rustfuzz 0.1.13

rapid fuzzy string matching

---

[!WARNING] 🚧 Under Heavy Construction

This library is actively being developed and APIs may change between releases. We're shipping fast — expect frequent updates, new features, and occasional breaking changes. Pin your version if stability matters to you: pip install rustfuzz==0.1.12

---

🤖 This project was built entirely by AI.

The idea was simple: could an AI agent beat RapidFuzz — one of the fastest fuzzy matching libraries in the world — by writing a Rust-backed Python library from scratch, guided only by benchmarks?

The development loop was: Research → Build → Benchmark → Repeat.

---

rustfuzz is a blazing-fast fuzzy string matching library for Python — implemented entirely in Rust. 🚀

Zero Python overhead. Memory safe. Pre-compiled wheels for every major platform.

The Challenge: Beat RapidFuzz

flowchart LR
    R["🔍 Research<br>Profiler output<br>& algorithm gaps"]
    B["🦀 Build<br>Rust implementation<br>via PyO3"]
    T["✅ Test<br>All tests must pass<br>before proceeding"]
    BM["📊 Benchmark<br>vs RapidFuzz<br>Numbers don't lie"]
    RP["🔁 Repeat<br>Find the next<br>bottleneck"]

    R --> B --> T --> BM --> RP --> R

    style R fill:#6366f1,color:#fff,stroke:none
    style B fill:#a855f7,color:#fff,stroke:none
    style T fill:#ef4444,color:#fff,stroke:none
    style BM fill:#22c55e,color:#fff,stroke:none
    style RP fill:#f59e0b,color:#fff,stroke:none

The goal: match or exceed RapidFuzz's throughput on ratio, partial_ratio, token_sort_ratio, and process.extract — all from Python. Each iteration starts with profiling, identifies the hottest path, and rewrites it deeper into Rust.

The Results: RustFuzz is Faster 🏆

We benchmarked process.extract on a 1,000,000 row corpus. Thanks to zero-overhead Rayon parallelization, lock-free global threshold shrinking (AtomicU64), and native query token caching, rustfuzz officially outperforms rapidfuzz.

Benchmark (1M rows)	RapidFuzz	RustFuzz (Parallel)
Raw Characters (ratio)	5506 ms	5253 ms
Complex Tokens (WRatio)	3032 ms	2716 ms

But that's not all. By utilizing the built-in BM25 Hybrid Pipeline, rustfuzz can complete the identical extraction task in a revolutionary 97 ms (a ~30x speedup over state-of-the-art fuzzy matching!).

Features

⚡ Blazing Fast	Core algorithms written in Rust — no Python overhead, no GIL bottlenecks
🧠 Smart Matching	Ratio, partial ratio, token sort/set, Levenshtein, Jaro-Winkler, and more
🔒 Memory Safe	Rust's borrow checker guarantees — no segfaults, no buffer overflows
🐍 Pythonic API	Clean, typed Python interface. Import and go
📦 Zero Build Step	Pre-compiled wheels on PyPI for Python 3.10–3.14 on all major platforms
🏔️ Big Data Ready	Excels in 1 Billion Row Challenge benchmarks, crushing high-throughput tasks
🔍 3-Way Hybrid Search	BM25 + Fuzzy + Dense embeddings via RRF — 25ms at 1M docs, all in Rust
📄 Document Objects	First-class Document(content, metadata) + LangChain compatibility
🧩 Ecosystem Integrations	BM25, Hybrid Search, and LangChain Retrievers for Vector DBs (Qdrant, LanceDB, FAISS, etc.)

Installation

pip install rustfuzz
# or, with uv (recommended — much faster):
uv pip install rustfuzz

Quick Start

import rustfuzz.fuzz as fuzz
from rustfuzz.distance import Levenshtein

# Fuzzy ratio
print(fuzz.ratio("hello world", "hello wrold"))          # ~96.0

# Partial ratio (substring match)
print(fuzz.partial_ratio("hello", "say hello world"))    # 100.0

# Token-order-insensitive match
print(fuzz.token_sort_ratio("fuzzy wuzzy", "wuzzy fuzzy")) # 100.0

# Levenshtein distance
print(Levenshtein.distance("kitten", "sitting"))         # 3

# Normalised similarity [0.0 – 1.0]
print(Levenshtein.normalized_similarity("kitten", "kitten")) # 1.0

Batch extraction

from rustfuzz import process

choices = ["New York", "New Orleans", "Newark", "Los Angeles"]
print(process.extractOne("new york", choices))
# ('New York', 100.0, 0)

print(process.extract("new", choices, limit=3))
# [('Newark', ...), ('New York', ...), ('New Orleans', ...)]

