sadit/similarity-search

Similarity Search

Nearest neighbor search benchmarks beyond vectors and beyond metrics. Seven datasets cover six kinds of objects: strings, probability distributions, point clouds, sets, vectors and bit codes. Each dataset comes with the dissimilarities it is meant to be searched with, 26 in all. Some are metrics, some are squares of metrics, and some are neither.

Every file holds a database, 1,000 queries held out from the same collection, and the exact 100 nearest neighbors of every query under every distance of the file. The layout follows the HDF5 files of the SISAP indexing challenges. The DBLP datasets and the bit codes come in several sizes. A larger database is always a superset of the smaller ones, with the same queries.

The datasets were built for the experiments on general dissimilarities of the manuscript A graph-based similarity search index with online automatic hyperparameter optimization (E. S. Tellez and G. Ruiz, under review). The tables below report exact search with ExhaustiveSearch, the BK-tree and the spatial access tree, and approximate search with SearchGraph tuned for a recall of 0.9, all from SimilaritySearch.jl. The code that builds the datasets, computes the ground truth and runs the benchmarks is in code/.

Contents

dataset	objects	database sizes	distances	source
dblp-author-names	DBLP author identifiers as strings	100K, 1M, 4M	Levenshtein, LCS, Jaro, Jaro–Winkler	DBLP, October 2026
dblp-author-topics	DBLP authors as distributions over 256 title terms	100K, 200K	Bray–Curtis, correlation, χ², Jensen–Shannon, Kullback–Leibler, Hellinger	DBLP, October 2026
dblp-author-years	DBLP authors as distributions over 92 publication years	100K, 200K	1-D earth mover's distance	DBLP, October 2026
dblp-author-clouds	DBLP authors as clouds of 20 title embeddings of 256 dimensions	100K, 200K	Hausdorff and Chamfer, over squared L2	DBLP, October 2026, and EmbeddingGemma-300m
uci-plants-sets	plants as sets of the regions where they grow, over 70 regions	33,781	Jaccard, Dice, Rogers–Tanimoto	UCI Plants
nasa	20-dimensional feature vectors	39,150	L0.5, L1, L2, squared L2, L∞, mean squared deviation, cosine	SISAP Metric Space Library
hiob-bits	384-bit binary sketches of text embeddings	100K, 1M, 3M	Hamming, Rogers–Tanimoto, Russell–Rao	GooAQ embeddings of the SISAP 2025 challenge

Every file has 1,000 queries. The files are <dataset>/<dataset>-n=<size>.h5, where the size label is the number of database objects: 100K is 100,000 and 1M is 1,000,000. The two collections used whole carry their exact size.

file	database	queries	distances	size	sha256
dblp-author-names/dblp-author-names-n=100K.h5	100,000	1,000	4	24 MB	a7979c838fab
dblp-author-names/dblp-author-names-n=1M.h5	1,000,000	1,000	4	109 MB	3468416af506
dblp-author-names/dblp-author-names-n=4M.h5	4,000,000	1,000	4	390 MB	67c4dff97266
dblp-author-topics/dblp-author-topics-n=100K.h5	100,000	1,000	6	126 MB	fae15a5e3bba
dblp-author-topics/dblp-author-topics-n=200K.h5	200,000	1,000	6	244 MB	dc219e9738fe
dblp-author-years/dblp-author-years-n=100K.h5	100,000	1,000	1	50 MB	618ae41b371b
dblp-author-years/dblp-author-years-n=200K.h5	200,000	1,000	1	97 MB	851ca72f694b
dblp-author-clouds/dblp-author-clouds-n=100K.h5	100,000	1,000	2	2.39 GB	fbdfab1da9a8
dblp-author-clouds/dblp-author-clouds-n=200K.h5	200,000	1,000	2	4.68 GB	91ad35f88e9c
uci-plants-sets/uci-plants-sets-n=33781.h5	33,781	1,000	3	6.9 MB	0806fd69f47d
nasa/nasa-n=39150.h5	39,150	1,000	7	8.9 MB	94f32c71a641
hiob-bits/hiob-bits-n=100K.h5	100,000	1,000	3	16 MB	da8bc3409943
hiob-bits/hiob-bits-n=1M.h5	1,000,000	1,000	3	60 MB	2ac70f7a90da
hiob-bits/hiob-bits-n=3M.h5	3,000,000	1,000	3	155 MB	30f78f2d9d83

