bayesian-bm25 0.3.0

Bayesian probability transforms for BM25 retrieval scores

Bayesian BM25

[Blog] [Papers]

A probabilistic framework that converts raw BM25 retrieval scores into calibrated relevance probabilities using Bayesian inference.

Overview

Standard BM25 produces unbounded scores that lack consistent meaning across queries, making threshold-based filtering and multi-signal fusion unreliable. Bayesian BM25 addresses this by applying a sigmoid likelihood model with a composite prior (term frequency + document length normalization) and computing Bayesian posteriors that output well-calibrated probabilities in [0, 1]. A corpus-level base rate prior further improves calibration by 68--77% without requiring relevance labels.

Key capabilities:

• Score-to-probability transform -- convert raw BM25 scores into calibrated relevance probabilities via sigmoid likelihood + composite prior + Bayesian posterior
• Base rate calibration -- corpus-level base rate prior estimated from score distribution decomposes the posterior into three additive log-odds terms, reducing expected calibration error by 68--77% without relevance labels
• Parameter learning -- batch gradient descent or online SGD with EMA-smoothed gradients and Polyak averaging, with three training modes: balanced (C1), prior-aware (C2), and prior-free (C3)
• Probabilistic fusion -- combine multiple probability signals using log-odds conjunction with optional per-signal reliability weights (Log-OP), which resolves the shrinkage problem of naive probabilistic AND
• Hybrid search -- cosine_to_probability() converts vector similarity scores to probabilities for fusion with BM25 signals via weighted log-odds conjunction
• WAND pruning -- wand_upper_bound() computes safe Bayesian probability upper bounds for document pruning in top-k retrieval
• Search integration -- drop-in scorer wrapping bm25s that returns probabilities instead of raw scores

Adoption

• MTEB -- included as a baseline retrieval model (bb25) for the Massive Text Embedding Benchmark
• txtai -- used for BM25 score normalization in hybrid search (normalize="bayesian-bm25")

Installation

pip install bayesian-bm25

To use the integrated search scorer (requires bm25s):

pip install bayesian-bm25[scorer]

Quick Start

Converting BM25 Scores to Probabilities

import numpy as np
from bayesian_bm25 import BayesianProbabilityTransform

transform = BayesianProbabilityTransform(alpha=1.5, beta=1.0, base_rate=0.01)

scores = np.array([0.5, 1.0, 1.5, 2.0, 3.0])
tfs = np.array([1, 2, 3, 5, 8])
doc_len_ratios = np.array([0.3, 0.5, 0.8, 1.0, 1.5])

probabilities = transform.score_to_probability(scores, tfs, doc_len_ratios)

End-to-End Search with Probabilities

from bayesian_bm25 import BayesianBM25Scorer

corpus_tokens = [
    ["python", "machine", "learning"],
    ["deep", "learning", "neural", "networks"],
    ["data", "visualization", "tools"],
]

scorer = BayesianBM25Scorer(k1=1.2, b=0.75, method="lucene", base_rate="auto")
scorer.index(corpus_tokens, show_progress=False)

doc_ids, probabilities = scorer.retrieve([["machine", "learning"]], k=3)

Combining Multiple Signals

import numpy as np
from bayesian_bm25 import log_odds_conjunction, prob_and, prob_or

signals = np.array([0.85, 0.70, 0.60])

prob_and(signals)                # 0.357 (shrinkage problem)
log_odds_conjunction(signals)    # 0.773 (agreement-aware)

Hybrid Text + Vector Search

import numpy as np
from bayesian_bm25 import cosine_to_probability, log_odds_conjunction

# BM25 probabilities (from Bayesian BM25)
bm25_probs = np.array([0.85, 0.60, 0.40])

# Vector search cosine similarities -> probabilities
cosine_scores = np.array([0.92, 0.35, 0.70])
vector_probs = cosine_to_probability(cosine_scores)  # [0.96, 0.675, 0.85]

# Fuse with reliability weights (BM25 weight=0.6, vector weight=0.4)
stacked = np.stack([bm25_probs, vector_probs], axis=-1)
fused = log_odds_conjunction(stacked, weights=np.array([0.6, 0.4]))

WAND Pruning with Bayesian Upper Bounds

from bayesian_bm25 import BayesianProbabilityTransform

transform = BayesianProbabilityTransform(alpha=1.5, beta=2.0, base_rate=0.01)

# Standard BM25 upper bound per query term
bm25_upper_bound = 5.0

# Bayesian upper bound for safe pruning -- any document's actual
# probability is guaranteed to be at most this value
bayesian_bound = transform.wand_upper_bound(bm25_upper_bound)

