1 00:00:04,960 --> 00:00:09,730 In this lesson we are going to talk about rack optimization. 2 00:00:16,309 --> 00:00:24,290 If you remember, in previous, uh, lessons, we have been talking about a rag a lot, and we said 3 00:00:24,290 --> 00:00:28,010 that there is a basic approach, like a basic. 4 00:00:28,910 --> 00:00:37,340 Range of techniques in order to apply the rack technique in your prototype or initial demo, and then 5 00:00:37,340 --> 00:00:49,100 you will have a many other advanced techniques in order to optimize your rack application for a professional 6 00:00:49,100 --> 00:00:52,280 a for a professional stage. 7 00:00:53,000 --> 00:01:00,380 So we have already been talking about some of the optimization techniques you need to consider. 8 00:01:00,380 --> 00:01:01,370 In this lesson. 9 00:01:01,370 --> 00:01:11,570 We are going to focus our attention in a couple of areas where you are going to see a probably when 10 00:01:11,570 --> 00:01:21,590 you study this, a this, this topic, uh, further, you are going to find a different terminologies, 11 00:01:21,590 --> 00:01:25,220 uh, uh, new names, new ways of doing things. 12 00:01:25,220 --> 00:01:30,680 And we are not going to, uh, explain all of them here in detail. 13 00:01:30,680 --> 00:01:37,280 We just want to give you a quick idea of the things you are going to find and the things that can be 14 00:01:37,280 --> 00:01:47,150 interesting for you to study in a more advanced stage when you are, uh, optimizing your application 15 00:01:47,150 --> 00:01:53,390 or when you are preparing the launch of your or your of your professional LM application. 16 00:01:53,390 --> 00:02:02,780 So one of the areas that we wanted to, uh, explore a little bit more with you is the area of the alternative 17 00:02:02,780 --> 00:02:06,500 techniques to index and store vectors. 18 00:02:06,500 --> 00:02:19,820 So remember that this is refer to how we a how we, uh, save and how we search data in vector databases, 19 00:02:19,820 --> 00:02:23,270 which, as you know, is an important part of the rack technique. 20 00:02:23,270 --> 00:02:32,660 So in this particular step, storing and indexing vectors, you may see a different terminology like 21 00:02:32,690 --> 00:02:43,730 hierarchy here hierarchical navigable small small world difficult time I repeat hierarchical navigable 22 00:02:43,730 --> 00:02:45,440 small world. 23 00:02:46,600 --> 00:02:47,620 Sometimes you will. 24 00:02:47,620 --> 00:02:53,650 You will see just the the NSW letters. 25 00:02:54,650 --> 00:02:58,250 This is one of the techniques for storing and indexing vectors. 26 00:02:58,250 --> 00:03:03,650 You may find as the advanced technique you would like to explore. 27 00:03:03,740 --> 00:03:09,620 You have another one like Locally Sensitive hashing or LSH. 28 00:03:10,870 --> 00:03:14,740 You have product quantization or PQ. 29 00:03:15,370 --> 00:03:21,640 And finally you have filtering pre and post filtering and random projections. 30 00:03:21,820 --> 00:03:28,270 All these some different advanced techniques for storing and indexing vectors. 31 00:03:28,270 --> 00:03:34,000 So and they are not all the the the ones you have available for you. 32 00:03:34,000 --> 00:03:35,440 These are the most popular. 33 00:03:35,440 --> 00:03:41,200 So these are the kind of techniques you will have to explore study and practice. 34 00:03:41,200 --> 00:03:47,260 And in many cases just just try and see, you know, the results of each of these techniques in your, 35 00:03:47,470 --> 00:03:54,610 uh, professional application in order to see how each of them impacts the performance of the application. 36 00:03:54,610 --> 00:03:58,660 In many cases, it's just a a case of trial and error. 37 00:03:58,660 --> 00:04:05,440 Try and compare, you know, the different techniques and see the impact in your application. 38 00:04:05,440 --> 00:04:14,260 The second area where you will probably explore different alternatives is the area, uh, when you are 39 00:04:14,260 --> 00:04:20,560 measuring the vector similarity or the search similarity, semantic similarity. 40 00:04:21,130 --> 00:04:31,090 So remember that the semantic similarity is the main technique to search data in vector databases. 41 00:04:31,090 --> 00:04:37,120 So for a for this purpose we have different ways of doing things. 42 00:04:37,120 --> 00:04:41,560 So we have different ways of doing semantic similarity. 43 00:04:41,560 --> 00:04:44,020 The most common the most popular. 44 00:04:44,020 --> 00:04:53,260 The ones that you will probably try first are cosine similarity and the second one Euclidean distance. 45 00:04:54,290 --> 00:04:59,540 So they are both popular in the math world. 46 00:04:59,570 --> 00:05:06,890 The cosine similarity measures the angle between vectors and the Euclidean distance measures. 47 00:05:06,890 --> 00:05:11,180 You know the straight line between items, between vectors as well. 48 00:05:11,180 --> 00:05:15,350 But don't worry about about, uh, math right now. 49 00:05:15,350 --> 00:05:21,830 This is this is something that you may find in the literature, you know, in the articles or the courses 50 00:05:21,830 --> 00:05:24,050 or advanced videos you may find about that. 51 00:05:24,050 --> 00:05:31,250 The important thing right now is for you to understand that when you are working with a Rag application, 52 00:05:31,250 --> 00:05:36,920 which is going to be very frequent when you are, uh, working in the professional world, you have 53 00:05:36,920 --> 00:05:44,270 two main stages the basic development stage of the Rag technique and then the advanced one, where you 54 00:05:44,270 --> 00:05:52,580 are going to try different approaches in order to optimize the results that the basic approach, uh, 55 00:05:52,580 --> 00:05:55,250 initially, initially gave you. 56 00:05:55,250 --> 00:05:55,820 Okay. 57 00:05:55,820 --> 00:06:05,030 So this is just a quick review of two areas where you probably would like to, uh, explore and different 58 00:06:05,030 --> 00:06:06,110 alternatives. 59 00:06:08,540 --> 00:06:15,950 In the next lesson, we are going to talk about a very important a question regarding LLM applications, 60 00:06:15,950 --> 00:06:18,080 which is the speed. 61 00:06:18,410 --> 00:06:22,340 And in some cases this is known as latency. 62 00:06:22,460 --> 00:06:27,170 We will see the relationship between speed and latency in the next lesson.