bb25 (Bayesian BM25)

bb25 is a fast, self-contained BM25 + Bayesian calibration implementation with a minimal Python API. It also includes a small reference corpus and experiment suite so you can validate the expected numerical properties.

Original author's implementation: The paper author (Jaepil Jeong, Cognica) maintains the reference Python implementation at cognica-io/bayesian-bm25. That library focuses on production-ready score-to-probability conversion with BM25 ranking order preservation, auto parameter estimation, online learning, and log-odds conjunction for hybrid fusion. If you need a drop-in probability transform for an existing search system, use the original. bb25 is a Rust-core experimental validation that prioritizes performance and end-to-end reproducibility of the paper's claims.

Install

pip install bb25

Quick start

Use the built-in corpus and queries

import bb25 as bb

corpus = bb.build_default_corpus()
docs = corpus.documents()
queries = bb.build_default_queries()

bm25 = bb.BM25Scorer(corpus, 1.2, 0.75)
score = bm25.score(queries[0].terms, docs[0])
print("score0", score)

Build your own corpus

import bb25 as bb

corpus = bb.Corpus()
corpus.add_document("d1", "neural networks for ranking", [0.1] * 8)
corpus.add_document("d2", "bm25 is a strong baseline", [0.2] * 8)
corpus.build_index()  # must be called before creating scorers

bm25 = bb.BM25Scorer(corpus, 1.2, 0.75)
print(bm25.idf("bm25"))

Bayesian calibration + hybrid fusion

import bb25 as bb

corpus = bb.build_default_corpus()
docs = corpus.documents()
queries = bb.build_default_queries()

bm25 = bb.BM25Scorer(corpus, 1.2, 0.75)
bayes = bb.BayesianBM25Scorer(bm25, 1.0, 0.5)
vector = bb.VectorScorer()
hybrid = bb.HybridScorer(bayes, vector)

q = queries[0]
prob_or = hybrid.score_or(q.terms, q.embedding, docs[0])
prob_and = hybrid.score_and(q.terms, q.embedding, docs[0])
print("OR", prob_or, "AND", prob_and)

Run the experiments

import bb25 as bb

results = bb.run_experiments()
print(all(r.passed for r in results))

Sample script

See docs/sample_usage.py for an end-to-end example using BM25, Bayesian calibration, and hybrid fusion.

Benchmarks (BM25 vs Bayesian)

See benchmarks/README.md for a lightweight runner that compares BM25 and Bayesian BM25 on your own corpora.

English Benchmark (SQuAD, 100 validation queries)

This is where BB25 shines: Bayesian Hybrid beats the classic BM25 Hybrid.

Method	NDCG@10	MRR@10	Notes
WS (BB25+Dense)	0.9149	0.8850	SOTA!
WS (BM25+Dense)	0.9051	0.8717
RRF (BM25+Dense)	0.8874	0.8483	RRF underperforms weighted sum

Conclusion

"Bayesian BM25 (bb25) has demonstrated the potential to outperform classic BM25 in hybrid search."

On the English dataset (SQuAD), combining bb25 with Dense (BGE-M3) achieves higher performance than the BM25 + Dense baseline (+1.0%p NDCG). This suggests the probabilistic score from bb25 blends more smoothly with vector scores (less scale mismatch than a simple weighted sum).

Original paper and implementations:

• Paper: Bayesian BM25: A Probabilistic Framework for Hybrid Text and Vector Search
• Author's reference implementation (Python): cognica-io/bayesian-bm25
• This implementation (Rust + Python bindings): instructkr/bb25

Build from source (Rust)

make build

PyPI publishing

Build a wheel with maturin:

python -m pip install maturin
maturin build --release

For Pyodide builds, see docs/pyodide.md.