reclink

Blazing-fast fuzzy matching and record linkage library powered by Rust.

Features

• 21 string similarity metrics — edit distance, token-based, subsequence, alignment, and hybrid metrics
• 10 phonetic algorithms — Soundex, Metaphone, Double Metaphone, NYSIIS, Caverphone, Cologne, Beider-Morse, Phonex, MRA, Daitch-Mokotoff
• Full record linkage pipeline — blocking, comparison, classification, and clustering
• 11 blocking strategies — exact, phonetic, sorted neighborhood, q-gram, LSH, canopy, trie, numeric, date, hybrid (union/intersection)
• 8 classifiers — threshold, weighted, bands, Fellegi-Sunter with EM estimation, logistic regression, decision tree
• 5 clustering algorithms — connected components, hierarchical, DBSCAN, OPTICS, incremental
• 7 index structures — BK-tree, VP-tree, N-gram index, memory-mapped N-gram, MinHash LSH, Bloom filter, inverted index
• DataFrame integration — pandas and polars accessors, native Polars plugin
• Parallel computation — Rayon-powered cdist and pipeline execution
• Scoring presets & composite scorer — pre-tuned configs for name, address, and general-purpose matching
• Extensible plugin system — register custom metrics, blockers, comparators, classifiers, and preprocessors
• WASM bindings — run reclink in the browser

Installation

pip install reclink

Build from source

git clone https://github.com/ByteVeda/reclink.git
cd reclink
uv sync --extra dev
maturin develop --release

Quick Start

Record linkage pipeline

import pandas as pd
from reclink.pipeline import ReclinkPipeline

df = pd.DataFrame({
    "id": ["1", "2", "3"],
    "first_name": ["Jon", "John", "Jane"],
    "last_name": ["Smith", "Smyth", "Doe"],
})

pipeline = (
    ReclinkPipeline.builder()
    .preprocess("first_name", ["fold_case", "strip_punctuation"])
    .preprocess("last_name", ["fold_case"])
    .block_phonetic("last_name", algorithm="soundex")
    .compare_string("first_name", metric="jaro_winkler")
    .compare_string("last_name", metric="jaro_winkler")
    .classify_threshold(0.85)
    .build()
)

matches = pipeline.dedup(df)
print(matches)
#    left_id right_id     score            scores
# 0       1        2  0.921...  [0.832..., 1.0...]

Direct metric usage

from reclink import jaro_winkler, soundex, cdist

jaro_winkler("Jon", "John")  # 0.93...
soundex("Smith") == soundex("Smyth")  # True
cdist(["Jon", "Jane"], ["John", "Janet"], scorer="jaro_winkler")  # 2x2 numpy array

Documentation

Full documentation at docs.byteveda.org/reclink, including:

• API Reference — every metric, algorithm, and class
• Guides — pipelines, preprocessing, DataFrames, custom plugins
• Interactive Playground — try reclink in your browser (WASM-powered)
• Changelog — release history

Performance

Pairwise comparison (10 string pairs, 500 iterations, microseconds per pair):

Metric	reclink	rapidfuzz	jellyfish	vs rapidfuzz	vs jellyfish
levenshtein	0.55	0.18	1.28	3.0x slower	2.3x faster
jaro	0.31	0.20	0.68	1.6x slower	2.2x faster
jaro_winkler	0.31	0.20	0.68	1.5x slower	2.2x faster
damerau_levenshtein	0.93	0.24	2.41	3.9x slower	2.6x faster

Batch matching (1,000 candidates, 50 iterations, microseconds per candidate):

Operation	reclink	rapidfuzz	thefuzz	vs rapidfuzz	vs thefuzz
match_batch (jaro_winkler)	0.32	0.13	1.60	2.4x slower	5.0x faster

Reproduce with python benchmarks/compare.py (requires pip install rapidfuzz jellyfish thefuzz).

Development

# Setup
uv sync --extra dev
uv run pre-commit install
maturin develop --release

# Rust
cargo test --workspace
cargo clippy -- -D warnings
cargo fmt --check

# Python
uv run pytest tests/python/ -v
uv run ruff check py_src/ tests/
uv run ruff format --check py_src/ tests/
uv run mypy py_src/reclink/

License

Apache-2.0