1
00:00:00,830 --> 00:00:07,380
In this video, we ll discuss React, which
is short from reasoning and acting.

2
00:00:08,650 --> 00:00:11,680
We will also create and use a React agent.

3
00:00:12,410 --> 00:00:15,300
React is a new approach that combines

4
00:00:15,310 --> 00:00:19,260
reasoning and acting capabilities of LLMs.

5
00:00:19,770 --> 00:00:24,260
It does this by allowing LLMs to generate

6
00:00:24,270 --> 00:00:29,740
reasoning traces and task -specific
actions in an interleaved manner.

7
00:00:30,590 --> 00:00:36,820
This means that the model can reason
about a task and then take actions in the

8
00:00:36,830 --> 00:00:40,640
real world to gather more information or
complete the task.

9
00:00:42,490 --> 00:00:49,140
This combination of reasoning and acting
allows React to overcome some of the

10
00:00:49,150 --> 00:00:54,860
limitations of existing LLMs, such as
hallucinations and error propagations.

11
00:00:56,010 --> 00:01:01,560
Lang chain agents are one of the key
components that make Lang chain so useful.

12
00:01:02,030 --> 00:01:07,940
They take Lang chain tools, which we ve
looked at in some of the other videos,

13
00:01:08,490 --> 00:01:11,800
and chains and combine them into one agent.

14
00:01:13,170 --> 00:01:15,880
We will create and use a React agent.

15
00:01:17,490 --> 00:01:23,980
The big problem when using only tools or
chains is that you probably won t want to

16
00:01:23,990 --> 00:01:27,080
use the same tool with every query you send.

17
00:01:27,630 --> 00:01:30,500
In our previous examples, we used as

18
00:01:30,510 --> 00:01:35,340
tools Wikipedia and DuckDuckGo, but not both.

19
00:01:36,250 --> 00:01:39,540
Using a React agent, it will check the

20
00:01:39,550 --> 00:01:45,920
input and then decide what is the best
tool to use based on the user s query.

21
00:01:46,410 --> 00:01:47,320
Let s get started.

22
00:01:48,270 --> 00:01:50,900
I will install the LangChainHub package.

23
00:01:51,990 --> 00:01:55,660
pip install LangChainHub -q.

24
00:01:56,290 --> 00:02:00,040
The LangChainHub is a hub for sharing and

25
00:02:00,050 --> 00:02:03,200
discovering resources related to Lang chain.

26
00:02:03,790 --> 00:02:07,840
It stores and allows access to pre

27
00:02:07,850 --> 00:02:12,980
-trained prompts, chains, and agents that
can be used in Lang chain projects.

28
00:02:13,890 --> 00:02:18,540
I am loading the OpenAI API key from .env.

29
00:02:21,170 --> 00:02:24,740
Next, let s import the modules for

30
00:02:24,750 --> 00:02:28,640
prompts, agents, tools, and external services.

31
00:02:29,310 --> 00:02:31,260
I am pasting the imports.

32
00:02:33,620 --> 00:02:35,530
I am setting up the LLM.

33
00:02:36,300 --> 00:02:39,230
LLM equals chat -open -ai -of.

34
00:02:40,120 --> 00:02:46,310
And I ll choose gpt4 -turbo, a powerful
language model for processing text and

35
00:02:46,320 --> 00:02:47,370
generating responses.

36
00:02:48,340 --> 00:02:50,930
If you don t have access to gpt4 -turbo,

37
00:02:51,360 --> 00:02:53,630
use gpt3 .5 -turbo instead.

38
00:02:56,010 --> 00:02:59,020
Model name equals gpt4 -turbo -preview.

39
00:03:02,710 --> 00:03:06,450
And the temperature equals 0.

40
00:03:07,940 --> 00:03:11,510
Once again, if you are using a free plan

41
00:03:11,520 --> 00:03:18,090
and don t have access to gpt4 -turbo, use
gpt3 .5 -turbo instead.

42
00:03:21,090 --> 00:03:26,460
In the next step, I will prepare the
prompt template for asking questions.

43
00:03:27,670 --> 00:03:33,220
This template will structure the way the
questions are presented to the agent.

44
00:03:34,690 --> 00:03:44,070
I will define a multi -line string,
answer the following questions as best

45
00:03:44,080 --> 00:03:54,740
you can, and the questions, colon, and
the Q in curly braces.

46
00:03:57,910 --> 00:04:08,120
I am converting the template into a
prompt template equals the prompt

47
00:04:08,130 --> 00:04:14,480
-template constructor from template of template.

