Development build. This page describes main, not a released package. The latest published Lodestar.Fuzzy is 0.5.0 — read its documentation.

Home › Fuzzy

Performance — Lodestar.Fuzzy

What Lodestar.Fuzzy costs against the library a reader would otherwise reach for. How to read a row, and what this page leaves out: docs/guides/performance.md.

Levenshtein.Distance against Fastenshtein 1.0.12, Quickenshtein 1.5.1 and F23.StringSimilarity 7.0.1

Machine: AMD Ryzen 7 8700G w/ Radeon 780M Graphics, 1 CPU, 16 logical and 8 physical cores (BenchmarkDotNet's own header), Ubuntu 26.04.1 LTS, .NET SDK 10.0.401, .NET 10.0.12 runtime, AVX-512. Window: BenchmarkDotNet 0.14.0, default job, 2026-09-14, one run per class. The class and its agreement check: bench/README.md section 15.

Length	Lodestar	Fastenshtein	Quickenshtein	F23.StringSimilarity
8	15.71 ns, 0 B	46.11 ns, 56 B	48.27 ns, 0 B	103.42 ns, 128 B
64	178.57 ns, 0 B	3,339.85 ns, 280 B	727.63 ns, 0 B	5,443.03 ns, 576 B
512	7,097.59 ns, 0 B	239,725.19 ns, 2,072 B	19,810.70 ns, 0 B	452,681.32 ns, 4,160 B

Ahead of all three at every length: 2.8× to 4.1× against Quickenshtein, the closest, and 2.9× to 63.8× against the other two. Neither Lodestar nor Quickenshtein allocates.

Fuzz against Raffinert.FuzzySharp 6.0.0

Machine: AMD Ryzen 7 8700G w/ Radeon 780M Graphics, 1 CPU, 16 logical and 8 physical cores (BenchmarkDotNet's own header), Ubuntu 26.04.1 LTS, .NET SDK 10.0.401, .NET 10.0.12 runtime, AVX-512. Window: BenchmarkDotNet 0.14.0, default job, 2026-09-14, one run per class. The class and its agreement check: bench/README.md section 15.

Operation	Lodestar	FuzzySharp	Faster
Fuzz.Ratio	62.18 ns, 0 B	129.69 ns, 80 B	Lodestar, 2.09×
Fuzz.PartialRatio	452.36 ns, 0 B	5,767.01 ns, 160 B	Lodestar, 12.75×
Fuzz.TokenSetRatio	653.21 ns, 1,448 B	1,159.41 ns, 1,944 B	Lodestar, 1.78×
Fuzz.WRatio	1,346.20 ns, 2,760 B	2,938.48 ns, 3,128 B	Lodestar, 2.18×

Ahead on all four, and allocating less on each.

The score matrix against FuzzySharp and rapidfuzz (issue #1123)

Full method and what each row does and does not compare: bench/README.md. Machine: the AMD Ryzen 7 8700G named above, on 2026-09-23. BenchmarkDotNet 0.14.0, default job. Phrases of three words and an index, so no pair scores 100 by accident.

The .NET incumbent publishes no matrix call at all, checked by reflection: Raffinert.FuzzySharp.Process exports ExtractAll, ExtractTop, ExtractSorted and ExtractOne, and nothing that takes two collections. Its row is therefore ExtractAll once per query — the loop that library forces.

size	operation	Lodestar	FuzzySharp 6.0.0	FuzzySharp / Lodestar
50 × 50	score matrix, Ratio	89.86 μs	255.91 μs	2.85
200 × 200	score matrix, Ratio	1,974.46 μs	4,459.54 μs	2.26

Allocation is the wider margin: 19.55 KB against 301.95 KB at 50 a side, and 312.62 KB against 4,723.45 KB at 200 — 15×, because the matrix is one array where a per-query ExtractAll materialises a result object per pair.

Process.Cdist costs exactly what the same double loop written by hand against Fuzz.Ratio costs — 1.00× and 1.01× at the two sizes, byte for byte the same allocation. It is worth saying plainly: this member buys the shape, the cutoff applied where the score is produced, and one allocation instead of a list per query. It does not buy speed over what a caller could already write.

Against rapidfuzz 3.14.6 through compare-cdist, one run of each side, milliseconds per operation, best of five, dtype=np.float64 and workers=1 on the Python side so both compute the same thing:

n a side	scorer	Lodestar	rapidfuzz	ratio
50	Ratio	0.088 ms	0.018 ms	0.21
50	WRatio	1.125 ms	1.148 ms	1.02
200	Ratio	1.575 ms	0.229 ms	0.15
200	WRatio	17.319 ms	17.658 ms	1.02
500	Ratio	9.393 ms	1.383 ms	0.15
500	WRatio	104.467 ms	107.631 ms	1.03

The gap is the bulk shape, and the two scorers prove it. cdist builds each query's bit-parallel equality table once and scans every choice against it, where Fuzz.Ratio rebuilds it per pair. On Ratio that fixed cost is most of what a cell costs and the reference is ~7× ahead; on WRatio, which inspects its input and computes several sub-ratios, the same fixed cost is amortised and the two are level. Two rows of the same two libraries on the same corpus say more than any cross-language loop could. Issue #1130 carries the hoisting, which needs new Lodestar.Text API and the release order that comes with it.

A cold probe nearly went into this document. One unwarmed Stopwatch pass over 200 × 200 read 25.5 ms against the reference's 0.5 — 51×, all of it JIT. The harness and BenchmarkDotNet agree with each other at 1.58 ms and 1.97 ms and not with it.

The score matrix under a cutoff, against rapidfuzz (issue #1134)

Full method, the corpus rule and what the cutoff does to each side: bench/README.md. Machine: the AMD Ryzen 7 8700G named above, on 2026-09-23. compare-cdist, one run of each side, milliseconds per operation, best of five, dtype=np.float64 and workers=1 so both compute the same thing. Phrases of one to six words and an index, which is what gives a length bound anything to reject — the three-word corpus above holds every phrase to within a word of every other.

n a side	cutoff	Lodestar	rapidfuzz 3.14.6	ratio
50	none	0.080 ms	0.015 ms	0.19
50	90	0.033 ms	0.015 ms	0.47
200	none	1.460 ms	0.185 ms	0.13
200	90	0.624 ms	0.185 ms	0.30
500	none	8.335 ms	1.096 ms	0.13
500	90	3.416 ms	1.098 ms	0.32

A cutoff of 90 buys this package 2.3× to 2.4× and buys the reference nothing. rapidfuzz reads 0.185 ms against 0.185 and 1.096 against 1.098 with the cutoff and without it — inside its own noise. The Indel distance is at least the difference in lengths, so a pair whose lengths alone miss the cutoff cannot reach it; that rejection is what this package now takes, and the deficit against the reference halves — 0.13× becomes 0.30× at 200 a side and 0.32× at 500.

It does not close the gap, and the reason is the one section #1123 already named. What remains is the per-pair equality table Fuzz.Ratio rebuilds where cdist builds it once per query, which no cutoff can remove — only the hoisting #1130 carries, and the Lodestar.Text release order it needs.

The bound holds for the Indel ratio alone, so the rows above are the default scorer and there is no WRatio row to put beside them: partial_ratio("cat", "the cat sat on the mat") is 100 against a ceiling of 24, and a scorer passed by the caller has every pair scored.