1
00:00:00,700 --> 00:00:06,590
We ve seen that LLMs lack the ability to
solve complex calculations, so we are

2
00:00:06,600 --> 00:00:08,090
going to take another approach.

3
00:00:08,960 --> 00:00:11,570
Instead of having the LLM generate the

4
00:00:11,580 --> 00:00:16,710
answer directly, which is difficult, we
ll instruct the LLM to generate the

5
00:00:16,720 --> 00:00:22,030
Python code to calculate the answer and
then run that code in a Python

6
00:00:22,040 --> 00:00:24,310
interpreter to get the answer.

7
00:00:24,320 --> 00:00:27,950
I am installing a package called

8
00:00:27,960 --> 00:00:33,790
langchain -experimental which contains
experimental components and features

9
00:00:33,800 --> 00:00:41,490
related to Langchain, pip install -q
-langchain -experimental.

10
00:00:45,000 --> 00:00:45,710
I am running it.

11
00:00:46,040 --> 00:00:49,870
It s important to note that this library

12
00:00:49,880 --> 00:00:56,830
is intended for research and exploration
and some parts are not yet stable or

13
00:00:56,840 --> 00:01:02,770
secure for production use, this is just a
demonstration of the Langchain capabilities.

14
00:01:03,840 --> 00:01:06,450
Let s see how to use Python -Apple.

15
00:01:06,780 --> 00:01:09,710
I am importing the Python -Apple class

16
00:01:09,720 --> 00:01:16,540
from the Langchain -Experimental
-Utilities module from Langchain

17
00:01:16,550 --> 00:01:23,640
-Experimental .utilities import Python -Apple.

18
00:01:26,050 --> 00:01:29,340
I am creating an instance of the Python

19
00:01:29,350 --> 00:01:33,400
-Apple class which provides a way to
execute Python code.

20
00:01:34,410 --> 00:01:40,540
Python underline Apple equals and I am
calling the constructor of the class.

21
00:01:41,270 --> 00:01:48,060
Now, I can execute a Python expression
within the Apple environment like this.

22
00:01:48,930 --> 00:01:54,920
Python -Apple .run and the Python
expression as a string.

23
00:01:55,970 --> 00:02:02,900
For example, this line of Python code
takes a list with all numbers divisible

24
00:02:02,910 --> 00:02:10,100
by 13 between 1 and 99 using list
comprehension and prints the list.

25
00:02:10,870 --> 00:02:11,520
I am running it.

26
00:02:12,530 --> 00:02:15,820
In this example, there is no LLM involved.

27
00:02:16,450 --> 00:02:18,780
Let s go ahead and see how to create an agent.

28
00:02:19,450 --> 00:02:22,800
Agents are what make Langchain so useful.

29
00:02:23,470 --> 00:02:26,660
Agents combine Langchain tools and chains.

30
00:02:27,910 --> 00:02:30,260
I am importing a function called

31
00:02:30,270 --> 00:02:35,280
createPythonAgent from the Experimental
-Agent toolkits.

32
00:02:37,980 --> 00:02:45,530
This function is used to create a special
type of Langchain agent that can interact

33
00:02:45,540 --> 00:02:46,530
with Python code.

34
00:02:47,340 --> 00:02:50,330
Next, I will import the Python -Apple

35
00:02:50,340 --> 00:02:51,030
tool class.

36
00:02:52,180 --> 00:02:54,890
This tool provides a Python -Apple

37
00:02:54,900 --> 00:03:00,850
environment within Langchain allowing the
agent to execute Python code directly.

38
00:03:01,960 --> 00:03:06,570
Finally, I will import chatOpenAi to
create the LLM.

39
00:03:09,120 --> 00:03:10,790
I am creating the LLM.

40
00:03:11,380 --> 00:03:13,390
LLM equals chatOpenAi.

41
00:03:13,820 --> 00:03:18,170
I will use gpt for turbo because we need
a powerful model.

42
00:03:19,240 --> 00:03:30,040
Model equals gpt for turbo preview and
the temperature equals 0.

43
00:03:33,650 --> 00:03:40,000
I am instantiating a Python agent
executor using the openAi LLM.

