Search in video

0:00
we have a brand new version of autogen
0:02
Studio that just dropped it has new
0:04
features expanded functionality and
0:07
we're going to build something for the
0:08
real world and this video is going to be
0:10
a little different I haven't built this
0:12
agent team myself yet so you're going to
0:14
build it with me and you're going to see
0:15
my mistakes let's get into it the team
0:18
we're going to be building today is
0:19
going to take a YouTube url grab the
0:21
transcript for it and then convert that
0:23
transcript into a compelling blog post
0:26
and a tweet thread this is something
0:28
that I need it's a real world use case
0:30
and we're going to build it together all
0:31
right so the first thing we're going to
0:32
do is install autogen Studio from
0:35
scratch now I had an older version of
0:37
autogen studio and then when I went to
0:39
upgrade to the newer version of autogen
0:41
Studio everything broke and I kept
0:43
getting all these database issues and
0:45
errors and it was just a nightmare and
0:47
what finally fixed it was uninstalling
0:50
autogen studio with this command pip
0:52
uninstall autogen studio and not only
0:55
that I actually had to go into my python
0:57
site packages and manually delete the
0:59
auto gen Studio folders so if you run
1:01
into any issues while upgrading to the
1:03
new autogen studio be sure to do that
1:05
and I already made a beginner tutorial
1:07
about autogen Studio I'll drop a link to
1:09
that in the description below so the
1:11
first thing we're going to do is create
1:12
a new cond environment condac create DN
1:15
AG python equal 3.11 hit enter so I
1:18
already have an existing environment we
1:20
are going to remove it and then accept
1:22
we're going to install everything we
1:23
need then we're going to grab this
1:25
Command right here cond to activate AG
1:27
paste it in hit enter and now we know
1:29
the environment is active because it's
1:31
right there next we're going to install
1:32
autogen studio and this is super simple
1:35
pip install autogen Studio then just hit
1:38
enter now once you're done with that
1:39
we're going to run autogen Studio ui--
1:42
Port 8081 and you can specify any port
1:45
you want sometimes if you close autogen
1:48
studio in the wrong way it'll say this
1:50
port is busy and then you just change it
1:52
to 882 or 883 or anything that you want
1:55
so I'm going to hit enter on 881 and
1:57
there is that issue so I closed down
1:59
Port 8 881 in the wrong way last time so
2:01
here is the error that you will see so
2:03
we're going to change that to 882 now
2:05
and now it works I'm going to hold down
2:07
command and I'm going to click on this
2:09
URL to open up local host in my browser
2:11
all right and here it is the newest
2:12
version of autogen studio so what is
2:15
different about it if you go to the
2:17
build tab you'll see we have new tabs
2:20
over here for one we have this models
2:22
tab now you can predefine different
2:25
models to use with your agents and you
2:27
can mix and match as you want you can
2:29
have specific specific models tailored
2:31
for specific agent use cases it is
2:33
brilliant so here you can see we have
2:35
GPT 4 GPT 4 preview and the BLS Zephyr
2:38
7B Alpha if you wanted to add a new
2:41
model we'll go do that right there but
2:42
we don't need to do that we're going to
2:44
be using chat GPT 4 today and we're
2:46
going to actually be using this preview
2:47
version the other thing that's new and
2:50
I'll get to it in a second is the fact
2:51
that you can have agent teams greater
2:54
than just two agents in the previous
2:56
version you could only have two agents a
2:58
sender and a receiver now you can have
3:00
agent groups where you can group
3:02
together multiple agents in addition to
3:05
your user proxy now I haven't tested
3:07
this yet I'm hoping it works but if it
3:10
doesn't you're going to see it so switch
3:11
over to open AI if you don't already
3:13
have an account go ahead and sign up
3:15
come to the API Keys section we're going
3:17
to create a new key I'm going to call it
3:19
autogen studio hit
3:23
enter and it's asking me to verify that
3:25
I'm a human so this is something I'd
3:27
usually cut out of the video use the
3:29
eror to rotate the object to face in the
3:31
direction of the hand interesting all
3:34
right verification complete all right
3:36
now I'm going to copy this yes I am
3:37
going to revoke this API key before
3:39
publishing the video I know you all are
3:41
sick of hearing me say that but if I
3:43
don't I get a bunch of comments about
