Formal models and scripts: word2vec-synonym-provider-vs-exact-cosine-reference

Tools: Python

These are the working models and scripts this part of the hunt used. Some encode the code as it is (a counterexample is the expected result), others a fixed or correct variant; read the header comment of each file. Paths inside them were replaced by placeholders such as $LUCENE (a Lucene checkout) and $WORK.

Findings

• Word2VecSynonymFilter documents minAcceptedSimilarity as a cosine similarity, but Word2VecSynonymProvider compares it with the normalised score (1 + cosine) / 2, so the effective cosine cutoff is 2m - 1 (the default 0.8 accepts cosine 0.6, 0.5 accepts orthogonal terms, values below 0.5 accept negatively correlated terms); documentation defect, the javadoc fix is the default proposal (finding LFV-870; repro · finding page)

Files

• word2vec-min-accepted-similarity-compared-as-score-not-cosine/model/threshold_closed_form.py (used for LFV-870)

How to run

• Python: python3 <file>.py; scripts that import z3 need the z3-solver package.

Snapshot of 2026-11-03. AI-assisted; not yet triaged by a human Lucene committer. Feedback: open an issue in lucene-ai-bug-hunt.