►
From YouTube: Building a command-line-interface tool designed to answer user queries using Microsoft Graph API
Description
In this 15-minute developer focused demo, Rohan Ankarigari Boda shows combining the power of Microsoft Graph and large language model (LLM) to deliver a powerful wizard – magi. Magi’s three building blocks – command line (LCLI), ChatGPT API and Microsoft Graph. The CLI skeleton is just 16 lines. OpenAI library provides a .NET focused experience. Prompt engineering is key to getting most out of the ChatGPT API. GraphClientFactory makes arbitrary requests to Graph API. A hackathon winning solution.
This PnP Community demo is taken from the weekly Microsoft 365 Platform Community call recorded on April 18, 2023.
Demo Presenter
• Rohan Ankarigari Boda
Supporting materials
• nuget package - System.CommandLine 2.0.0-beta4.22272.1 | https://www.nuget.org/packages/System.CommandLine/2.0.0-beta4.22272.1
• nuget package - OpenAI 1.7.2 - C#/.NET SDK for accessing the OpenAI GPT-3 API, ChatGPT, and DALL-E 2 | https://www.nuget.org/packages/OpenAI/
• nuget package - Microsoft.Graph 5.6.0 | https://www.nuget.org/packages/Microsoft.Graph
• Documentation - System.CommandLine overview | https://learn.microsoft.com/dotnet/standard/commandline/
• Code snippet - Rahtoken/cli.cs - System.CommandLine Example | https://gist.github.com/Rahtoken/45587e6a55f0a7e3fcb45d8eee53b1d4
• Repo – library - C#/.NET SDK for accessing the OpenAI GPT-3 API, ChatGPT, and DALL-E 2 | https://github.com/OkGoDoIt/OpenAI-API-dotnet
• Documentation - Microsoft Graph SDK overview | https://learn.microsoft.com/graph/sdks/sdks-overview
• Documentation - GraphClientFactory Class | https://learn.microsoft.com/dotnet/api/microsoft.graph.graphclientfactory
• Slides – magi | https://www.chotu.me/magi.pdf
• Repo - Code – magi | https://github.com/Rahtoken/magi-msgraph-hackathon
Learn more
• Microsoft 365 Unified Sample gallery - https://aka.ms/community/samples
• Microsoft 365 Platform Community in YouTube - https://aka.ms/community/videos
• Microsoft 365 Platform Community - http://aka.ms/community/home
A
Foreign
over
the
last
few
weeks,
we've
seen
AI
do
everything
right
from
generating
images
to
generating
audio
summarizing
text.
One
thing
that
we
haven't
seen
in
doing,
though,
is
to
call
the
Microsoft
graph.
Some
say
that
you
could
do
it,
and
actually
you
can
so
you
are
about
to
see
a
demo
where
you
can
use
AI
to
call
the
Microsoft
graph,
and
that
is
really
really
cool.
So
with
that
over
to
Europe.
B
Kind
of
disclaimer
this
is
my
first
Community
call
and
I'm
glad
to
be
presenting
so
I'll
just
quickly
go
ahead
of
there
might
be
minor
hiccups,
as
is
with
live
presentation.
So
please
bear
with
me
quickly
moving
on
to
the
screen
share
so
for
the
Microsoft
hack,
together,
hackathon
I,
built
Magi,
which
is
basically
a
hack
which
combines
the
power
of
Microsoft
graph
and
hooks
it
up
with
llm
magic.
B
B
So
how
Magi
was
born
was
it
was
basically
my
answer
to
the
question.
What
if
I
can
just
talk
with
the
Microsoft
graph,
so
Microsoft
graph
is
a
impressive
piece
or
impressive
framework
for
billing
intelligent
apps.
It's
got
tremendous
documentation
and
excellent
toolkit
and
there's
also
they're
already
awake
of
the
graph
Explorer
to
play
around
with
the
Microsoft
apis.
B
However,
still
more
steps
than
we
would
like
to
a
user
might
not
want
to
figure
out
all
the
authentication,
all
the
permissions
required
for
a
particular
API
and
what
what
are
the
permutations?
You
can
call
the
API
with
what,
if
all
that
could
be
simplified
to
the
user,
where
you
could
just
in
a
way
command
the
Microsoft
graph
API.