44
00:03:40,630 --> 00:03:44,860
This allows us to have the language model
execute Python code.

45
00:03:46,030 --> 00:03:55,230
AgentExecutor equals createPythonAgent
and the arguments.

46
00:03:57,040 --> 00:04:07,880
LLM equals LLM, tool equals Python -Apple
tool and the verbose equals to.

47
00:04:10,920 --> 00:04:16,430
In the function createPythonAgent, there
is an argument called tool.

48
00:04:17,120 --> 00:04:19,030
Let s talk about this for a moment.

49
00:04:19,880 --> 00:04:23,050
Tools are essentially functions that

50
00:04:23,060 --> 00:04:26,830
agents can use to interact with the
outside world.

51
00:04:27,540 --> 00:04:32,950
These can range from general utilities
such as a math calculator or a search

52
00:04:32,960 --> 00:04:36,350
function to other chase or even other agents.

53
00:04:38,610 --> 00:04:40,800
I will execute the Python agent.

54
00:04:41,350 --> 00:04:45,540
The problem to solve is given in natural
language format.

55
00:04:46,950 --> 00:04:51,240
AgentExecutor .invoke and a task.

56
00:04:53,360 --> 00:04:55,510
Calculate the square root of the

57
00:04:55,520 --> 00:04:59,690
factorial of 12 and display it with 4
decimal points.

58
00:05:00,860 --> 00:05:01,730
I am running the code.

59
00:05:04,680 --> 00:05:08,130
You are now looking at the process of how

60
00:05:08,140 --> 00:05:10,550
an LLM generates Python code.

61
00:05:16,810 --> 00:05:21,680
The LLM first makes a plan, then an

62
00:05:21,690 --> 00:05:25,280
observation and finally a thought.

63
00:05:26,190 --> 00:05:29,240
It runs the code and we receive the

64
00:05:29,250 --> 00:05:29,840
correct answer.

65
00:05:33,950 --> 00:05:36,660
I am solving the same problem in Python

66
00:05:36,670 --> 00:05:38,260
idle to check the result.

67
00:05:38,790 --> 00:05:40,320
I ve got the same result.

68
00:05:42,100 --> 00:05:44,570
I am assigning another task to it.

69
00:05:45,520 --> 00:05:51,590
Response equals agentExecutor .invoke and

70
00:05:51,600 --> 00:05:51,990
a task.

71
00:05:52,680 --> 00:05:57,650
This time it s to calculate 5 .1 to the

72
00:05:57,660 --> 00:05:59,290
power of 7 .3.

73
00:05:59,880 --> 00:06:03,290
Both the base and the exponent are floats

74
00:06:03,300 --> 00:06:07,810
and LLM alone might fail to provide the
correct answer.

75
00:06:09,020 --> 00:06:10,550
I am executing the code.

76
00:06:14,450 --> 00:06:17,840
Notice how it constructs its chain of thought.

77
00:06:20,040 --> 00:06:26,810
If the generated code doesn t work
initially and it detects an error, it

78
00:06:26,820 --> 00:06:32,650
will attempt to resolve the issue and
generate another piece of code, hopefully functional.

79
00:06:34,060 --> 00:06:41,970
I am running it in idle to 5 .1 to the
power of 7 .3.

80
00:06:45,170 --> 00:06:47,500
We see that the response was correct.

81
00:06:52,670 --> 00:06:55,780
Note that response is a Python dictionary

82
00:06:55,790 --> 00:06:57,680
with two key value pairs.

83
00:06:58,650 --> 00:07:03,280
The keys are input, representing what the

84
00:07:03,290 --> 00:07:09,240
user asked and output which contains the
LLM response.

85
00:07:18,010 --> 00:07:18,840
Great job.

86
00:07:19,350 --> 00:07:21,520
Congratulations on reaching this far.

87
00:07:22,110 --> 00:07:26,140
You ve made fantastic progress and
learned a lot about langchain.

88
00:07:26,330 --> 00:07:32,440
You ve learned about connecting langchain
to LLMs, creating and using prompt

89
00:07:32,450 --> 00:07:38,520
templates and building both simple and
sequential chains using langchain agents

90
00:07:38,530 --> 00:07:38,980
and more.