Fix for propagating filters from compound to inner retrievers

[pmpailis]

This PR aims to fix an issue where the inner filters are not properly added to the children of compound retrievers if they require different number of rewrite rounds.

[elasticsearchmachine]

Hi @pmpailis, I've created a changelog YAML for you.

[pmpailis]

The prefilters were not actually used anywhere for the RankDocsRetrieverBuilder as they had already been accounted for when computing the parent results.

[elasticsearchmachine]

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

[pmpailis]

run elasticsearch-ci/part-3

[benwtrent]

Could you add a test that exposes the bug? My understanding was that it appeared when using a top level filter with a query_vector_builder in kNN. Lets add a test for that somewhere. It can be a simple single node test (not yaml).

To make the query vector builder stuff easier to test, you can maybe add a "test query vector builder" that simply returns a constant vector.

[pmpailis]

Could you add a test that exposes the bug? My understanding was that it appeared when using a top level filter with a query_vector_builder in kNN. Lets add a test for that somewhere. It can be a simple single node test (not yaml).

To make the query vector builder stuff easier to test, you can maybe add a "test query vector builder" that simply returns a constant vector.

That's a good idea. The query_vector_builder is a sub-case of the actual bug, but it wasn't so straightforward to add it as a yaml test. Will add a standard single node test instead as suggested, thanks! :)