So
that's
the
core
reason
for
building
Magi
and
it
has
three
building
blocks:
it's
a
command
and
application.
B
It
uses
a
chat,
GPD
API
and
it
relies
on
Microsoft
graph
I'll
go
into
detail
in
each
of
these
billing
blocks
and
there
will
be
code.
So
I
will
share
the
Snippets
of
code.
We
used
to
build
a
hack,
so
here's
the
flow
of
how
Magi
works.
So
first,
a
user
issues,
a
query
to
match
I
so
and
the
secrets
also
Magi
is:
it
relies
on
the
chat,
GPD
API
to
actually
decode
the
user
query
into
HTTP
requests
which
it
can
for
Microsoft
graph.
B
So
after
the
charge,
GPD
API
Returns,
the
HTTP
request
for
Microsoft
graph
manager
takes
care
of
all
the
Authentication
permissions
and
all
the
scaffolding
and
makes
a
graph
request
and
surfaces.
The
graph
response
to
the
user
Magi
was
built
in
approximately
two
days
and
the
reason
I
was
able
to
build
this
in
such
a
quick
time
with
quick
turnaround
was
because
of
the
vibrant,
ordinary
EOS
system.
I
had
for
all
the
functionality
I
wanted.
B
There
was
a
package
to
get
me
started
and
so
that
I
didn't
have
to
deal
with
it
in
details
of
what
I
need
of
of
say,
stuff,
which
is
required,
but
rather
I
could
focus
on
the
stuff
which
I
wanted
to
do
so
for
the
command
line.
I
relied
on
a
new
system.com
online
package,
which
is
in
preview
the
charge
GPD
API
was
delivered
through
an
open,
AI
community
supported
package
and
Microsoft
graph
was
built.
B
B
You
don't
have
to
take
read
the
arcs,
which
you
pass
into
main
anymore
and
deal
ensure
that
all
the
types
are
right.
All
the
options
are
given
correctly
and
do
the
validations
and
all
it's
a
fluent
interface
for
building
CLI
apps
in
a
query,
quick
manner
in
it's
such
quick
that
the
CLI
skeleton
of
magic
was
just
16
lines.
B
How
do
I
make
chat,
GPT
understand,
Microsoft
graph,
so
llms
are
tremendous
technology
and
we
still
discovering
what
all
is
possible
through
them.
When
openai
announced
the
charge
GPD
API
I
was
intrigued
and
I
wanted
to
see
how
I
could
fit
in
the
API
with
the
hackathon
and
that's
how
I
got
started
on
it.
B
Net
native
interface,
for
working
with
the
API
moving
on
to
the
most
important
part
of
the
hack
was
how
do
I
get
chargpd
to
generate
the
correct.
Http
requests
go
ahead
graph,
so
the
key
was
prompt
engineering,
which
is
basically,
you
instruct
chatgpt
to
do
something,
and
then
you
iterate
upon
those
prompts
so
as
to
you,
so
that
you
arrive
at
something
which
gives
you
what
you
want
to
acceptable
manner.
B
B
I,
basically
asked
that
GPD
or
zoomed
a
role
of
Magi
and
I,
guided
it
how
it
would
work
an
important
part
of
making
it
work
was
so
that
I
could
identify
the
HTTP
request
correctly.
So
what
I
instructed
was?
You
should
only
provide
me.
The
responses
in
this
form
HTTP
verb
or
an
end
point
I
had
to
do
this,
because
if,
if
you
have
used
rgbt
so
far,
you
know
that
it
is
verbose
and
it
could
go
on
and
on
and
on
about
various
stuff
related
to
the
question
you
ask.
B
Instead
of
directly
answering
it,
I
wanted
a
straightforward
response
so
that
I
could
easily
parse
it
and
make
the
Microsoft
graph
request
coming
to
how
the
graph
SDK
worked.
So
one
piece
which
I
have
to
take
care
of
is
that
the
graph
SDK
is
very
structured
in
order
to
make
a
request
for
the
users
API,
you
would
use
graphclient.uses
dot
getasync
or
something
like
that.
