Support for PostgreSQL to_tsvector function with multiple fields

[Charnelx]

Hi.
I added SearchVectorFilter to support to_tsvector function for full text search. Theare two problems I faced:

• this thing is vendor-specific (yep, only PostgreSQL)
• to search by multiple fields we need to create one more argument (e.x. field_names) or use existing field_name with additional type check

Pros:

• fast search
• search by multiple fields at one's

Cons:

• used only with PostgreSQL
• dual meaning of field_name argument wich may lead to misunderstanding and mistakes

Case of use:

models.py

class Person(models.Model):

    name = models.CharField(max_length=128, verbose_name='person name')
    surname = models.CharField(max_length=128, verbose_name='person surname')

views.py

class PersonFilter(filters.FilterSet):
    person = SearchVectorFilter(field_names=['name', 'surname'], lookup_expr='icontains')

[rpkilby]

Hi. I've added more detailed feedback below, but in short, overloading field_name isn't the correct thing to do here. Instead, extend the Filter to use additional arguments.

Also, tests will need to be provided.

[rpkilby]

get_field_parts is not intended to return a list of field names, but a list of the individual parts of a related field name (FK, m2m, 1to1). This result would be nonsensical, as first_name is a regular field on the author model, and does not represent a relationship to an additional Name model or something.

Either way, as explained above, overloading field_name is not the correct path, and these changes are a symptom of that. I would revert these changes.

[Charnelx]

You truth, I have read method descption about "the traversable relationships" and..ignored it.

[rpkilby]

I would inherit CharFilter, as you are validating text.

[Charnelx]

Reasonably and thus accepted.

[rpkilby]

Instead of overloading the purpose of field_name, create a custom __init__ that requires users to pass in a list of search fields. eg,

# Note that this provides a default `field_name` to be used for the annotation
# and query, and that `search_fields` will be a required keyword argument.
def __init__(self, field_name='search_vector', lookup_expr='exact', *, search_fields, **kwargs):
    super().__init__(field_name, lookup_expr, **kwargs)

    self.search_fields = search_fields

From here, it's really simple to overload the .filter() method to add in the annotation

def filter(self, qs, value):
    # Note that the annotation is operating on the filter's `field_name`, which was
    # set above to 'search_vector' as the filter's default value.
    qs.annotate(self.field_name=SearchVector(*self.search_fields))
    
    return super().filter(qs, value)

[Charnelx]

Yes, this was the way it should be implemented. Thanks.

[Charnelx]

@rpkilby great thanks for detailed response!
Well, adding additional argument to Filter was my first idea but I thought that I would be able to make something that not require any external change at all. Now I see the failure of this approach.

I'll take into account your remarks and make another try + will write tests.