Do we something similar to opensearch's "Normalization processor" in elasticsearch?

Elastic Stack Elasticsearch
1

One of the major pain point with hybrid-search is how to combine BM25 (lexical) score with vector similarity score. BM25 scores are unbounded and their absolute value does not indicate degree of relevance. While on the other hand, vector (cosine) similarity score is bounded between 0 and 1, and its absolute value indicates degree of relevance.

Due to difference in scale of their scores, it is difficult to combine them.

Opensearch provides score normalization feature called, Normalization processor. Is there something similar in elasticsearch?

PS: RRF is one of the way to combine BM25 and vector similarity score. But currently this feature is in technical preview and also RRF is not be suitable for all the use-cases.

3

Actually it does not compare scores but only document ranks for every query (vector, BM25, text expansion...).

May be this could help: Similarity module | Elasticsearch Guide [8.13] | Elastic. I can see for example some normalization in IB and DFR. But I can't comment more as I never played with that :wink:

4

Thanks for the reply @dadoonet . The link you shared are on how to set parameters for 'similarity' formula. However, I'm looking for normalization technique that uses score of all the matched document. For example "min-max" normalization uses score across matched documents, and generates their normalize_score.