3-Way Hybrid Search (BM25 + Fuzzy + Dense)

from rustfuzz.search import Document, HybridSearch

# Create documents with metadata
docs = [
    Document("Apple iPhone 15 Pro Max 256GB", {"brand": "Apple", "price": 1199}),
    Document("Samsung Galaxy S24 Ultra", {"brand": "Samsung", "price": 1299}),
    Document("Google Pixel 8 Pro", {"brand": "Google", "price": 699}),
]

# Optional: add dense embeddings for semantic search
embeddings = [[1.0, 0.0, 0.0], [0.9, 0.1, 0.0], [0.1, 0.9, 0.0]]

hs = HybridSearch(docs, embeddings=embeddings)

# Handles typos via fuzzy, keywords via BM25, meaning via dense — all in Rust
results = hs.search("appel iphon", query_embedding=[1.0, 0.0, 0.0], n=1)
text, score, meta = results[0]
print(f"{text} — ${meta['price']}")
# Apple iPhone 15 Pro Max 256GB — $1199

Also works with LangChain Document objects — no dependency required, auto-detected via duck-typing!

With Real Embeddings (FastEmbed)

Use FastEmbed for lightweight, local, ONNX-based embeddings — no GPU needed:

from fastembed import TextEmbedding
from rustfuzz.search import Document, HybridSearch

model = TextEmbedding("BAAI/bge-small-en-v1.5")  # ~33 MB, CPU-only

docs = [
    Document("Apple iPhone 15 Pro Max 256GB", {"brand": "Apple"}),
    Document("Samsung Galaxy S24 Ultra",      {"brand": "Samsung"}),
    Document("Sony WH-1000XM5 Headphones",    {"brand": "Sony"}),
]

embeddings = [e.tolist() for e in model.embed([d.content for d in docs])]
hs = HybridSearch(docs, embeddings=embeddings)

query = "wireless noise cancelling headset"
query_emb = list(model.embed([query]))[0].tolist()

results = hs.search(query, query_embedding=query_emb, n=1)
text, score, meta = results[0]
print(f"{text} — {meta['brand']}")
# Sony WH-1000XM5 Headphones — Sony

Supported Algorithms

Module	Algorithms
rustfuzz.fuzz	ratio, partial_ratio, token_sort_ratio, token_set_ratio, token_ratio, WRatio, QRatio, partial_token_*
rustfuzz.distance	Levenshtein, Hamming, Indel, Jaro, JaroWinkler, LCSseq, OSA, DamerauLevenshtein, Prefix, Postfix
rustfuzz.process	extract, extractOne, extract_iter, cdist
rustfuzz.search	BM25, BM25L, BM25Plus, BM25T, HybridSearch, Document
rustfuzz.utils	default_process

The BM25 Search Engines

rustfuzz.search implements lightning-fast Text Retrieval mathematical variants. The core differences:

• BM25 (Okapi): The industry standard. Employs term frequency saturation (logarithmic decay) and document length normalization.
• BM25L: Focuses on length penalization corrections. Introduces a static term shift delta, guaranteeing that matching terms yield a minimum baseline score even in massive documents where normalisation would normally suppress them.
• BM25Plus: Also creates a lower-bound for any given matching term, but applies the shift after term saturation. Widely considered the best default for highly mixed-length corpuses.
• BM25T: Introduces Information Gain adjustments to dynamically calculate the saturation limit $k_1$ per term, restricting dominant variance. rustfuzz hyper-optimises this by pre-computing term limits natively within the inverted index.

You can see an end-to-end benchmark comparison of these algorithms resolving the BEIR SciFact dataset in examples/bench_retrieval.py.

Documentation

Full cookbook with interactive examples and benchmark results: 👉 bmsuisse.github.io/rustfuzz

License

MIT © BM Suisse