File format

Each file follows the layout of the SISAP indexing challenges, with one group per distance:

path	content
train	the database, one object per row in h5py and NumPy, one per column in Julia
itest/queries	the 1,000 queries, in the same representation
itest/<distance>/knns	for each query, the 100 nearest database objects, as 1-based rows of train (int32, 1,000 × 100)
itest/<distance>/dists	their distances (float32), in ascending order
itest/<distance>/at_k, upto_k	for each query, the number of database objects at exactly, and at most, the distance of its 10th neighbor
meta/...	per-object metadata (meta/train/..., meta/itest/...) and the provenance of the sample

The queries are held out from the same collection before the database is drawn, so they follow the distribution of the database. The SISAP files call such queries itest, and we keep that name. The neighbor ids are 1-based, as in the SISAP files. Ties at equal distance are broken by ascending row. Each itest/<distance> group has these attributes: distance (the Julia expression that computes it), definition, metric, integer_valued, k (10), kgold (100), algo (exhaustive), querytime (seconds) and ties_at_k (the share of queries whose 10th neighbor is tied). The root attributes describe the dataset: the object and how to read it in Julia and Python, the list of distances, the source, the license, and how the sample was drawn.

The representation of train depends on the kind of object:

dataset	train	one object
dblp-author-names	UTF-8 strings, (n,)	the string; the distances compare Unicode characters
dblp-author-topics	float32, (n, 256)	a probability vector over the 256 terms of meta/vocabulary
dblp-author-years	float32, (n, 92)	a cumulative distribution over the years 1936–2027 (meta/years)
dblp-author-clouds	float32, (n, 20, 256)	20 points of 256 dimensions, unit norm
uci-plants-sets	variable-length int32, (n,)	a sorted set of 1-based region ids (meta/universe_names)
nasa	float32, (n, 20)	a vector
hiob-bits	uint64, (n, 6)	384 bits

Reading a file in Python:

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download("sadit/similarity-search", "dblp-author-names/dblp-author-names-n=100K.h5",
                       repo_type="dataset")
with h5py.File(path) as h:
    names = h["train"].asstr()[:]                        # 100,000 strings
    queries = h["itest/queries"].asstr()[:]              # 1,000 strings
    knns = h["itest/levenshtein/knns"][:] - 1            # 1,000 x 100, now 0-based rows of names
    dists = h["itest/levenshtein/dists"][:]
    print(dict(h["itest/levenshtein"].attrs))

Reading a file in Julia, and checking it with an exact search:

using HDF5, SimilaritySearch

h = h5open("dblp-author-names-n=100K.h5")
db = VectorDatabase(collect.(read(h["train"])))         # each name as a Vector{Char}
Q = VectorDatabase(collect.(read(h["itest/queries"])))
knns = read(h["itest/levenshtein/knns"])                 # 100 x 1,000, 1-based
ids, dists = searchbatch(ExhaustiveSearch(SimilaritySearch.Dist.Seqs.Levenshtein(), db), GenericContext(), Q, 10)

Point clouds read as a 256 × 20 × n array in Julia: MatrixDatabase(A[:, :, i]) is the cloud i, one point per column. code/hf/benchmark.jl reads every dataset this way.

Distances

The metric column says whether a distance is a metric on the objects (yes), the square of a metric (√d: its square root is a metric and ranks the neighbors in the same order), or neither (no). Exact metric indexes apply directly to the metrics, and to the square roots of the √d distances. The classification follows the manuscript, which also reports a random search for triangles that break the inequality. "Integer" marks the distances that take only integer values, which a BK-tree needs. Kullback–Leibler is not symmetric: the query is always the first argument.

