v0.6.0

Balanced log-odds fusion for hybrid search

balanced_log_odds_fusion() combines Bayesian BM25 probabilities with dense cosine similarities by normalizing both signals in logit space. Min-max normalization equalizes voting power, preventing heavy-tailed sparse logits (from sigmoid unwrapping) from drowning the dense signal while preserving document-length and term-frequency priors.

BEIR hybrid search benchmark

Evaluated on 5 BEIR datasets with retrieve-then-evaluate protocol (top-1000 per signal, union candidates, pytrec_eval). Dense encoder: all-MiniLM-L6-v2.

Method	ArguAna	FiQA	NFCorpus	SciDocs	SciFact	Average
BM25	36.16	25.32	31.85	15.65	67.91	35.38
Dense	36.98	36.87	31.59	21.64	64.51	38.32
Convex	40.03	37.10	35.61	19.65	73.38	41.15
RRF	39.61	36.85	34.43	20.09	71.43	40.48
Bayesian-Balanced	37.28	40.57	35.63	21.55	71.75	41.36
LO-Local	39.63	37.20	34.10	19.50	73.81	40.85
Bayesian-LogOdds	37.17	33.12	35.25	18.52	72.25	39.26

Bayesian-Balanced leads in average NDCG@10 (41.36%), MAP@10 (30.23%), and Recall@10 (49.92%).

New features

• balanced_log_odds_fusion() in bayesian_bm25.fusion
  • Accepts weight parameter for asymmetric signal weighting (default 0.5)
  • Composes existing library functions (logit, cosine_to_probability, _clamp_probability)
• BEIR hybrid search benchmark (benchmarks/hybrid_beir.py)
  • 9 fusion methods: BM25, Dense, Convex, RRF, Bayesian-OR, Bayesian-LogOdds, LO-Local, Bayesian-LO-BR, Bayesian-Balanced
  • Snowball English stemmer + stop word removal (matching BEIR official BM25 baseline)
  • Embedding cache (.npz) to skip re-encoding across runs

Install

pip install bayesian-bm25==0.6.0