rg.load takes too long because of the vector field

[cceyda]

Is your feature request related to a problem? Please describe.
rg.load takes too long because of the vector field, even when I don't need it.

Describe the solution you'd like
Allow passing a param or through the query
use the elasticsearch _source query field to exclude vector when not needed.

"_source": {        
        "exclude": ["vector"]
    }

Describe alternatives you've considered
wait

Additional context
Low priority

[davidberenstein1957]

#2261

[davidberenstein1957]

Hi @cceyda, thanks for picking this up! I think we need to review the rg.log/rg.load performance in general, given that a lot of users have had some issues with this for larger records.

Perhaps we can implement something like you suggested, or some lazy loading. At least, I expect to be able to fix some of these load issues by implementing this #2341

[orasik]

Excluding vectors would be great. Technically vectors will not be required in some cases when training a model.

The option vectors=None in rg.load could be used to exclude the vector of the record? @davidberenstein1957

[dvsrepo]

thanks @cceyda and @orasik , this is a needed and requested feat. @frascuchon we discussed this a long time ago (get a projection of only some fields when using rg.load) Is there a relatively easy solution to this?

[davidberenstein1957]

I agree with the idea about the projection but IMO they are 2 separate things. Perhaps we can also think about quantization of the vectors to reduce the size a bit. Ideally, I also see this as a broader topic where we potentially can choose which attributes to load/update for a list of records.  On 24 Feb 2023, at 21:40, Daniel Vila Suero ***@***.***> wrote: thanks @cceyda and @orasik , this is a needed and requested feat. @frascuchon we discussed this a long time ago (get a projection of only some fields when using rg.load) Is there a relatively easy solution to this? —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>

[orasik]

@davidberenstein1957 @frascuchon I found a bug when I tried to hard-code a list of fields to test.

This line

argilla/src/argilla/client/client.py

Line 707 in c35d63a

projection={"*"},

projection={"*"}

is leading to this code for selecting which fields to include

argilla/src/argilla/client/apis/datasets.py

Line 185 in c35d63a

"fields": list(projection) if projection else ["id"],

"fields": list(projection) if projection else ["id"]

If you make "fields": ["id"]

You will receive this error:

records = [sdk_record_class.parse_obj(r).to_client() for r in records]
  File "pydantic/main.py", line 527, in pydantic.main.BaseModel.parse_obj
  File "pydantic/main.py", line 342, in pydantic.main.BaseModel.__init__
pydantic.error_wrappers.ValidationError: 1 validation error for TextClassificationRecord

Not sure what is the validation behind parse_obj that is causing this error.

[orasik]

Found the issue.

inputs is a required field in the validation. If I change the line

argilla/src/argilla/client/apis/datasets.py

Line 185 in c35d63a

"fields": list(projection) if projection else ["id"],

to:

"fields": ["id", "inputs"]

it works fine.

I will raise a PR soon

[orasik]

PR #2417

[davidberenstein1957]

@frascuchon I think we should add some w.r.t. which fields we update during a rg.log afterwards to avoid overwriting records without vector. Can this be done easily?