dataset	distance	definition	metric	integer	Julia
names	levenshtein	the fewest insertions, deletions and substitutions of Unicode characters	yes	yes	Dist.Seqs.Levenshtein()
names	lcs	|a| + |b| − 2 LCS(a, b), the fewest insertions and deletions	yes	yes	Dist.Seqs.LCS()
names	jaro	1 − Jaro similarity	no	no	StringDistances.Jaro()
names	jaro-winkler	1 − Jaro–Winkler similarity: prefix scale 0.1 over at most 4 characters, applied when the Jaro distance is at most 0.3	no	no	StringDistances.JaroWinkler()
topics	bray-curtis	Σ|x − y| / Σ(x + y)	√d	no	Distances.BrayCurtis()
topics	correlation	1 − Pearson correlation	√d	no	Distances.CorrDist()
topics	chi-squared	Σ (x − y)² / (x + y)	√d	no	Distances.ChiSqDist()
topics	jensen-shannon	½ KL(x ‖ m) + ½ KL(y ‖ m), m = (x + y) / 2, natural logarithm	√d	no	Distances.JSDivergence()
topics	kullback-leibler	KL(query ‖ object) = Σ x log(x / y), natural logarithm	no	no	Distances.KLDivergence()
topics	hellinger	√(1 − Σ √(x y))	yes	no	Distances.HellingerDist()
years	emd-1d-years	L1 between the cumulative distributions: the earth mover's distance on the unit-spaced year axis	yes	no	Dist.L1()
clouds	hausdorff	max over the two directions of the max over one cloud of the squared distance to the nearest point of the other	√d	no	Dist.Cloud.Hausdorff(Dist.SqL2())
clouds	chamfer	mean over each cloud of the squared distance to the nearest point of the other, the two means added	no	no	Dist.Cloud.Chamfer(Dist.SqL2())
plants	jaccard	1 − |A ∩ B| / |A ∪ B|	yes	no	Dist.Sets.Jaccard()
plants	dice	1 − 2 |A ∩ B| / (|A| + |B|)	√d	no	Dist.Sets.Dice()
plants	rogers-tanimoto	2m / (70 + m), m = |A △ B|	yes	no	Dist.Sets.RogersTanimoto(70)
nasa	l0.5	(Σ |x − y|^0.5)²	√d	no	Dist.Lp(0.5)
nasa	l1	Σ |x − y|	yes	no	Dist.L1()
nasa	l2	√Σ (x − y)²	yes	no	Dist.L2()
nasa	sql2	Σ (x − y)²	√d	no	Dist.SqL2()
nasa	linf	max |x − y|	yes	no	Dist.LInfty()
nasa	msd	Σ (x − y)² / 20	√d	no	Distances.MeanSqDeviation()
nasa	cosine	1 − cos(x, y), on the raw vectors	√d	no	Dist.Cosine()
bits	hamming	the number of differing bits, 0 to 384	yes	yes	Dist.Bits.Hamming()
bits	rogers-tanimoto	2h / (384 + h), h the Hamming distance	yes	no	Dist.Bits.RogersTanimoto()
bits	russell-rao	1 − (bits set in both) / 384; d(x, x) is not 0	no	no	Dist.Bits.RussellRao()

Dist is SimilaritySearch.Dist. Distances.jl and StringDistances.jl distances work with SimilaritySearch.jl unchanged. Some pairs rank the neighbors in the same order: L2, squared L2 and MSD; Hamming and Rogers–Tanimoto on bit codes; and Jaccard and Dice. They are kept because an approximate search compares distance values, not only their order, so a monotone change of the distance changes the search even when the true neighbors stay the same.

Ground truth

For each query and distance, the gold standard is the exhaustive search over the whole database. Each distance is stored in single precision. The distances are sorted with a stable sort, so ties go to the lower row, and the 100 nearest are kept. The at_k and upto_k counts show how many database objects tie with the 10th neighbor. Ties are frequent under the integer-valued distances and on the sets. There, many different answers are equally correct, and recall by identifiers depends on which tied objects the gold standard happens to list. The benchmark tables therefore report two recalls:

• by identifiers, the share of the 10 true identifiers that the search returns, as in the SISAP challenges;
• with ties, the share of the 10 answers whose distance is at most the distance of the 10th true neighbor. This one is 1 for every exact search, and it is the one to read when the ties column is high.

Benchmarks

Measured on 3 identical nodes (Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz), 64 threads, Julia 1.12.7, SimilaritySearch.jl 1.6.8.

Each index answers the 1,000 queries with k = 10, in one batch over all the threads. Queries per second is the median of at least three timed passes. The percentage after it is the number of distance evaluations per query as a share of the database.

