1
00:00:00,700 --> 00:00:08,370
In the last video, we developed a basic
RAC system using Lankchain, OpenAI and Pinecon.

2
00:00:09,360 --> 00:00:14,710
RAC stands for Retrieval Augmented
Generation and is a technique that

3
00:00:14,720 --> 00:00:18,590
combines an LLM with a way to search for information.

4
00:00:19,640 --> 00:00:22,770
This lets the model look at stuff like

5
00:00:22,780 --> 00:00:26,810
private documents or databases while its
generating text.

6
00:00:28,890 --> 00:00:35,740
RAC helps overcome knowledge limits,
makes answers more factual and lets the

7
00:00:35,750 --> 00:00:37,740
model handle complex questions.

8
00:00:38,930 --> 00:00:41,420
We used Pinecon which is a powerful

9
00:00:41,430 --> 00:00:44,740
vector store great for big systems and apps.

10
00:00:45,750 --> 00:00:47,920
In this video, I ll show you an

11
00:00:47,930 --> 00:00:50,100
alternative to Pinecon called Chroma.

12
00:00:50,810 --> 00:00:53,100
Chroma is an open source in -memory

13
00:00:53,110 --> 00:00:59,080
vector store making it a better fit for
small to medium -sized projects, you

14
00:00:59,090 --> 00:01:04,360
might not need a separate server or
hosting for it, plus you don t have to

15
00:01:04,370 --> 00:01:09,620
mess around with indexes and namespaces
which saves time and effort.

16
00:01:10,810 --> 00:01:17,260
To install Chroma, just run pip install
-q chroma -db.

17
00:01:22,700 --> 00:01:28,190
Finally, I cleaned up the code a bit by
getting rid of stuff outside the

18
00:01:28,200 --> 00:01:31,550
functions I have defined, these functions.

19
00:01:35,000 --> 00:01:38,170
I am defining a new function that creates

20
00:01:38,180 --> 00:01:45,130
the embeddings using the OpenAIEmbeddings
class, saves them in a Chroma database

21
00:01:45,140 --> 00:01:51,630
and returns the database, df
createEmbeddings Chroma.

22
00:01:58,730 --> 00:02:05,960
The parameters are the chunks and a
persistent directory used for storage,

23
00:02:09,670 --> 00:02:14,020
its default value will be chroma -db in
the current directory.

24
00:02:24,040 --> 00:02:32,180
From langchain -vector -stores, I am
importing Chroma and from langchain

25
00:02:32,190 --> 00:02:49,520
-OpenAI, I am importing OpenAIEmbeddings,
I am instantiating an embedding model to

26
00:02:49,530 --> 00:02:52,220
convert text to numerical representations.

27
00:02:54,850 --> 00:03:01,460
Embeddings equals OpenAIEmbeddings of

28
00:03:01,470 --> 00:03:13,000
model equals and the embedding model,
text -embedding -rismol and the

29
00:03:13,010 --> 00:03:19,860
dimensions equals 1536.

30
00:03:23,220 --> 00:03:25,830
Now I am creating a ChromaVectorStore

31
00:03:25,840 --> 00:03:30,830
object, storing the embeddings and
linking them back to the original text,

32
00:03:31,480 --> 00:03:40,930
so vectorStore equals Chroma from
documents and the arguments, chunks,

33
00:03:42,160 --> 00:03:50,930
embeddings and persistent directory
equals persistent directory, the

34
00:03:50,940 --> 00:03:51,850
functions parameter.

35
00:03:54,040 --> 00:04:01,900
I am returning the vectorStore, that's it

36
00:04:01,910 --> 00:04:02,800
for this function.

37
00:04:02,810 --> 00:04:09,780
I am also defining a function, that loads

38
00:04:09,790 --> 00:04:18,120
the existing embeddings from disk, to a
vectorStore object, def

39
00:04:18,130 --> 00:04:27,080
loadEmbeddingsChroma and the function
will have one argument, the persistent directory.

40
00:04:30,670 --> 00:04:33,380
It will be by default, chroma -db.

41
00:04:37,930 --> 00:04:41,960
I am copy -pasting the first three lines

42
00:04:41,970 --> 00:04:44,580
of code from the previous function.

43
00:04:49,120 --> 00:04:51,810
I am loading the existing Chroma database

44
00:04:51,820 --> 00:04:58,150
from the persistent directory, specifying
the embedding, so vectorStore equals

45
00:04:58,160 --> 00:05:09,130
Chroma of persistent directory equals
persistent directory and the embedding

46
00:05:09,140 --> 00:05:20,260
function equals embeddings, the object, I
have just defined one line of code above.