48
00:04:17,410 --> 00:04:20,420
I am retrieving the structure of the

49
00:04:20,430 --> 00:04:23,620
react prompt from the online -hub.

50
00:04:23,810 --> 00:04:40,260
Prompt equals I am running the code.

51
00:04:46,160 --> 00:04:49,370
Let's see the input variables of the prompt.

52
00:04:50,600 --> 00:04:53,790
Print prompt .inputVariables.

53
00:05:04,870 --> 00:05:10,600
If you want to see the template, use
prompt .template.

54
00:05:11,830 --> 00:05:13,980
Print prompt .template.

55
00:05:17,060 --> 00:05:19,650
You can see the chain of docs.

56
00:05:23,450 --> 00:05:24,180
Take a look here.

57
00:05:26,420 --> 00:05:27,570
Ok, let's move on.

58
00:05:30,850 --> 00:05:37,460
In this example, I will use three tools,
the Python -Apple tool for executing

59
00:05:37,470 --> 00:05:47,720
Python code, the Wikipedia tool for
searching Wikipedia, and the DuckDuckGo

60
00:05:47,730 --> 00:05:50,060
search tool for general web searches.

61
00:05:51,650 --> 00:05:56,100
The agent will choose the best tool based

62
00:05:56,110 --> 00:05:57,800
on the user's input query.

63
00:05:59,070 --> 00:06:01,060
I am creating the first tool.

64
00:06:02,030 --> 00:06:05,940
I am creating an object of type Python
-Apple tool.

65
00:06:06,890 --> 00:06:11,380
Python -Apple equals and I am calling the
constructor of the class.

66
00:06:13,920 --> 00:06:15,150
And now the tool.

67
00:06:16,560 --> 00:06:23,830
Python -Apple tool equals tool of and the

68
00:06:23,840 --> 00:06:27,890
arguments are name equals Python -Apple.

69
00:06:29,570 --> 00:06:32,160
Give the tool a descriptive name.

70
00:06:32,890 --> 00:06:35,920
The second argument will be func.

71
00:06:36,620 --> 00:06:39,560
This is the function to run when using

72
00:06:39,570 --> 00:06:40,300
this tool.

73
00:06:41,610 --> 00:06:43,860
Python -Apple .run.

74
00:06:45,030 --> 00:06:47,500
And finally, a description.

75
00:06:51,290 --> 00:06:54,180
Useful when you need to run Python to

76
00:06:54,190 --> 00:06:55,120
answer a question.

77
00:06:56,550 --> 00:06:58,480
You should input Python code.

78
00:06:59,570 --> 00:07:02,200
Let's do the same for the Wikipedia tool.

79
00:07:04,000 --> 00:07:06,470
I am creating an API wrapper.

80
00:07:08,360 --> 00:07:14,740
I am calling the constructor and an
Wikipedia query run object.

81
00:07:15,350 --> 00:07:19,080
Wikipedia equals Wikipedia query run.

82
00:07:20,540 --> 00:07:24,740
And the API wrapper equals API wrapper,

83
00:07:25,030 --> 00:07:27,300
the object I have just defined.

84
00:07:30,900 --> 00:07:33,650
Next, I will create the Wikipedia tool.

85
00:07:34,160 --> 00:07:45,200
Wikipedia tool equals tool of and the
arguments name equals Wikipedia, func

86
00:07:45,210 --> 00:07:52,010
equals Wikipedia run and the description.

87
00:07:56,920 --> 00:07:57,330
Very well.

88
00:07:58,100 --> 00:08:01,470
Let's create the DuckDuckGo search tool.

89
00:08:02,240 --> 00:08:09,670
Search equals DuckDuckGo search run and

90
00:08:09,680 --> 00:08:09,990
the tool.

91
00:08:11,680 --> 00:08:12,810
DuckDuckGo tool.

92
00:08:16,910 --> 00:08:23,040
I am calling the constructor of the tool
class and the arguments are name equals

93
00:08:23,050 --> 00:08:35,460
DuckDuckGo search, func equals search
.run and the description.

94
00:08:42,470 --> 00:08:48,040
Pay attention to single quotes if you
define the string using single quotes.

95
00:08:48,770 --> 00:08:51,220
Use a backslash before it.

96
00:08:53,910 --> 00:08:56,500
We have created the tools, the Python or

97
00:08:56,510 --> 00:09:01,320
Apple tool, the Wikipedia tool and the
DuckDuckGo search tool.

98
00:09:05,850 --> 00:09:09,420
Next, I will collect all the tools into a list.