3:45
not showing my API key so we'll come
3:47
back to autogen Studio we're going to
3:50
click on this model right here and then
3:52
we're going to enter our API key I don't
3:54
believe we need to do anything else for
3:55
GPT 4 everything else should stay the
3:58
same because the chat GPT T API is the
4:00
default then we'll hit enter model
4:02
created successfully now the first thing
4:05
we want to do is create a skill and the
4:08
skill that I need created today is a
4:10
function to go grab the transcript from
4:13
a YouTube video now how would I go about
4:16
writing that I'm actually going to use
4:17
chat GPT to write it for me now just to
4:20
make sure it is in the format that is
4:22
similar to something I already know
4:24
works I clicked over to the skills tab
4:26
I'm looking at this fetch profile skill
4:29
clicking into it and this is pretty
4:31
similar where you pass in a URL and it
4:33
goes and gets something so this is
4:35
probably what I'm going to want to do
4:37
something similar to so I'm going to go
4:39
ahead and copy it I switched over to
4:41
chat GPT I'm going to paste it in and
4:44
then under it I'm going to say here's an
4:46
example script make something similar to
4:48
this but instead make it accept a
4:50
YouTube url as a parameter it goes and
4:53
gets the transcript from the URL via the
4:55
YouTube API and then Returns the
4:57
transcript let's see what we get okay so
4:59
it looks like this method is going to
5:01
accept a video URL so that's right so
5:03
far giving me some documentation that's
5:05
fine it's going to extract the video ID
5:08
from the URL that looks good okay and
5:10
it's going to use the Google api's
5:13
YouTube API so hopefully this works
5:15
we'll see we might need to iterate on
5:17
this if it doesn't work I've seen
5:18
another method written for this exact
5:20
use case to go grab the transcript and
5:22
it was a little bit different from this
5:24
so I don't know if it's going to work
5:25
okay so this one says the YouTube data
5:27
API requires an API key which you can
5:30
obtain from the Google Cloud console so
5:32
this is different I don't want to have
5:34
to do that let's try again show me
5:36
another way to do it that doesn't
5:38
require an API key there should be a
5:40
YouTube API that you can hit directly
5:43
okay and this looks more akin to what I
5:45
saw when I did it last time so it's
5:47
using this Library called YouTube
5:50
transcript API and hopefully it's able
5:53
to do it just with that and it says I
5:55
need to install the YouTube transcript
5:57
API with Pip but I'm going to try it
5:59
without that that so I'm going to go
6:00
ahead and copy the code and I believe
6:02
autogen Studio when I paste this in will
6:05
automatically know that it needs to
6:07
install that Library so I copied it
6:09
switch back to autogen Studio we're
6:11
going to create a new skill I'm going to
6:13
call it fetch YouTube transcript I'm
6:15
going to take what's in here which is a
6:16
sample I'm going to delete it then I'm
6:17
going to paste in what I just had it has
6:19
an example of how to use it at the
6:21
bottom I might as well leave that in and
6:23
then I'm going to hit okay so that's
6:25
done so now we have our model and our
6:27
skill created now let's create an agent
6:30
we have our primary assistant and our
6:32
user proxy but I want to create an agent
6:35
team and the Agents that I want are one
6:38
I want an agent to go and get the
6:39
YouTube transcript and remove the
6:41
timestamps and then return it the second
6:43
agent I want is to take that text and
6:46
then convert it into a blog post and
6:48
into a Twitter thread let's see if we
6:51
can do that so I'm going to click new
6:52
agent the agent name I'm going to call
6:54
it transcript getter for the description
6:57
takes a YouTube URL and gets the
6:59
transcript from a video I'm going to say
7:01
Max consecutive auto reply just leave it
7:03
as the default human input mode never
7:06
and then the system message this is
7:08
where I believe we need to actually tell
7:10
it which tool to use you are an AI agent
7:13
that uses the fetch YouTube transcript
7:15
skill to get a YouTube transcript for
7:18
subsequent processing we're going to be
7:19
using this model now if you wanted to
7:21
add a different model you do so right
7:23
here now here for the skills this is
7:25
where we need to add our skill so I go
7:27
ahead and click plus here we go fetch
7:29
YouTube transcript I'm going to go and
7:32
say add skill and there it is then I
7:35
click okay so now we have our transcript
7:37
