Alessandro Benedetti created LUCENE-10593:
---------------------------------------------
Summary: VectorSimilarityFunction reverse removal
Key: LUCENE-10593
URL: https://issues.apache.org/jira/browse/LUCENE-10593
Project: Lucene - Core
Issue Type: Improvement
Reporter: Alessandro Benedetti
org.apache.lucene.index.VectorSimilarityFunction#EUCLIDEAN similarity behaves
in an opposite way in comparison to the other similarities:
A higher similarity score means higher distance, for this reason, has been
marked with "reversed" and a function is present to map from the similarity to
a score (where higher means closer, like in all other similarities.)
Having this counterintuitive behavior with no apparent explanation I could
find(please correct me if I am wrong) brings a lot of nasty side effects for
the code readability, especially when combined with the NeighbourQueue that has
a "reversed" itself.
In addition, it complicates also the usage of the pattern:
Result Queue -> MIN HEAP
Candidate Queue -> MAX HEAP
In HNSW searchers.
The proposal in my Pull Request aims to:
1) the Euclidean similarity just returns the score, in line with the other
similarities, with the formula currently used to move from distance to score
2) simplify the code, removing the bound checker that's not necessary anymore
3) refactor here and there to be in line with the simplification
4) refactor of NeighborQueue to clearly state when it's a MIN_HEAP or MAX_HEAP,
now debugging is much easier and understanding the HNSW code is much more
intuitive
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
