alessandrobenigni/bm25-turbo-rust-python-wasm-cli

BM25涡轮 ⚡ Rust ·Python ·WASM ·CLI 最快的BM25计分引擎。就这样。

---

**每秒28,217次查询，880万份文档。8.6毫秒P50延迟。预计算稀疏BM25，支持BMW剪枝、内存映射持久性和零拷贝索引加载。专为RAG流水线、搜索应用和机器学习工作流构建。

为什么是BM25涡轮？

大多数 BM25 库在查询时计算分数——每次请求扫描倒置索引。BM25 Turbo 采取相反方法：在索引时预先计算每个 BM25 分数，形成压缩稀疏列（CSC）矩阵。查询变成稀疏向量查找，服务时间无数学计算。

这使得BM25 Turbo成为以下条件的合适选择：

• 你多次查询同一个索引（RAG、重新排序、批处理评估）
• 你需要确定性、可复现的评分（机器学习流水线、实验）
• 你需要最简单的API（3行索引，1行查询）
• 你是在构建一个搜索功能，不需要完整的搜索服务器

表演

基于MS MARCO（8,841,823份文档，509,962次查询）——标准信息检索基准测试。支持宝马修剪。单线程。

查询吞吐量

BM25 Turbo ⚡ Rust · Python · WASM · CLI The fastest BM25 scoring engine. Period.

---

28,217 queries/second on 8.8 million documents. 8.6ms P50 latency. Precomputed sparse BM25 with BMW pruning, memory-mapped persistence, and zero-copy index loading. Built for RAG pipelines, search applications, and ML workflows.

use bm25_turbo::{BM25Builder, Method};

let index = BM25Builder::new()
    .method(Method::Lucene)       // Robertson, Lucene, ATIRE, BM25L, BM25+
    .k1(1.5).b(0.75)
    .build_from_corpus(&[
        "Rust is a systems programming language",
        "BM25 is a ranking function used in information retrieval",
        "Machine learning models benefit from fast retrieval",
    ])?;

let results = index.search("information retrieval", 10)?;
for (id, score) in results.doc_ids.iter().zip(results.scores.iter()) {
    println!("doc {} → {:.4}", id, score);
}

Why BM25 Turbo?

Most BM25 libraries compute scores at query time — scanning inverted indexes on every request. BM25 Turbo takes the opposite approach: precompute every BM25 score at index time into a compressed sparse column (CSC) matrix. Queries become sparse vector lookups with no math at serving time.

This makes BM25 Turbo the right choice when:

• You query the same index many times (RAG, reranking, batch evaluation)
• You need deterministic, reproducible scores (ML pipelines, experiments)
• You want the simplest possible API (3 lines to index, 1 line to query)
• You’re building a search feature and don’t need a full search server

Performance

Benchmarked on MS MARCO (8,841,823 documents, 509,962 queries) — the standard information retrieval benchmark. BMW pruning enabled. Single-threaded.

Query Throughput

Query Latency

Scaling Across Corpus Sizes