goldenfuzz

Fast, byte-identical-to-rapidfuzz fuzzy-string scorers, plus a one-vs-many extract / cdist / BatchComparator API. A thin PyO3 wheel over the pyo3-free goldenfuzz-core Rust crate.

import goldenfuzz as gf

gf.jaro_winkler("jonathan", "jonathon")      # -> 0.95...
gf.levenshtein("kitten", "sitting")          # normalized similarity in [0, 1]
gf.indel("fuzzy wuzzy", "wuzzy fuzzy")

# one-vs-many top-k (query bitmap built once)
gf.extract("jonathan smith",
           ["jon smith", "jane doe", "jonathan smith"],
           scorer="jaro_winkler", score_cutoff=0.7, limit=2)
# -> [(2, 1.0), (0, 0.9...)]

# reuse a prepared query across many choices
bc = gf.BatchComparator("acme corporation")
[bc.jaro_winkler(c) for c in choices]

Scorers: jaro_winkler | levenshtein | indel, each returning normalized similarity in [0, 1], byte-identical to the corresponding rapidfuzz metric (proven by an oracle fuzz in goldenfuzz-core). On short strings (names, addresses) goldenfuzz is faster than rapidfuzz; on documents it matches/beats on jaro-winkler and levenshtein.

fuzz.* composite scorers (drop-in for rapidfuzz fuzz)

The full weighted/token/partial family, each returning a score in [0, 100]:

gf.ratio("fname", "first_name")              # normalized indel, x100
gf.partial_ratio("fname", "first_name")      # best alignment of the shorter in the longer
gf.token_sort_ratio("a b c", "c b a")        # -> 100.0
gf.token_set_ratio("fuzzy was a bear", "fuzzy fuzzy was a bear")  # -> 100.0
gf.WRatio("fname", "first_name")             # weighted composite (rapidfuzz's fuzz.WRatio)
gf.QRatio("this is a test", "this is a test!")

Also token_ratio, partial_token_sort_ratio, partial_token_set_ratio, partial_token_ratio. Every one is verified byte-identical to rapidfuzz.fuzz over a 6.5k-pair corpus (all 10 scorers, worst abs diff 0.0), so it is a true drop-in — the value is the same, computed by our own kernel with no rapidfuzz at runtime.

Part of the golden suite.