112 00:05:38.160 --> 00:05:39.380 as we go along. 113 00:05:40.680 --> 00:05:43.160 What we can do is, of course, we 114 00:05:43.160 --> 00:05:45.280 can go in and show that same vector 115 00:05:45.280 --> 00:05:48.860 and we'll just see a whole lot of 116 00:05:48.860 --> 00:05:53.560 numbers again with different values in terms of 117 00:05:53.560 --> 00:05:55.040 the different dimensions. 118 00:05:55.620 --> 00:05:57.780 But again, it's not like just because it 119 00:05:57.780 --> 00:05:59.920 was a longer text that it's a longer 120 00:05:59.920 --> 00:06:00.380 vector. 121 00:06:00.920 --> 00:06:03.060 They will be exactly the same length. 122 00:06:06.680 --> 00:06:10.920 So now we have some vectors and I've 123 00:06:10.920 --> 00:06:13.980 just asked ChatGBT to make me a co 124 00:06:13.980 --> 00:06:19.900 -similarity search score that can compare two vectors 125 00:06:19.900 --> 00:06:21.600 to each other and see how close they 126 00:06:21.600 --> 00:06:22.140 are together. 127 00:06:22.360 --> 00:06:24.520 Because that's how things work. 128 00:06:25.780 --> 00:06:28.480 So I won't claim that I understand this 129 00:06:28.480 --> 00:06:31.820 code specifically, but it's going in and figuring 130 00:06:31.820 --> 00:06:34.740 out how close are two vectors to each 131 00:06:34.740 --> 00:06:38.900 other and giving a value between zero, meaning 132 00:06:38.900 --> 00:06:41.040 it had absolutely nothing to do with each 133 00:06:41.040 --> 00:06:43.700 other, and one being it was the exact 134 00:06:43.700 --> 00:06:47.580 same sentence that was vectorised. 135 00:06:48.620 --> 00:06:51.840 So if we go in and ask, how 136 00:06:51.840 --> 00:06:56.280 close are our Wi-Fi data with the 137 00:06:56.280 --> 00:06:56.860 book data? 138 00:06:57.400 --> 00:07:01.580 And hopefully you will intuitively know that in 139 00:07:01.580 --> 00:07:04.040 private purchase, they probably don't talk a lot 140 00:07:04.040 --> 00:07:06.840 about Wi-Fi and offices and so on. 141 00:07:07.020 --> 00:07:09.020 So they're probably not too similar. 142 00:07:10.340 --> 00:07:13.440 And what we see is that that's correct. 143 00:07:13.600 --> 00:07:16.120 It's like 0.10 out of 1. 144 00:07:16.860 --> 00:07:18.160 So a fairly low score. 145 00:07:20.600 --> 00:07:23.100 But if we want to check a bit 146 00:07:23.100 --> 00:07:27.000 more realistic, we can make a small vector 147 00:07:27.000 --> 00:07:30.600 here that says, how can my customer use 148 00:07:30.600 --> 00:07:31.220 the Wi-Fi? 149 00:07:32.180 --> 00:07:35.380 This is what a real person would perhaps 150 00:07:35.380 --> 00:07:36.180 ask for. 151 00:07:36.180 --> 00:07:38.300 They wouldn't know to ask, what is the 152 00:07:38.300 --> 00:07:39.840 Wi-Fi for the office and so on. 153 00:07:39.920 --> 00:07:42.840 They would just, I have a customer at 154 00:07:42.840 --> 00:07:44.640 the office at the moment. 155 00:07:44.760 --> 00:07:46.400 They need to get onto the Wi-Fi. 156 00:07:47.280 --> 00:07:49.020 They don't specify the word guest. 157 00:07:49.220 --> 00:07:52.360 They don't specify office and anything. 158 00:07:53.140 --> 00:07:55.600 So we can make a vector of that 159 00:07:55.600 --> 00:07:59.640 and check the similarity search between our Wi 160 00:07:59.640 --> 00:08:00.800 -Fi data and the question. 161 00:08:01.440 --> 00:08:04.380 And here we get a better score, 0 162 00:08:04.380 --> 00:08:08.840 .45, which is a rather okay score, but 163 00:08:08.840 --> 00:08:09.880 of course it's not a one. 164 00:08:10.480 --> 00:08:12.820 But again, a one would never really happen, 165 00:08:13.420 --> 00:08:17.720 because if that were to happen, the end 166 00:08:17.720 --> 00:08:20.600 user would have written exactly this sentence with 167 00:08:20.600 --> 00:08:23.240 colons in exactly the same places and so 168 00:08:23.240 --> 00:08:23.480 on. 169 00:08:24.380 --> 00:08:28.120 So a 45 is a pretty okay, and 170 00:08:28.120 --> 00:08:30.180 if we only had these two, this would 171 00:08:30.180 --> 00:08:31.600 of course be the best match. 172 00:08:32.919 --> 00:08:35.059 We can ask in a different way, what 173 00:08:35.059 --> 00:08:36.720 is the office Wi-Fi? 174 00:08:37.580 --> 00:08:41.120 And if we do that, we get an 175 00:08:41.120 --> 00:08:44.840 even better score, because there's more similarity of 176 00:08:44.840 --> 00:08:48.740 this sentence compared to our original Q&A 177 00:08:48.740 --> 00:08:51.520 in terms of this question. 178 00:08:53.380 --> 00:08:56.600 But again, if the book is totally different, 179 00:08:57.000 --> 00:09:01.960 so we had these three, then this one 180 00:09:01.960 --> 00:09:04.380 would be the best way of curing to 181 00:09:04.380 --> 00:09:05.800 get our right information. 182 00:09:07.660 --> 00:09:11.180 So it's all about these numbers, and in 183 00:09:11.180 --> 00:09:13.200 reality, we are not sitting and doing the 184 00:09:13.200 --> 00:09:14.480 match score ourselves. 185 00:09:14.800 --> 00:09:18.940 That's the job of a vector store, and 186 00:09:18.940 --> 00:09:21.240 we are just doing here a sample. 187 00:09:21.920 --> 00:09:25.140 I've never in real life written code like 188 00:09:25.140 --> 00:09:28.220 this, but in order for us to quickly 189 00:09:28.220 --> 00:09:30.720 see that we have different scores for different 190 00:09:30.720 --> 00:09:34.160 vectors when they're compared, it's nice to know. 191 00:09:34.820 --> 00:09:36.480 But again, when we go to the vector 192 00:09:36.480 --> 00:09:40.280 store in the next lecture, we will let 193 00:09:40.280 --> 00:09:42.520 that do this similarity score. 194 00:09:43.440 --> 00:09:46.160 In this case, we're using cosine similarity, where 195 00:09:46.160 --> 00:09:48.320 it's between 0 and 1. 196 00:09:49.280 --> 00:09:51.840 Others use from minus 1 to 1. 197 00:09:52.280 --> 00:09:54.160 Some use from 0 to 2. 198 00:09:54.580 --> 00:09:55.800 So there's different ways. 199 00:09:55.900 --> 00:09:57.760 So you need to check on your vector 200 00:09:57.760 --> 00:10:01.520 store what is a good score, because sometimes 201 00:10:01.520 --> 00:10:03.360 a negative score is a good score. 202 00:10:04.520 --> 00:10:08.020 That's all depending on the implementation of how 203 00:10:08.020 --> 00:10:10.820 the mathematician have said, this is the best 204 00:10:10.820 --> 00:10:13.360 way for me to check two vectors together. 205 00:10:15.180 --> 00:10:17.740 But now we should be ready to go 206 00:10:17.740 --> 00:10:18.940 to our vector store part.