• ExhaustiveSearch evaluates every database object. It applies to every distance.
• BKT is a BK-tree. It applies to the integer-valued metrics: Levenshtein, LCS and Hamming.
• SAT is the spatial access tree (SpatialAccessTree.Sat, default construction). It applies to the metrics, and to the √d distances through their square root. Its answers are then exact for the original distance.
• SearchGraph is built incrementally with a logarithmic neighborhood of base 1.5 and the library's construction target, MinRecall(0.97). Its adjacency is then frozen, and optimize_index! tunes the search for MinRecall(0.9) on up to 1,024 database objects drawn with a fixed seed and masked from their own search. The tuning never sees the queries, so the recall on the queries can land a little below or above 0.9. The last column gives the construction and tuning times.

The ties column is the share of queries whose 10th neighbor is tied with the 11th or a later one. A dash means that the index does not apply to that distance.

dblp-author-names

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
Levenshtein	yes	100K	93%	229 qps	270 qps, 80%	247 qps, 87%	0.745 / 0.995	8.6K qps, 1.7%	26 + 5.1
Levenshtein	yes	1M	94%	23 qps	34 qps, 54%	30 qps, 64%	0.721 / 0.992	6.1K qps, 0.19%	201 + 8.1
Levenshtein	yes	4M	94%	6.1 qps	12 qps, 37%	10.0 qps, 46%	0.727 / 0.992	2.8K qps, 0.07%	1,376 + 21
LCS	yes	100K	91%	245 qps	293 qps, 76%	259 qps, 90%	0.811 / 0.996	10.0K qps, 1.9%	28 + 4.9
LCS	yes	1M	92%	24 qps	38 qps, 50%	27 qps, 71%	0.804 / 0.993	6.3K qps, 0.25%	227 + 8.0
LCS	yes	4M	91%	7.1 qps	14 qps, 34%	8.1 qps, 58%	0.808 / 0.998	2.6K qps, 0.18%	789 + 17
Jaro	no	100K	38%	184 qps	—	—	0.920 / 0.954	12K qps, 1.5%	41 + 4.8
Jaro	no	1M	49%	28 qps	—	—	0.926 / 0.975	8.3K qps, 0.23%	288 + 7.3
Jaro	no	4M	52%	4.0 qps	—	—	0.901 / 0.961	7.5K qps, 0.04%	1,589 + 16
Jaro–Winkler	no	100K	26%	220 qps	—	—	0.930 / 0.948	16K qps, 1.1%	36 + 4.7
Jaro–Winkler	no	1M	40%	28 qps	—	—	0.932 / 0.964	15K qps, 0.14%	205 + 6.7
Jaro–Winkler	no	4M	46%	6.3 qps	—	—	0.911 / 0.952	7.2K qps, 0.04%	625 + 16

dblp-author-topics

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
Bray–Curtis	√d	100K	0.60%	3.7K qps	—	1.1K qps, 100%	0.927 / 0.928	36K qps, 1.6%	18 + 4.2
Bray–Curtis	√d	200K	0.60%	1.5K qps	—	606 qps, 100%	0.925 / 0.926	24K qps, 1.1%	25 + 4.2
correlation	√d	100K	0%	23 qps	—	23 qps, 100%	0.909 / 0.909	4.0K qps, 0.55%	84 + 7.4
correlation	√d	200K	0.10%	12 qps	—	12 qps, 100%	0.919 / 0.919	2.9K qps, 0.36%	191 + 11
χ²	√d	100K	0%	3.3K qps	—	1.1K qps, 100%	0.933 / 0.933	45K qps, 2.0%	20 + 4.2
χ²	√d	200K	0.10%	1.6K qps	—	543 qps, 100%	0.904 / 0.904	29K qps, 1.1%	26 + 4.2
Jensen–Shannon	√d	100K	0%	41 qps	—	39 qps, 100%	0.942 / 0.942	1.6K qps, 2.1%	144 + 8.0
Jensen–Shannon	√d	200K	0%	20 qps	—	20 qps, 100%	0.915 / 0.915	1.5K qps, 1.1%	308 + 12
Kullback–Leibler	no	100K	0%	94 qps	—	—	0.914 / 0.914	4.9K qps, 1.5%	67 + 6.0
Kullback–Leibler	no	200K	0.10%	46 qps	—	—	0.915 / 0.915	3.9K qps, 0.99%	158 + 7.1
Hellinger	yes	100K	0%	914 qps	—	560 qps, 100%	0.949 / 0.949	22K qps, 2.2%	17 + 4.2
Hellinger	yes	200K	0%	434 qps	—	270 qps, 100%	0.911 / 0.911	20K qps, 1.1%	39 + 4.4

