►
Description
The AI Assisted group continues to work on the Code Suggestions project. We have taught the model to identify and mask personally identifiable information such as email and IP addresses. We are also working on how to improve the quality of predictions in autofill mode.
The product demoed in this video is a CLOSED BETA feature. It is not widely available to all customers at this point. More details about an open beta will come during Q1
Detailed written instructions: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/issues/27#how
B
A
A
Visual
Studio
code
on
my
environment
I
go
to
extensions,
I
search
for
the
gitlab
workload,
extension
I
click
install
and
we
are
ready
and
now
I
have
to
set
up
a
token
for
the
extension
I
come
here.
I
press,
Ctrl
shift,
p
shortcut
to
add
a
gitlab
account
to
vs
code.
I
choose
gitlab.com,
then
I
need
a
token
so
to
up
to
make
a
token
I
go
to
my
project.
A
A
In
the
extension
I
go
there
to
the
extension
settings
and
I
have
one
setting
which
is
enabled
code
completion,
which
is
already
enabled.
Now,
when
I
go
to
a
gitlab
project,
it
must
be
a
gitlab
repository
and
I
click.
A
certain
file
like
hello.js
I,
can
come
here
and
first
write.
Let's
say
hello
in
JavaScript
and
then
I
can
use
the
code.
Simplex
and
writer
create
a
View
application.
A
B
Thanks
a
lot
helper
for
this
example
for
the
tutorial,
how
to
enable
cut
completions
and
the
GitHub
haircut
extension.
So
in
my
work,
I
would
like
to
to
show
how
I'd
like
to
show
completely
another
example:
let's
say
in
Python
and
we
can
apply
multi-term
recommendations
so,
for
instance,
we
would
like
to
to
write
an
application
by
requesting
recommendations
Slide
by
line.
So
let's
start,
and
so
here
is
their
thing-
that
we
should
receive
from
the
model
and,
let's
start
start
line
by
line.
B
B
Doesn't
work:
okay,
let's
import
buy
ourselves
and
then
okay,
let's
define
a
regular
expression
that
matches
an
email
address
yeah.
Here
we
go
get
this
regular
expression,
okay
done.
Let's
then
search
for
an
email
in
our
inputs
and
store
the
first
match
to
variable
address
so
suppose
that
our
input
is
abc
dot
XYZ
at
email.com.
Right.
B
B
B
Now
we
fail
here:
okay,
let's
ask
again
yeah
done.
Sometimes
it
happens,
since
we
have
some
kind
of
Randomness
in
our
recommendations
and
yeah,
let's
print
the
results.
So
that's
good
just
recall
that
in
this
situation
we
were
trying
to
improve
the
quality
of
the
model.
We
made
lots
of
things
so,
first
of
all,
we
were
exploring
how
to
what
is
the
best
strategy
to
decode.
B
The
next
token,
please
consider
that
the
current
model-
this
is
the
model
converted
by
Brandon
architecture,
to
the
gptj
architecture,
supported
by
Triton
that
we
use
to
serve
the
model
in
this
iteration.
We
offer
we
author,
implemented
several
algorithms
to
to
identify
personal
information
and
then
mask
this
personal
information.
So,
for
instance,
let's
say
that
we
also
have
this
third
Python
and
we
would
like
to
get
someone
similar's
email
address.
So
let's
say
this
will
be
my
email
address.
B
B
Yeah
yeah,
we
triggered
this
recommendation.
So
if
you,
if
you
see
that
line,
if
you
check
this
line,
You'll
see
that
we
have
email
placeholders,
so
we
don't
recommend
email
addresses,
we
replace.
We
identify
all
email
addresses,
as
well
as
secrets
and
IP
addresses
with
placeholders,
and
then
we
bring
them.