47
00:05:20,270 --> 00:05:35,640
I am returning the vectorStore, I am
running the code, let's load a PDF file,

48
00:05:41,300 --> 00:05:46,970
this is about REX systems powered by
Google search technology, it is worth

49
00:05:46,980 --> 00:05:53,970
reading the document and you will find it
in the resources, let's get back to code.

50
00:05:54,580 --> 00:06:05,050
Data equals loadDocument of and the path
to the files and the reg powered by

51
00:06:05,060 --> 00:06:06,450
Google search .pdf.

52
00:06:08,400 --> 00:06:10,790
I am splitting the document into chunks,

53
00:06:11,420 --> 00:06:21,840
chunks equals chunkData of data and the
chunkSize equals 256.

54
00:06:21,850 --> 00:06:29,880
This is the default value, anyway,
finally, I am creating the vectorStore,

55
00:06:30,330 --> 00:06:38,120
vectorStore equals
createEmbeddingsChroma, I am calling this

56
00:06:38,130 --> 00:06:57,700
function and the functions argument are
the chunks, I am running this code, I

57
00:06:57,710 --> 00:07:04,540
noticed a small mistake, the argument of
the function from documents is persist,

58
00:07:04,770 --> 00:07:15,860
not persistent, so I will change
everywhere to persist directory, I am

59
00:07:15,870 --> 00:07:27,260
running the code, the data from the PDF
was loaded, chunked and embedded.

60
00:07:28,130 --> 00:07:42,800
Let's ask the first question, Q equals
what is vertexAsearch and I will call the

61
00:07:42,810 --> 00:07:50,930
function askAndGetAnswer, the one I have
defined in the previous lecture, answer

62
00:07:50,940 --> 00:07:58,510
equals askAndGetAnswer of vectorStore and
Q, the question.

63
00:07:59,280 --> 00:08:07,760
And I am printing the answer, I am
running the code, I got a good answer,

64
00:08:08,290 --> 00:08:17,720
that matched what was in the document, in
the PDF document about regs powered by

65
00:08:17,730 --> 00:08:28,590
Google search, if I ask childGPT the same
question, what is vertexAsearch, it will

66
00:08:28,600 --> 00:08:35,310
say something like this, as of my last
update in January 2022, vertexAsearch is

67
00:08:35,320 --> 00:08:42,310
not a specific product or service
provided by Google, that is how it is

68
00:08:42,320 --> 00:08:48,930
supposed to work, the application is
pulling the answer from the PDF, not its

69
00:08:48,940 --> 00:08:49,710
general knowledge.

70
00:08:52,400 --> 00:08:55,380
Let's test the function that loads the

71
00:08:55,390 --> 00:09:00,660
embeddings from disk into a vectorStore
object, the loadEmbeddingsChroma

72
00:09:00,670 --> 00:09:12,480
function, db equals loadEmbeddingsChroma,
I will use the default value for the

73
00:09:12,490 --> 00:09:22,760
persist directory, ChromaDb in the
current directory, this one, and the

74
00:09:22,770 --> 00:09:31,240
question, how many pairs of questions and
answers had the StackOverflow dataset, I

75
00:09:31,250 --> 00:09:39,960
am calling the askAndGetAnswer function
and printing the answer, I am running the

76
00:09:39,970 --> 00:09:47,780
code, alright, I got the good answer, the
StackOverflow dataset had 8 million pairs

77
00:09:47,790 --> 00:09:59,980
of questions and answers, let's check the
answer in the PDF, ok, the answer was

78
00:09:59,990 --> 00:10:06,540
correct, the StackOverflow dataset
contained 8 million pairs of questions

79
00:10:06,550 --> 00:10:13,700
and answers, that is how it works with
Chroma, however, there is a drawback, if

80
00:10:13,710 --> 00:10:18,840
I want to send a follow -up question, it
will not have access to the previous chat

81
00:10:18,850 --> 00:10:27,470
history and will respond that it doesn't
know the context, for example, Q equals,

82
00:10:28,440 --> 00:10:39,620
multiply that number by 2, the number is
the one from the last question, 8 million

83
00:10:39,630 --> 00:10:54,790
pairs, it answered, I don't have a
specific number to multiply by 2.

84
00:10:56,240 --> 00:11:01,990
Let's take a quick break and in the next
video we will learn how to add memory to

85
00:11:02,000 --> 00:11:08,110
our rack system, this way it will
remember our old questions and we can ask

86
00:11:08,120 --> 00:11:08,950
follow -ups.