dblp-author-years

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
EMD 1-D	yes	100K	1.3%	6.6K qps	—	15K qps, 15%	0.934 / 0.936	308K qps, 0.31%	12 + 4.4
EMD 1-D	yes	200K	1.1%	3.3K qps	—	6.9K qps, 12%	0.927 / 0.929	319K qps, 0.18%	12 + 4.3

dblp-author-clouds

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
Hausdorff	√d	100K	0.50%	13 qps	—	13 qps, 100%	0.900 / 0.901	142 qps, 6.9%	1,662 + 47
Hausdorff	√d	200K	0.60%	6.1 qps	—	6.0 qps, 100%	0.886 / 0.886	110 qps, 4.5%	4,910 + 59
Chamfer	no	100K	0%	12 qps	—	—	0.937 / 0.937	616 qps, 1.6%	275 + 18
Chamfer	no	200K	0%	6.6 qps	—	—	0.940 / 0.940	376 qps, 1.3%	601 + 25

uci-plants-sets

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
Jaccard	yes	33,781	88%	16K qps	—	29K qps, 44%	0.670 / 0.995	222K qps, 2.7%	8.2 + 4.0
Dice	√d	33,781	88%	18K qps	—	20K qps, 57%	0.646 / 0.991	232K qps, 2.7%	8.3 + 4.0
Rogers–Tanimoto	yes	33,781	98%	15K qps	—	26K qps, 55%	0.553 / 0.999	252K qps, 2.9%	8.0 + 4.0

nasa

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
L0.5	√d	39,150	0%	1.1K qps	—	1.4K qps, 94%	0.902 / 0.902	141K qps, 0.67%	9.2 + 4.0
L1	yes	39,150	0%	31K qps	—	24K qps, 47%	0.946 / 0.946	507K qps, 0.59%	8.1 + 4.2
L2	yes	39,150	0%	39K qps	—	37K qps, 38%	0.935 / 0.935	705K qps, 0.53%	8.0 + 4.1
L2²	√d	39,150	0%	43K qps	—	39K qps, 36%	0.959 / 0.959	587K qps, 0.56%	8.2 + 4.5
L∞	yes	39,150	0%	39K qps	—	51K qps, 26%	0.956 / 0.956	570K qps, 0.81%	8.5 + 4.2
MSD	√d	39,150	0%	32K qps	—	35K qps, 37%	0.960 / 0.960	695K qps, 0.55%	8.0 + 4.0
cosine	√d	39,150	0%	3.2K qps	—	6.9K qps, 38%	0.931 / 0.931	305K qps, 0.51%	8.6 + 4.1

hiob-bits

distance	metric	n	ties	ExhaustiveSearch	BKT	SAT	SearchGraph recall (ids / ties)	SearchGraph	build + tune (s)
Hamming	yes	100K	39%	38K qps	9.7K qps, 100%	9.1K qps, 99%	0.947 / 0.972	492K qps, 0.62%	9.5 + 4.1
Hamming	yes	1M	43%	3.8K qps	434 qps, 95%	589 qps, 87%	0.935 / 0.960	264K qps, 0.07%	24 + 4.2
Hamming	yes	3M	40%	1.2K qps	116 qps, 91%	180 qps, 73%	0.877 / 0.900	336K qps, 0.02%	82 + 4.1
Rogers–Tanimoto	yes	100K	39%	14K qps	—	5.2K qps, 100%	0.933 / 0.959	456K qps, 0.55%	9.5 + 3.9
Rogers–Tanimoto	yes	1M	43%	1.5K qps	—	322 qps, 97%	0.949 / 0.976	246K qps, 0.06%	28 + 4.0
Rogers–Tanimoto	yes	3M	40%	482 qps	—	97 qps, 91%	0.942 / 0.964	246K qps, 0.02%	88 + 4.1
Russell–Rao	no	100K	60%	29K qps	—	—	0.906 / 0.958	462K qps, 0.83%	9.5 + 4.1
Russell–Rao	no	1M	60%	3.3K qps	—	—	0.889 / 0.936	257K qps, 0.08%	28 + 4.1
Russell–Rao	no	3M	62%	1.0K qps	—	—	0.908 / 0.954	195K qps, 0.03%	169 + 4.2