getter then let's create our second
7:39
agent and this one is going to be
7:41
content writer takes raw YouTube video
7:44
transcripts and converts it into blog
7:46
posts and tweet threads Max consecutive
7:48
auto reply stays the same human input
7:50
mode never now for the system message
7:53
you are an insightful intelligent and
7:55
witty content writer who is able to take
7:57
raw YouTube video transcripts and turn
7:59
them into blog posts and tweet threads
8:01
we do not need a skill for this guy and
8:03
we're going to leave the model the same
8:04
as before so now we have our two agents
8:07
right there now the last thing we need
8:09
to do is create a workflow so we'll
8:11
click over to workflows and we're going
8:12
to click new workflow now this is
8:14
something new previously when you
8:16
clicked workflow it just automatically
8:18
opened up and you can only have two
8:20
agents now you can have two agents and
8:22
you can also do this thing called group
8:24
chat so we're going to use group chat
8:25
because we're going to have three agents
8:27
the content writer the YouTube scraper
8:30
and then the user proxy so we're going
8:32
to hit group chat we're going to call
8:34
this YT script to content workflow
8:37
description so we're going to say takes
8:39
a YouTube url gets the transcript of the
8:42
video and then creates a blog post and
8:44
tweet thread from that transcript
8:46
summary method we're going to leave as
8:47
last which is the default the sender is
8:49
going to be the user proxy and then the
8:52
receiver is going to be the group chat
8:53
assistant but we need to fill out the
8:56
group chat assistant so go ahead click
8:58
into it now we we have two primary
9:00
assistants here so I'm going to go ahead
9:02
and delete both of those and we're going
9:03
to add the content writer and the
9:06
transcript getter I'm going to keep it
9:08
the same name I'm going to leave the
9:10
description default and then for the
9:11
system message I'm going to be very
9:13
descriptive okay I'm being very explicit
9:16
here I actually don't know if you need
9:17
to be That explicit in the system
9:19
message cuz typically a system message
9:21
you're just setting up the role you're
9:22
telling the agent who to behave as not
9:25
necessarily describing what you want
9:27
them to do which we will likely do again
9:29
when we actually submit the prompt you
9:31
are a helpful assistant skilled at okay
9:33
coordinating spelled wrong coordinating
9:35
a group of other assistants to solve a
9:37
task the task you will solve is taking a
9:40
YouTube url having an agent use the
9:42
fetch YouTube transcript skill to get
9:44
the transcript from the YouTube video
9:46
pass that transcript to a Content writer
9:49
and then having the content writer
9:50
create a blog post and tweet thread
9:52
based on that transcript now it's
9:54
interesting the group chat manager also
9:56
can take a skill and I don't think we
9:59
need to give it a skill because the
10:01
agent itself has the skill we'll find
10:04
out ah and here's a bug that I've
10:05
noticed so if you remember I I removed
10:09
these primary assistants and I added
10:11
other assistants but when I start typing
10:13
it reverts back to these primary
10:15
assistance so remember fill out
10:17
everything first and then modify your
10:19
group agents so I'm going to delete
10:20
these again and I'm going to add the
10:22
content writer back and the transcript
10:24
getter and then we're going to hit okay
10:26
and okay again let's double check that
10:28
it actually does have the correct agent
10:30
so I'm going to click here click there
10:33
and it does perfect okay so now we're
10:34
going to get out of here and let's test
10:37
it out I hope it works we're going to go
10:39
to the playground we're going to click
10:41
new session and we're going to select YT
10:43
script to content and then hit create so
10:46
here we go take this YouTube URL and I
10:48
paste it in one of my YouTube videos use
10:51
the fetch YouTube transcript skill to
10:53
get the transcript from the YouTube
10:54
video remove the timestamps if any then
10:57
create a blog post and tweet thread
10:59
based on that transcript now the one
11:02
thing that's kind of difficult about
11:04
autogen is it doesn't have explicit
11:06
delegation options I can't explicitly
11:09
say this group chat agent should
11:11
delegate to the content writer or the
11:13
YouTube transcript getter you can't do
11:15
that and I think we just have to be
11:17
extra explicit in our descriptions to
11:19
get it to work that way all right enough
11:21
