1
00:00:05,160 --> 00:00:11,370
In this lesson, we are going to talk about the components of the rack technique.

2
00:00:14,160 --> 00:00:20,970
So remember that the main purpose of the drag technique is to overcome the limits of the context window.

3
00:00:21,660 --> 00:00:26,340
A the steps of the technique are very simple.

4
00:00:26,640 --> 00:00:35,820
A first we load in the data, uh, like remember, for example, in the seq uh insights application,

5
00:00:35,820 --> 00:00:41,940
when we loaded the financial documents from a couple of companies Apple and Amazon.

6
00:00:41,940 --> 00:00:42,360
Right.

7
00:00:42,360 --> 00:00:48,900
So we load these documents and then we apply the rack technique.

8
00:00:49,020 --> 00:00:53,610
And the rack technique has um, um, mostly three steps.

9
00:00:53,610 --> 00:01:00,960
The first step is to divide the data into small segments, small chunks.

10
00:01:01,230 --> 00:01:07,290
The second step is to convert the small segments into numbers.

11
00:01:08,820 --> 00:01:12,930
We convert these small segments of text into numbers.

12
00:01:12,930 --> 00:01:19,200
Because computers and vector databases work much better with numbers than with text.

13
00:01:20,200 --> 00:01:30,010
So when we convert the small segments of text into numbers and we call these numbers embeddings.

14
00:01:30,010 --> 00:01:37,090
And the third step of the technique is to load the embeddings into a vector database.

15
00:01:37,720 --> 00:01:43,000
So these are the three initial steps of the RAC technique.

16
00:01:43,300 --> 00:01:46,210
Divide the data into small segments.

17
00:01:46,210 --> 00:01:54,940
Convert the small segments into numbers, and load the embeddings into a vector database.

18
00:01:56,110 --> 00:02:06,760
Now when the user asks anything about the data, what the LM application does is to go to the vector

19
00:02:06,760 --> 00:02:15,340
database and to search for data that only answer the question of the user.

20
00:02:15,760 --> 00:02:24,550
So what we are going to do here is to use a technique called semantic similarity search.

21
00:02:24,550 --> 00:02:27,790
We will learn more about that later.

22
00:02:27,910 --> 00:02:28,840
So.

23
00:02:29,990 --> 00:02:36,890
The rack technique is the essential technique for most LM applications.

24
00:02:36,890 --> 00:02:39,170
The essential the core technique.

25
00:02:39,170 --> 00:02:43,820
That's why we will focus on mastering this technique.

26
00:02:44,090 --> 00:02:53,120
Apart from embeddings, vector databases, and semantic similarity search, we will be learning more

27
00:02:53,120 --> 00:02:59,360
about queries, indexing, orchestration frameworks, etc. but let's.

28
00:02:59,980 --> 00:03:05,770
Learn a little bit more about embeddings and databases now.

29
00:03:05,950 --> 00:03:08,890
So embeddings a.

30
00:03:10,320 --> 00:03:18,300
Ah, if you remember the result to convert small segments of text into numbers.

31
00:03:18,300 --> 00:03:26,730
So remember, computers work better with numbers, and that's why they convert text into numbers.

32
00:03:26,730 --> 00:03:35,340
They do the same with images, audio, video, etc. embeddings are more than numbers.

33
00:03:35,340 --> 00:03:38,490
They are vectors of numbers.

34
00:03:38,610 --> 00:03:48,360
For example, the word hello is converted into an embedding like one comma four comma six.

35
00:03:48,360 --> 00:03:49,650
This is a vector.

36
00:03:49,650 --> 00:03:51,390
This is just an example.

37
00:03:51,390 --> 00:03:56,070
So embeddings are vectors of numbers.

38
00:03:56,070 --> 00:04:03,210
We use embeddings because computers and vector databases are much faster working with numbers than with

39
00:04:03,210 --> 00:04:03,870
text.

40
00:04:03,870 --> 00:04:08,640
A little bit about vector databases.

41
00:04:08,940 --> 00:04:18,269
So vector databases are specialized in working with hundreds of millions of embeddings.

42
00:04:18,690 --> 00:04:23,430
They are much faster than conventional databases.

43
00:04:23,430 --> 00:04:29,250
They are optimized for storing, indexing, and retrieving.

44
00:04:30,780 --> 00:04:38,010
Semantic similarity is a search technique in the vector databases.

45
00:04:38,010 --> 00:04:44,910
So vector databases group embeddings by their semantic similarity.

46
00:04:44,910 --> 00:04:54,300
For example, the embeddings, the embeddings of dog and cat, which are semantically similar as both

47
00:04:54,300 --> 00:04:59,760
are animals, will be grouped together in the vector database.

48
00:04:59,940 --> 00:05:10,080
So using the semantic similarity search vector databases are only going to search for data that are

49
00:05:10,080 --> 00:05:13,530
similar to the question of the user.

50
00:05:15,110 --> 00:05:21,710
So in this lesson we have been started talking about the rack components.

51
00:05:21,710 --> 00:05:26,300
It's just a theory theory theoretical introduction.

52
00:05:26,300 --> 00:05:33,650
But I think it's important for us to become familiar with the main concepts and con and components of

53
00:05:33,650 --> 00:05:37,280
the rack technique before going into practice.

54
00:05:37,910 --> 00:05:45,050
In the next lesson, we are going to talk briefly about the main challenges of the rack technique.

55
00:05:45,050 --> 00:05:48,320
This is important to know at this point.