What the tables show. These are the measured ranges over the 49 pairs of file and distance.

• The spatial access tree evaluates between 12% and 100% of the database. It prunes well only on the year distributions, at 12% to 15%. On the topic profiles, the point clouds and the 100K bit codes it evaluates 99% or more.
• The BK-tree prunes more as the database grows. Under Levenshtein it evaluates 80% of the names at 100K, 54% at 1M and 37% at 4M. On the bit codes it evaluates 91% to 100%.
• In 22 of the 45 tree runs, the tree answers fewer queries per second than the exhaustive search.
• SearchGraph evaluates between 0.017% and 6.9% of the database, and it answers 10 to 1,900 times as many queries per second as the exhaustive search. Its recall with ties is between 0.886 and 0.999. Two cases land just below 0.9: Hausdorff on the 200K clouds, at 0.886, and Hamming on the 3M bit codes, at 0.877 by identifiers and 0.900 with ties.

The raw records, with more measures per index, are in results/.

How the datasets were built

Each dataset has a datasheet in code/ with every filter, what it discards, the statistics and the known issues. This section summarizes them. Every random draw uses a fixed seed, and every build records the checksums of its inputs and the commit of its code (see provenance/).

DBLP base tables

The three DBLP datasets start from the monthly release of the dblp computer science bibliography of 2026-10-03 (DOI 10.4230/dblp.xml.2026-10-03, CC0 1.0). The XML is read as a stream. It keeps the articles, conference papers and book chapters that have no publtype (CoRR preprints and other informal entries are left out) and that have at least one author. This leaves 7,418,235 papers. A second table lists every DBLP author identifier, 4,087,894 in all, with the author's papers. An identifier is DBLP's person name, with its 4-digit suffix when DBLP has told apart several people with the same name (Wei Wang 0001). A bare name next to suffixed ones is DBLP's catch-all profile for that name, and it can mix several people. The datasets keep such profiles and mark them in meta/.../kind. See code/dblp-2026-10-papers/README.md and code/dblp-2026-10-authors/README.md.

dblp-author-names

The objects are the author identifiers as strings, with their suffixes, capitalization and diacritics. One permutation of all 4,087,894 identifiers is drawn with numpy.random.default_rng(20261016) and stored. The queries are its first 1,000 entries, and the database is the next n entries. The 1M and 4M versions are therefore supersets of the 100K version, with the same queries. The median name has 14 characters. See code/dblp-author-names/README.md.

dblp-author-topics and dblp-author-years

Both datasets hold the same authors in the same order: the 243,930 authors with at least 20 papers and at least 20 occurrences of the vocabulary terms in their titles, permuted with the same seed. The queries are the first 1,000.

• Topics. Titles are lowercased, accents are folded, and the tokens of three or more letters are kept, minus English stop words and a short list of title boilerplate. The vocabulary is the 256 terms with the highest document frequency. An author's profile is the count of each term over all their titles, smoothed with 0.5 and normalized, so that Kullback–Leibler is defined.
• Years. An author's papers per year, from 1936 to 2027, normalized. train stores the cumulative distribution, so that the L1 distance between two rows is the 1-D earth mover's distance between the two careers.

