Release v0.4.0

bb25 v0.4.0

New Features (since v0.2.0)

Probabilistic Fusion

• Multi-Head Attention fusion (MultiHeadAttentionLogOddsWeights) — multiple heads independently learn query-dependent signal weighting, averaged in log-odds space
• GELU gating (logit × σ(1.702 × logit)) for smoother noise suppression
• Generalized Swish with configurable β parameter

Retrieval Efficiency

• BlockMaxIndex (BMW) — block-level max score precomputation for safe early termination in top-k retrieval
• Pruning for Attention & Multi-Head modules — upper-bound filtering to skip low-potential candidates

Score Calibration

• PlattCalibrator — sigmoid scaling via gradient descent
• IsotonicCalibrator — PAVA-based non-parametric calibration with binary search interpolation

Temporal Modeling

• TemporalBayesianTransform — exponential decay weighting with configurable half-life for recency-sensitive search

Flexibility

• AttentionLogOddsWeights: configurable seed and base_rate bias
• LearnableLogOddsWeights: base_rate support
• Custom prior_fn for BayesianProbabilityTransform

Benchmarks

• Multi-head attention fusion (4 heads) evaluation
• GELU gating benchmark
• Updated scorer list to 10 methods