Online Learning from User Feedback

from bayesian_bm25 import BayesianProbabilityTransform

transform = BayesianProbabilityTransform(alpha=1.0, beta=0.0)

# Batch warmup on historical data
transform.fit(historical_scores, historical_labels)

# Online refinement from live feedback
for score, label in feedback_stream:
    transform.update(score, label, learning_rate=0.01, momentum=0.95)

# Use Polyak-averaged parameters for stable inference
alpha = transform.averaged_alpha
beta = transform.averaged_beta

Training Modes

from bayesian_bm25 import BayesianProbabilityTransform

transform = BayesianProbabilityTransform(alpha=1.0, beta=0.0)

# C1 (balanced, default): train on sigmoid likelihood
transform.fit(scores, labels, mode="balanced")

# C2 (prior-aware): train on full Bayesian posterior
transform.fit(scores, labels, mode="prior_aware", tfs=tfs, doc_len_ratios=ratios)

# C3 (prior-free): train on likelihood, inference uses prior=0.5
transform.fit(scores, labels, mode="prior_free")

Benchmarks

Evaluated on BEIR datasets (NFCorpus, SciFact) with k1=1.2, b=0.75, Lucene BM25. Queries are split 50/50 for training and evaluation. "Batch fit" uses gradient descent on training labels; all other Bayesian methods are unsupervised.

Ranking Quality

Base rate prior is a monotonic transform -- it does not change document ordering.

Method	NFCorpus NDCG@10	NFCorpus MAP	SciFact NDCG@10	SciFact MAP
Raw BM25	0.5023	0.4395	0.5900	0.5426
Bayesian (auto)	0.5050	0.4403	0.5791	0.5283
Bayesian (auto) + base rate	0.5050	0.4403	0.5791	0.5283
Bayesian (batch fit)	0.5041	0.4400	0.5826	0.5305
Bayesian (batch fit) + base rate	0.5041	0.4400	0.5826	0.5305

Probability Calibration

Expected Calibration Error (ECE) and Brier score. Lower is better.

Method	NFCorpus ECE	NFCorpus Brier	SciFact ECE	SciFact Brier
Bayesian (no base rate)	0.6519	0.4667	0.7989	0.6635
Bayesian (base_rate=auto)	0.1461 (-77.6%)	0.0619	0.2577 (-67.7%)	0.1308
Bayesian (base_rate=0.001)	0.0081 (-98.8%)	0.0114	0.0354 (-95.6%)	0.0157
Batch fit (no base rate)	0.0093 (-98.6%)	0.0114	0.0103 (-98.7%)	0.0051
Batch fit + base_rate=auto	0.0085 (-98.7%)	0.0096	0.0021 (-99.7%)	0.0013

Threshold Transfer

F1 scores using the best threshold found on training queries, applied to evaluation queries. Smaller gap indicates better generalization.

Method	NFCorpus Train F1	NFCorpus Test F1	SciFact Train F1	SciFact Test F1
Bayesian (no base rate)	0.1607	0.1511	0.3374	0.2800
Batch fit (no base rate)	0.1577	0.1405	0.2358	0.2294
Batch fit + base_rate=auto	0.1559	0.1403	0.3316	0.3341

Reproduce with python benchmarks/base_rate.py (requires pip install ir_datasets). The base rate benchmark also includes Platt scaling, min-max normalization, and prior-aware/prior-free training mode comparisons.

Additional benchmarks (no external datasets required):

• python benchmarks/weighted_fusion.py -- weighted vs uniform log-odds fusion across noise scenarios
• python benchmarks/wand_upper_bound.py -- WAND upper bound tightness and skip rate analysis

Citation

If you use this work, please cite the following papers:

@preprint{Jeong2026BayesianBM25,
  author    = {Jeong, Jaepil},
  title     = {Bayesian {BM25}: {A} Probabilistic Framework for Hybrid Text
               and Vector Search},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.18414940},
  url       = {https://doi.org/10.5281/zenodo.18414940}
}

@preprint{Jeong2026BayesianNeural,
  author    = {Jeong, Jaepil},
  title     = {From {Bayesian} Inference to Neural Computation: The Analytical
               Emergence of Neural Network Structure from Probabilistic
               Relevance Estimation},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.18512411},
  url       = {https://doi.org/10.5281/zenodo.18512411}
}

License

This project is licensed under the Apache License 2.0.

Copyright (c) 2023-2026 Cognica, Inc.