The raw counts are in meta/train/term_counts and meta/train/year_counts. See code/dblp-author-distributions/README.md.

dblp-author-clouds

Each author is the cloud of the embeddings of 20 of their paper titles. The candidates are the 244,783 authors with at least 20 papers whose titles are not generic. A title is generic when it occurs at least 50 times in DBLP, such as "Editorial" or "Preface". The 20 papers of each author are drawn with a generator seeded by the author, so a cloud does not depend on which other authors are drawn. Each title is embedded with EmbeddingGemma-300m with its clustering prompt. The output is truncated to its first 256 of 768 components and renormalized to unit length. The distance between two points is the squared Euclidean distance, which on unit vectors equals 2 − 2 cos. The 200K version reuses the embeddings of the papers already in the 100K version and embeds the rest in the same way. The papers behind each cloud are in meta/train/cloud and meta/papers (DBLP key and title). See code/dblp-author-clouds/README.md.

uci-plants-sets

Each plant of the UCI Plants data, from the USDA PLANTS database, is the set of the US states and Canadian provinces where it occurs, over 70 regions. All 34,781 plants are used, with no deduplication. 1,000 of them, drawn with default_rng(20261016), are the queries, and the other 33,781 are the database. Many plants share the same set. 75% of the queries have an exact copy of their set in the database, so ties are massive. See code/uci-plants-sets/README.md.

nasa

The NASA collection of the SISAP Metric Space Library: 40,150 feature vectors of 20 dimensions extracted from images of NASA's archives. It is used as distributed, with 39,150 database vectors and 1,000 queries. See code/nasa/README.md.

hiob-bits

384-bit binary sketches of the GooAQ text embeddings of the SISAP 2025 indexing challenge (development set, 3,001,496 database vectors and 11,000 in-distribution queries). The sketches were trained and binarized with HIOB, with cosine as the metric of the embeddings. The database is a seeded sample of the codes: the first n entries of a stored permutation, sorted by source position, so the larger versions are supersets. The 1,000 queries are a seeded sample of the 11,000 query codes. See code/hiob-bits/README.md.

Reproducing

The builds use Python for the data preparation and Julia for the gold standards and the benchmarks.

• Python: code/pyproject.toml and code/uv.lock (uv).
• Julia 1.12: code/Project.toml and code/Manifest.toml, with SimilaritySearch.jl 1.6.8, Distances.jl, StringDistances.jl and HDF5.jl.

code/REPRODUCE.md lists the commands, from the raw sources to the published files and the tables. A benchmark of one file, for example:

julia --project=code -t 64 code/hf/benchmark.jl dblp-author-names/dblp-author-names-n=100K.h5 levenshtein

Known issues

• Ties. Recall by identifiers is ill-defined for most queries under Levenshtein and LCS and on the plant sets. Read the recall with ties or a distance-based error instead.
• People and identifiers. A DBLP catch-all identifier can mix several people. About 4% of the authors of the topics, years and clouds datasets are catch-alls. A few identifiers are organizations.
• Generic research words. The topic vocabulary keeps words such as "systems" or "analysis". Only explicit title boilerplate was removed. The vocabulary is English.
• Repeated points. About 4% of the clouds contain two papers with the same title, hence two identical points.
• NASA. How the library split the 1,000 queries from the 40,150 vectors is not recorded.

Licenses

The code in code/ and the benchmark records in results/ are under the MIT license, in code/LICENSE. The data files keep the licenses of their sources:

dataset	license of the data
dblp-author-names, dblp-author-topics, dblp-author-years	CC0 1.0, as the dblp data
dblp-author-clouds	CC0 1.0, as the dblp data; the embeddings were produced with EmbeddingGemma-300m under the Gemma Terms of Use
uci-plants-sets	CC BY 4.0, as UCI Plants. Cite: USDA, NRCS. 2008. The PLANTS Database (http://plants.usda.gov/, 31 December 2008). National Plant Data Center, Baton Rouge, LA 70874-4490 USA
nasa	none stated by the SISAP Metric Space Library
hiob-bits	Apache 2.0, as GooAQ

Citation

@misc{similarity-search,
  author       = {Tellez, Eric S. and Ruiz, Guillermo},
  title        = {Similarity Search: nearest neighbor benchmarks for strings, distributions, point clouds, sets, vectors and bit codes},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/sadit/similarity-search}}
}

Please also cite the sources: the dblp computer science bibliography (DOI 10.4230/dblp.xml.2026-10-03), the USDA PLANTS database as above, the SISAP Metric Space Library (K. Figueroa, G. Navarro and E. Chávez), GooAQ (D. Khashabi et al., 2021) and the SISAP 2025 indexing challenge.