99
00:09:10,670 --> 00:09:12,900
Tools equals and the list.

100
00:09:13,810 --> 00:09:21,770
Python or Apple tool, Wikipedia tool and
the DuckDuckGo tool.

101
00:09:23,380 --> 00:09:28,150
To create the agent, there are basically
two main things you have to do.

102
00:09:28,680 --> 00:09:35,750
Load the tools the agent is going to use
and initialize the agent using the agent

103
00:09:35,760 --> 00:09:37,690
executor and those tools.

104
00:09:38,260 --> 00:09:45,870
First, agent equals createReactAgent of

105
00:09:45,880 --> 00:09:48,750
LLM, tools and prompt.

106
00:09:51,940 --> 00:09:55,950
The createReact function was imported

107
00:09:55,960 --> 00:09:58,730
from langchain .agents.

108
00:10:01,110 --> 00:10:05,480
This line of code combines the LLM, the

109
00:10:05,490 --> 00:10:08,400
tools and the prompt to create a React agent.

110
00:10:11,270 --> 00:10:15,080
Next, I will set up an agent executor to

111
00:10:15,090 --> 00:10:16,000
run the agent.

112
00:10:18,940 --> 00:10:22,810
The executor examines the input and

113
00:10:22,820 --> 00:10:26,550
determines the best tool to use based on
the user's query.

114
00:10:27,220 --> 00:10:33,130
I am creating the executor to manage the
agent's actions and output.

115
00:10:34,740 --> 00:10:40,710
AgentExecutor equals and I am calling the
constructor of the AgentExecutor class

116
00:10:42,550 --> 00:10:52,160
and the arguments, agent equals agent,
this is the object returned by the

117
00:10:52,170 --> 00:11:04,470
createReactAgent function, tools equals
tools, my list of tools and the verbose

118
00:11:04,480 --> 00:11:05,110
equals true.

119
00:11:10,250 --> 00:11:13,080
I will add two more important arguments,

120
00:11:14,010 --> 00:11:25,240
handle parsingErrors equals true and
maxIterations equals 10.

121
00:11:28,980 --> 00:11:35,110
HandleParsingErrors controls how the
agent executor handles potential parsing

122
00:11:35,120 --> 00:11:40,730
errors that might occur during the
execution of tools or the agent itself.

123
00:11:40,960 --> 00:11:48,270
If it is set to true, the executor will
attempt to gracefully handle parsing

124
00:11:48,280 --> 00:11:54,070
errors by logging the error message for
debugging purposes, continuing with the

125
00:11:54,080 --> 00:12:00,630
execution process, potentially skipping
the affected tool or part of the agent's response.

126
00:12:01,340 --> 00:12:05,570
It will also try to recover and proceed
with the remaining steps.

127
00:12:06,320 --> 00:12:12,170
If it is set to false, which is the
default, parsing errors will cause the

128
00:12:12,180 --> 00:12:15,290
execution to stop immediately, raising an exception.

129
00:12:16,120 --> 00:12:19,390
Using this argument can make agents more

130
00:12:19,400 --> 00:12:24,390
robust and less prone to unexpected
failures due to parsing issues,

131
00:12:24,600 --> 00:12:29,410
especially when working with external
tools or unpredictable input.

132
00:12:30,440 --> 00:12:37,710
MaxIterations sets a maximum limit on the
number of iterations or steps the agent

133
00:12:37,720 --> 00:12:40,330
can take during a single execution.

134
00:12:41,180 --> 00:12:44,310
Using this argument prevents agents from

135
00:12:44,320 --> 00:12:50,330
running indefinitely, which could lead to
performance issues, resource exhaustion

136
00:12:50,340 --> 00:12:51,990
or getting stuck in loops.

137
00:12:52,820 --> 00:12:56,230
Once the specific number of iterations is

138
00:12:56,240 --> 00:13:02,670
reached, the execution will stop, even if
the agent hasn't fully completed its task.

139
00:13:03,220 --> 00:13:08,270
Note that the optimal value for
MaxIterations depends on the complexity

140
00:13:08,280 --> 00:13:14,630
of the task and the expected execution
time, you might need to adjust it based

141
00:13:14,640 --> 00:13:16,030
on the specific use case.

142
00:13:16,980 --> 00:13:20,210
Ok, we've successfully built the React

143
00:13:20,220 --> 00:13:24,570
agent, let's take a break and in the next
video we'll test it.

144
00:13:25,180 --> 00:13:28,110
I'm really excited to see what it can do,
see you