1
00:00:00,06 --> 00:00:02,05
- [Instructor] In this lesson, we'll build a simple

2
00:00:02,05 --> 00:00:04,08
knowledge retrieval workflow from scratch.

3
00:00:04,08 --> 00:00:08,00
This is the first tool our agent will be able to access.

4
00:00:08,00 --> 00:00:11,05
Create a new workflow,

5
00:00:11,05 --> 00:00:17,06
and let's name this, knowledge tool.

6
00:00:17,06 --> 00:00:20,08
So quick heads up, this is a very lightweight example.

7
00:00:20,08 --> 00:00:24,01
Perfect for short documents with just a few pages

8
00:00:24,01 --> 00:00:26,03
that don't require a complex RAG,

9
00:00:26,03 --> 00:00:28,08
retrieval augmented generation setup.

10
00:00:28,08 --> 00:00:30,03
For a full retrieval setup,

11
00:00:30,03 --> 00:00:33,08
see my other course, build a custom chat bot with n8n.

12
00:00:33,08 --> 00:00:36,02
The goal of this workflow is to grab a file

13
00:00:36,02 --> 00:00:39,09
from an external source, in this case GitHub.

14
00:00:39,09 --> 00:00:41,07
Let's see our file.

15
00:00:41,07 --> 00:00:44,05
Navigate to tools,

16
00:00:44,05 --> 00:00:47,05
knowledge, docs,

17
00:00:47,05 --> 00:00:50,07
and then the conference details markdown file.

18
00:00:50,07 --> 00:00:54,04
So this is the file we want our agent to be able to query

19
00:00:54,04 --> 00:00:57,07
to get information about our conference as needed.

20
00:00:57,07 --> 00:01:00,00
So let's head back to n8n

21
00:01:00,00 --> 00:01:03,06
and start with a workflow trigger.

22
00:01:03,06 --> 00:01:10,05
We want to choose when executed by another workflow.

23
00:01:10,05 --> 00:01:12,07
For input data mode, make sure you select

24
00:01:12,07 --> 00:01:15,02
define using fields below,

25
00:01:15,02 --> 00:01:20,09
and then add one field, which you call question.

26
00:01:20,09 --> 00:01:23,04
Make sure type is set to string.

27
00:01:23,04 --> 00:01:25,03
Click execute step.

28
00:01:25,03 --> 00:01:27,03
And as you can see, this is currently empty.

29
00:01:27,03 --> 00:01:29,02
Now we want to pin a simple test question,

30
00:01:29,02 --> 00:01:32,02
so we can actually use that during development.

31
00:01:32,02 --> 00:01:35,02
Hit edit,

32
00:01:35,02 --> 00:01:41,02
and then enter a small test question like

33
00:01:41,02 --> 00:01:44,07
what halls are available.

34
00:01:44,07 --> 00:01:48,03
Hit save, and you can now see that this data

35
00:01:48,03 --> 00:01:50,06
is pinned for test executions.

36
00:01:50,06 --> 00:01:51,09
Great.

37
00:01:51,09 --> 00:01:55,05
Next, let's close this and add the next node.

38
00:01:55,05 --> 00:01:57,08
So next, let's add a node to GitHub

39
00:01:57,08 --> 00:02:00,02
so we can download our documents.

40
00:02:00,02 --> 00:02:03,02
Search for GitHub,

41
00:02:03,02 --> 00:02:05,04
and then select list files.

42
00:02:05,04 --> 00:02:08,01
Here make sure you're connected to a GitHub account.

43
00:02:08,01 --> 00:02:09,04
If that's not the case,

44
00:02:09,04 --> 00:02:10,09
just click this little dot drop menu,

45
00:02:10,09 --> 00:02:13,01
click create new credentials,

46
00:02:13,01 --> 00:02:15,01
select this one OAuth2

47
00:02:15,01 --> 00:02:17,02
and select connect my account.

48
00:02:17,02 --> 00:02:19,02
Make sure you already logged into GitHub,

49
00:02:19,02 --> 00:02:22,02
and then you should see this success message.

