v0.3.1

What's Changed

• Optimize posterior computation using two-step Bayes update (Remark 4.4.5)
  • Replaces sigmoid(logit(L) + logit(br) + logit(p)) with two sequential Bayes updates using only multiplication and division
  • score_to_probability() delegates to posterior() instead of duplicating base_rate logic
• Vectorize scorer internals for faster retrieval
  • _scores_to_probabilities() processes all k documents per query in one vectorized numpy call instead of a scalar-by-scalar inner loop
  • Add _compute_tf_batch() for batch term frequency computation
  • Deduplicate pseudo-query sampling: _sample_pseudo_query_scores() is called once during indexing instead of separately by _estimate_parameters() and _estimate_base_rate()
• Add calibration metrics to the main package
  • expected_calibration_error(), brier_score(), reliability_diagram() are now importable from bayesian_bm25 directly
  • benchmarks/metrics.py re-exports from the main package for backward compatibility
• Fix norm_prior docstring to correctly describe peak at 0.5 and floor at 0.0/1.0

Install

pip install bayesian-bm25==0.3.1