talk let's see what
11:25
happens okay so we got an error and it
11:28
says that the API key is not there which
11:31
is interesting so I'm going to go back
11:32
to build I'm going to go to models here
11:34
it is and it is there let me see what
11:37
that message is one more time so I click
11:39
here and I'm just going to say retry so
11:42
the API client option must be either set
11:44
by passing environment variable okay
11:47
that's weird I don't understand why I
11:50
have to set the environment variable
11:51
since I have it set in the model right
11:53
there but that's okay we're going to do
11:55
it anyways so I'm going to switch back
11:57
to the terminal going to go go to a new
11:59
tab and I'm going to type then I'm going
12:01
to do cond to activate AG then I'm going
12:04
to type export open AI API key and then
12:06
equals and then the API key and then hit
12:09
enter okay so it should be done now I
12:11
think I have to restart autogen studio
12:13
so I'm going to do that so I come here
12:15
and I'm going to click contrl C okay now
12:18
we're going to restart it there it is
12:20
going to go back to autogen studio now
12:22
going to refresh the page going to go to
12:24
playground and let's click retry nope
12:28
let's try that again I'm going to quit
12:30
out of here again and I'm going to do
12:32
export then the API key one more time
12:34
and then we'll start it up and let's
12:37
refresh the page and let's see if it
12:39
worked that time it did okay so for some
12:42
reason going into a new tab made it so
12:44
that it didn't actually see the API key
12:47
which is weird so just keep that in mind
12:49
if you run into that issue now we can
12:51
tell it's working because we can see
12:52
these three dots bouncing right there so
12:54
switching over to the terminal I can
12:56
actually see the code and the back and
12:58
forth between the agents as it's going
13:00
so it is saying yes use the fetch
13:03
YouTube transcript skill so that's good
13:05
and it does know to install the YouTube
13:07
transcript API now one thing to note is
13:10
that in autogen it doesn't actually show
13:12
you the back and forth conversation I do
13:14
know that there's an open ticket to make
13:16
it do that so you don't have to look at
13:18
the terminal while it's going but here
13:19
it is so there it got the YouTube
13:21
transcript that's perfect okay it's done
13:24
look at that all right so here it is
13:28
let's take a look look at what we got so
13:29
here's the code output okay we got all
13:31
that that's great let's look at the
13:33
actual agent messages so the user proxy
13:35
says take this URL so that's me typing
13:38
take the URL use the fetch YouTube
13:39
transcript skill get the transcript
13:41
remove the timestamps create a blog post
13:43
and a tweet thread then the transcript
13:46
getter to the content writer is actually
13:48
saying to accomplish this task we will
13:50
follow these steps so the transcript
13:52
getter is actually planning out what we
13:54
need to do the transcript getter also
13:56
installed everything we needed writes a
13:58
p python script to use that skill that
14:00
we gave it and then here it is so here
14:02
is the actual transcript from the video
14:05
and then interestingly it didn't give us
14:07
anything after that so it didn't
14:09
actually output it okay so interesting
14:11
for some reason it didn't actually
14:13
output the final blog post and tweet
14:16
thread in the UI itself however if I
14:19
switch over to terminal here it is
14:20
here's a suggested structure for the
14:22
blog post and then it actually writes
14:25
the blog post and this is pretty good
14:27
and then we have a tweet thread so
14:29
here's tweet one tweet two uses emojis
14:32
uses hashtags I mean this is great now
14:34
the one thing I don't understand is why
14:36
it didn't actually output that
14:37
information into here but I don't think
14:39
I'm going to be able to solve that in
14:40
this video we do know it worked though
14:43
so that's it now you know how to use the
14:45
new version of autogen Studio you know
14:47
how to create a skill for your agents to
14:48
use you know how to assign that skill to
14:51
the right agent you know how to create
14:52
the agent team and you know how to
14:54
execute the task and it did so pretty
14:57
darn well I'm happy with this result if
14:59
you want to see me create more of these
15:01
tutorials with real world use cases let
15:04
me know in the comments below if you
15:06
enjoyed this video please consider
15:07
giving a like And subscribe and I'll see
15:09
you in the next one