bm25-mojo

A drop-in faster replacement for the rank_bm25 package, powered by a clean-room Mojo kernel — with a vendored pure-Python fallback for platforms without a native build (including Windows — tested there in CI: windows-fallback).

from bm25_mojo import BM25Okapi  # same API as rank_bm25.BM25Okapi

bm25 = BM25Okapi(tokenized_corpus)
scores = bm25.get_scores(tokenized_query)        # np.ndarray, float64
top = bm25.get_top_n(tokenized_query, documents, n=5)
panel = bm25.get_scores_batch(list_of_queries)   # batch API: bit-identical rows

• Same results: scores match rank_bm25 element-wise (last-ulp level; the differential suite asserts agreement within 1e-8 on both the native and fallback backends). BM25Okapi, BM25L, and BM25Plus are provided.
• Much faster scoring: the kernel walks postings lists with the idf-weighted scores baked in at index time — 149x-13,137x faster get_scores than rank_bm25, and 1.1x-1.95x faster than bm25s 0.3.11 (numba, its recommended config) across the whole 1k-100k doc grid (Apple M4 Max; full method and numbers in the repository README and benchmarks/AUTOPSY-bm25s.md).
• No toolchain needed: per-platform wheels ship the compiled kernel. Everywhere else the package transparently uses its pure-Python fallback.
• Force the fallback with BM25_MOJO_DISABLE_NATIVE=1; inspect the active backend with bm25_mojo.backend_info().

Source, benchmarks, and development: https://github.com/thyn-ai/mojo-kernels

License: Apache-2.0, © 2026 Algenta