Rank_feature query should support sparse_vector field type

[kderusso]

The rank_feature query boosts relevance scores based on numeric values of rank_feature or rank_features fields.

We should update this so it will also work with sparse_vector fields.

[elasticsearchmachine]

Pinging @elastic/es-search (Team:Search)

[elasticsearchmachine]

Pinging @elastic/ent-search-eng (Team:Enterprise Search)

[benwtrent]

This is indeed interesting. I would hope we move sparse_vector further away from rank_features in the future. This could be done by adjusting how values are indexed, etc. and forcing us to have to support rank_feature queries could hamper that for very little gain.

What do you think @jpountz ?

[jpountz]

@benwtrent Agreed, I'd rather preserve the ability to diverge these two fields in the future as they feel like they address different problems. That said, do we already have a query that allows computing the maximum inner product between a query vector and indexed vectors easily? Something like that may be a useful addition (and effectively be the same as a rank_features query with linear function internally)?

[benwtrent]

That said, do we already have a query that allows computing the maximum inner product between a query vector and indexed vectors easily?

I don't think so. Right now the way to do it is a "bool[term1(boost1),...,termn(boostn)]". So, maybe we add a "sparse_vector" query?

[kderusso]

Closing this issue, as we decided to introduce a sparse_vector query instead.