50
00:02:22,02 --> 00:02:24,03
Let's close this window for now

51
00:02:24,03 --> 00:02:28,08
and then put your own GitHub username in this field.

52
00:02:28,08 --> 00:02:33,09
Next, select the forked repository here from your list.

53
00:02:33,09 --> 00:02:36,07
Then we put in the path to our knowledge documents.

54
00:02:36,07 --> 00:02:38,07
In this case that's

55
00:02:38,07 --> 00:02:46,00
tools/knowledge/docs.

56
00:02:46,00 --> 00:02:51,01
Hit execute step.

57
00:02:51,01 --> 00:02:55,04
Let's fix the typo here.

58
00:02:55,04 --> 00:02:58,00
Hit execute step.

59
00:02:58,00 --> 00:02:58,09
And there we are.

60
00:02:58,09 --> 00:03:00,06
So you should now see the markdown document

61
00:03:00,06 --> 00:03:02,02
listed here on the right side.

62
00:03:02,02 --> 00:03:03,04
Let's close this window.

63
00:03:03,04 --> 00:03:07,00
And next, let's add a node to download this file actually.

64
00:03:07,00 --> 00:03:08,07
We don't have to insert everything manually,

65
00:03:08,07 --> 00:03:12,08
just duplicate this node

66
00:03:12,08 --> 00:03:14,05
and let's wire them up.

67
00:03:14,05 --> 00:03:18,06
Open this one, and instead of listing file operation,

68
00:03:18,06 --> 00:03:21,09
now select Get.

69
00:03:21,09 --> 00:03:24,08
We now have to replace the file path with the file path

70
00:03:24,08 --> 00:03:27,08
of the actual document.

71
00:03:27,08 --> 00:03:31,03
We can simply pull that over from here.

72
00:03:31,03 --> 00:03:33,02
And there we go.

73
00:03:33,02 --> 00:03:34,00
That's it.

74
00:03:34,00 --> 00:03:36,01
Just hit execute step.

75
00:03:36,01 --> 00:03:37,08
And you should see that we have now downloaded

76
00:03:37,08 --> 00:03:39,05
this markdown file.

77
00:03:39,05 --> 00:03:41,08
So let's close this.

78
00:03:41,08 --> 00:03:43,08
And since this is a binary file,

79
00:03:43,08 --> 00:03:50,00
let's now extract the text from this file.

80
00:03:50,00 --> 00:03:53,05
Search for extract from file,

81
00:03:53,05 --> 00:03:58,01
and then select extract from text file.

82
00:03:58,01 --> 00:04:02,01
We can leave the standard settings here, hit execute.

83
00:04:02,01 --> 00:04:04,02
And there you see, this is our file content

84
00:04:04,02 --> 00:04:06,05
in plain text format.

85
00:04:06,05 --> 00:04:07,09
So what do we do with this?

86
00:04:07,09 --> 00:04:10,03
We now feed in this entire text file into

87
00:04:10,03 --> 00:04:14,01
the context of an LLM, so the LLM can then answer questions

88
00:04:14,01 --> 00:04:18,02
based on that information.

89
00:04:18,02 --> 00:04:19,07
In this case, we don't need an agent,

90
00:04:19,07 --> 00:04:21,08
we just need a call to a simple model.

91
00:04:21,08 --> 00:04:26,04
So look for Gemini and select this Google Gemini node.

92
00:04:26,04 --> 00:04:29,02
This will allow us to message a model

93
00:04:29,02 --> 00:04:32,03
and just send text to it and get some text back.

94
00:04:32,03 --> 00:04:35,01
Select message a model, make sure it's connected

95
00:04:35,01 --> 00:04:37,06
to your Gemini API key.

96
00:04:37,06 --> 00:04:41,02
The resource should be text, operation is message a model.

97
00:04:41,02 --> 00:04:43,01
And then let's pick a model from list.

98
00:04:43,01 --> 00:04:46,02
We are also using Gemini 2.5 flash here.

99
00:04:46,02 --> 00:04:49,09
These models support context windows up to 1 million tokens.

