Vector rescoring oversamples k instead of num_candidates

[carlosdelest]

It makes more sense to apply rescoring to an oversampled k instead of num_candidates, as rescoring just a fraction of the candidates will be more performant and offer good recall, specially for smaller k sizes compared to number of candidates.

API changes so we use oversample instead of num_candidates_factor:

GET msmarco-v2-bbq/_search
{
    "query": {
        "knn": {
            "field": "emb",
            "query_vector": [...],
            "k": 10,
            "num_candidates": 100,
            "rescore_vector": {
                "oversample": 2.5
            }
        }
    }
}

This will mean rescoring k * oversample from the num_candidates retrieved on each shard, and returning the top k out of them.

Follow up to #116663

[elasticsearchmachine]

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

[benwtrent]

Great stuff! You may have issues with the api compat tests and thus they may need to be muted before backport. I am not sure.

[carlosdelest]

You may have issues with the api compat tests and thus they may need to be muted before backport. I am not sure.

I hope not as I changed the capability name - I'll keep an eye on this 👍

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	8.x