The
issue
I
had
a
user
of
Magi
could
make
a
query
for
any
arbitrary
endpoint
it
could.
B
It
could
range
from
users
to
to
do
and
all
mapping
all
these
quests
to
a
specific
graph
request
is
not
trivial,
because
there
can
be
n
number
of
permutations.
So
what
I
did
was
I
used
the
graph
client
Factory
to
get
me
an
instance
of
the
graph
based
client,
which
is
basically
an
HTTP
client
on
steroids,
which,
with
an
authentication
provider
and
other
niceties
which,
with
which
you
can
easily
make
graph
requests.
B
So
once
you
obtain
a
client
graph
client
from
the
graph
client
Factory
or
it's
all
a
matter
of
populating
the
HTTP
request
message
and
making
that
request
before
I
delve
into
questions
I'll
go
a
bit
deeper
into
the
code.
So
what
you're
seeing
right
now
is
Magi
in
its
entirety.
So
all
these
are
the
Imports.
And
then
this
is
system.
Message
is
a
prompt
which
I
shared
earlier,
but
I'm
basically
instructing
gbd.
B
To
give
me
the
response
in
this
manner
going
down,
we
have
all
the
CLI
stuff
the
query
argument
which
to
take
query
from
the
user.
The
config
file
argument
option,
and
so
what
happens
is
after
I
hit
the
charge,
GPD
API
with
the
user's
query
and
it
Returns
the
HTTP
request:
I
pass
the
request,
pass
the
response,
sorry
and
then
I
create
a
glass
client
and
make
the
HTTP
request
message.
So
I
check
the
response
and
then
I
create
a
graph
request.
B
B
D
Question
I
can
I
guess
I
hear
I
can
I
can
work
as
a
kind
of
a
few
comments
and
questions
I
have
to
call
out
the
amount
of
code
within
your
solution
is
just
spot
on
because
it
actually
shows
the
power
of
sdks
right
so
you're,
basically
using
the
existing
combination,
but
that's
the
beauty
of
modern
developers
as
well.
It's
not
about
the
amount
of
code.
It's
about.
D
How
do
you
combine
the
existing
Library
so
really
awesome
stuff
there
and
somebody
was
saying:
are
you
a
Harry
Potter
so
because
of
that
you
know
the
amount
of
the
the
things
work
surprisingly.
Well
is
there
a
documentation
on
this
and
GitHub
have
I
know
that
you
have
a
repo
available
on
that
solution
and
it
has
some
level
of
a
documentation
again,
it's
magical,
how
simple
it
is,
but
it's
still
incredibly
powerful.
D
D
B
That
it
was
an
iterative
process
with
a
lot
of
trial
and
error.
I
know
one
thing
I
found,
which
helped
me
was
I.
Try
to
view
the
mental
model
of
like
I
thought
of
myself
as
strategy
PT
and
say
what
would
be
the
specific
set
of
instructions
which
are
so
tightly
scoped
to
what
I
want
yeah
so
like
wearing
a
hat
and
like
thinking
about
it,
helps
yeah.
D
And
I
I
have
to
say
this
is
an
absolutely
brilliant
example
with
people
there's
this
classic
ongoing
meme
joke
related
on.
Everybody
will
be
a
prompt
engineer
at
some
points
because
now
the
AI
has
taken
over
in
the
world,
but
but
that's
part
of
the
game
right.
So
it's
part
of
the
the
now
the
tool
set.
Are
you
afraid,
just
out
of
curiosity,
for
before
we
go
to
that
close
up?
The
call?
D
D
Absolutely
that
I
personally
I
I
do
agree
on
that
thinking.
It's
the
I
think
whenever
we
introduced
cars
back
in
19
or
whatever
10
people
are
afraid
that
people,
you
know
horses
are
no
longer
needed
and
everybody
is
losing
their
job
and
blah
blah
blah.
That's
not
the
case.
It's
Evolution,
it's
an
abstraction
of
things
and
we
can
focus
on
more
meaningful
functionalities
so
or
or
tasks
as
a
human,
so
really
really
cool.
Thank
you,
Rohan,
absolutely
brilliant,
really,
really
cool
demo
and
and
a
great
demo
a
great
solution.