100
00:04:49,09 --> 00:04:53,01
So there's plenty of text you can put in here.

101
00:04:53,01 --> 00:04:55,09
Now, as a prompt, we want to insert the question

102
00:04:55,09 --> 00:04:58,02
that we got at the beginning of the workflow.

103
00:04:58,02 --> 00:05:02,06
Let's pull that over and you can see the user

104
00:05:02,06 --> 00:05:05,02
asks now what halls are available.

105
00:05:05,02 --> 00:05:07,03
And to answer that, of course, we need the context,

106
00:05:07,03 --> 00:05:09,06
which is the document information.

107
00:05:09,06 --> 00:05:14,03
And to do this, let's add options, the system message,

108
00:05:14,03 --> 00:05:16,01
and you can see the system message prompt here.

109
00:05:16,01 --> 00:05:19,04
Let's make this a little bigger and leave this open

110
00:05:19,04 --> 00:05:22,04
and go to our GitHub repository.

111
00:05:22,04 --> 00:05:26,07
So in here, under tools, knowledge,

112
00:05:26,07 --> 00:05:29,07
you will see the knowledge system prompt text file.

113
00:05:29,07 --> 00:05:32,09
This is a pretty short system prompt to just prompt the LLM

114
00:05:32,09 --> 00:05:35,06
to behave like a Q&A chatbot in this case.

115
00:05:35,06 --> 00:05:36,09
So given a question,

116
00:05:36,09 --> 00:05:40,05
it'll provide an answer given the provided document.

117
00:05:40,05 --> 00:05:42,08
So let's copy this one.

118
00:05:42,08 --> 00:05:48,03
Go over here and put it into the system message.

119
00:05:48,03 --> 00:05:49,01
There you go.

120
00:05:49,01 --> 00:05:52,04
So this is how the system message right now looks.

121
00:05:52,04 --> 00:05:56,00
It contains the original prompt and our information here

122
00:05:56,00 --> 00:05:58,08
because it pulls that from the previous node.

123
00:05:58,08 --> 00:06:00,05
Now, let's try this out.

124
00:06:00,05 --> 00:06:03,01
Click execute step.

125
00:06:03,01 --> 00:06:05,01
And you can see the answer to our question

126
00:06:05,01 --> 00:06:08,00
that the convention center features two, exhibition halls

127
00:06:08,00 --> 00:06:09,07
A and B.

128
00:06:09,07 --> 00:06:11,05
Great, the heavy lifting is done.

129
00:06:11,05 --> 00:06:14,00
What's remaining now is to return this information

130
00:06:14,00 --> 00:06:15,09
in a nice and clean way.

131
00:06:15,09 --> 00:06:22,05
So what we wanna do is to add a simple set field node

132
00:06:22,05 --> 00:06:26,07
where we define a field that is maybe named response,

133
00:06:26,07 --> 00:06:31,02
and it contains the text value from the LLM.

134
00:06:31,02 --> 00:06:32,07
Execute this step.

135
00:06:32,07 --> 00:06:36,02
And you can see this is the output of our whole workflow.

136
00:06:36,02 --> 00:06:37,06
So if we zoom out a little bit,

137
00:06:37,06 --> 00:06:41,05
what this workflow is doing is it takes a question

138
00:06:41,05 --> 00:06:44,00
that is coming in, it pulls information

139
00:06:44,00 --> 00:06:45,02
from an external file.

140
00:06:45,02 --> 00:06:46,09
In this case GitHub.

141
00:06:46,09 --> 00:06:49,02
The model will generate the answer,

142
00:06:49,02 --> 00:06:51,00
the question given the information,

143
00:06:51,00 --> 00:06:54,05
and then it returns the answer as a simple text response.

144
00:06:54,05 --> 00:06:57,00
And that's exactly the answer that our agent needs.

145
00:06:57,00 --> 00:06:58,08
When it asks a question to their tool,

146
00:06:58,08 --> 00:07:00,02
it gets the answer back.

147
00:07:00,02 --> 00:07:02,03
In the next video, we will give our agent

148
00:07:02,03 --> 00:07:04,00
access to this tool.