pith. machine review for the scientific record. sign in

arxiv: 2604.02801 · v1 · submitted 2026-04-03 · 💻 cs.DB

Recognition: no theorem link

Distance Comparison Operations Are Not Silver Bullets in Vector Similarity Search: A Benchmark Study on Their Merits and Limits

Zhuanglin Zheng , Yuxiang Zeng , Chenchen Liu , Yunzhen Chi , Binhan Yang , Yongxin Tong
Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:55 UTC · model grok-4.3

classification 💻 cs.DB
keywords vector similarity searchdistance comparison operationsbenchmark studyperformance evaluationproduction deploymenthigh-dimensional dataindex constructiondata updates
0
0 comments X

The pith

Distance comparison operations prove unstable and often slower than full scans in vector similarity search benchmarks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper benchmarks eight distance comparison operation algorithms across ten datasets with up to 100 million vectors and 12,288 dimensions, running on CPUs with and without SIMD as well as GPUs. Results show these methods are highly sensitive to dimensionality and query distribution, with performance degrading sharply on out-of-distribution queries and varying across hardware. While DCOs can accelerate index construction and data updates, their query performance instability, which sometimes exceeds full-dimensional scan times, supports the conclusion that recent DCO advances are not ready for production vector database deployment.

Core claim

Distance Comparison Operations (DCOs) decide whether the distance between a data vector and a query lies within a threshold without full-dimensional computation. Comprehensive evaluation reveals that DCO efficiency depends heavily on data dimensionality, degrades under out-of-distribution queries, and fluctuates across hardware platforms. These operations speed up index construction and updates yet can run slower than complete distance scans, indicating they remain unsuitable for production vector database systems.

What carries the argument

Distance Comparison Operations (DCOs) that decide if the distance between a data vector and a query is within a threshold, acting as a potential shortcut to avoid full-dimensional distance computations.

If this is right

  • Index construction in vector databases can run faster when using DCO methods.
  • Data update operations benefit from the reduced computation offered by DCOs.
  • Query latency stays unpredictable and can exceed full-scan costs on high-dimensional or shifted data.
  • Hardware differences such as SIMD availability or GPU usage strongly affect DCO results.
  • Production vector systems should avoid relying on current DCO implementations for query workloads.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Adaptive DCO variants that adjust thresholds based on detected query distribution shifts could reduce the observed performance drops.
  • Hybrid query pipelines that fall back to full scans when DCO confidence is low might deliver more reliable speedups.
  • Broader testing on streaming or user-generated query logs from deployed systems would clarify real-world applicability.
  • Vector database designers may focus on other indexing or pruning techniques until DCO stability improves.

Load-bearing premise

The chosen datasets, query distributions, and hardware configurations sufficiently represent real-world production vector database workloads and edge cases.

What would settle it

A new experiment on large-scale data showing consistent DCO speedups over full scans across all dimensions, query distributions, and hardware platforms would challenge the instability finding.

Figures

Figures reproduced from arXiv: 2604.02801 by Binhan Yang, Chenchen Liu, Yongxin Tong, Yunzhen Chi, Yuxiang Zeng, Zhuanglin Zheng.

Figure 1
Figure 1. Figure 1: Instances of Distance Comparison Operation (DCO) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: The three categories of DCO methods: core ideas and main workflows [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance on ultra-high-dimensional dataset XUltra 50 at [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results on in-distribution and OOD queries [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Dimension pruning via DCO and its impact on recall [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Index construction time using DCO Index Construction Time [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance on distance metric of Inner Product (IP) [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: QPS, recall, and total update time of DCO methods [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Performance with limited initial data Performance with Limited Initial Dataset. Most methods, except for FDScanning and PDScanning, rely heavily on the initial data for training prediction models or computing PCA matrices. To evaluate their sensitivity, we test their performance across a wide range of initial dataset sizes (i.e., from 0.01% to 100% of the full dataset) and report the resulting recall and … view at source ↗
Figure 4
Figure 4. Figure 4: This performance inversion is driven by the IVF index’s [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 14
Figure 14. Figure 14: Recommendations of DCO methods The optimal DCO algorithm selection depends on three primary factors: data dimensionality, hardware configurations, and query distribution. For example, on CPUs with SIMD support, the best choice for in-distribution queries shifts from FDScanning to DDCpca, and then to PDScanning, as data dimensionality increases. Overall, simple scanning based methods, FDScanning and PDScan… view at source ↗
Figure 15
Figure 15. Figure 15: Impact of initial step size ∆0 Overall Query Performance without SIMD. Compared to the result in [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗
Figure 18
Figure 18. Figure 18: Performance comparison on AVX2 85 90 95 100 Recall (%) 200 1000 2000 QPS FDScanning PDScanning PDScanning+ ADSampling DADE DDCres DDCpca DDCopq 98.0 98.5 99.0 99.5 100.0 Recall (%) 200 500 1500 QPS (a) Laion (in-distribution) 70 80 90 100 Recall (%) 200 500 1500 QPS (b) Laion (OOD) 97.5 98.0 98.5 99.0 99.5 100.0 Recall (%) 300 1000 1500 QPS (c) Text2Image (in-distribution) 80 85 90 95 100 Recall (%) 200 1… view at source ↗
Figure 19
Figure 19. Figure 19: Results on in-distribution and OOD queries without SIMD [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Search performance using cosine similarity [PITH_FULL_IMAGE:figures/full_fig_p018_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Comparisons of DCO methods using the IVF index on XUltra dataset across different hardware environments [PITH_FULL_IMAGE:figures/full_fig_p018_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Dimension pruning via DCO and its impact on recall on IVF [PITH_FULL_IMAGE:figures/full_fig_p019_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Impact of initial step size ∆0 32 64 96 128 160 192 224 0 600 800 1000 1200 QPS FDScanning PDScanning PDScanning+ ADSampling DADE DDCres DDCpca 8 16 32 64 96 128 160 d 500 750 1000 1250 QPS (a) HNSW 32 64 96 128 160 192 224 d 100 200 300 QPS (b) IVF [PITH_FULL_IMAGE:figures/full_fig_p019_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Impact of incremental step size ∆d Overall Query Performance without SIMD. Compared to the result in [PITH_FULL_IMAGE:figures/full_fig_p019_24.png] view at source ↗
Figure 26
Figure 26. Figure 26: Performance comparison on AVX2 85 90 95 100 Recall (%) 200 1000 2000 QPS FDScanning PDScanning PDScanning+ ADSampling DADE DDCres DDCpca DDCopq 98.0 98.5 99.0 99.5 100.0 Recall (%) 200 500 1500 QPS (a) Laion (in-distribution) 70 80 90 100 Recall (%) 200 500 1500 QPS (b) Laion (OOD) 97.5 98.0 98.5 99.0 99.5 100.0 Recall (%) 300 1000 1500 QPS (c) Text2Image (in-distribution) 80 85 90 95 100 Recall (%) 200 1… view at source ↗
Figure 27
Figure 27. Figure 27: Results on in-distribution and OOD queries without SIMD [PITH_FULL_IMAGE:figures/full_fig_p022_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Search performance using cosine similarity [PITH_FULL_IMAGE:figures/full_fig_p023_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Comparisons of DCO methods using the IVF index on XUltra dataset across different hardware environments [PITH_FULL_IMAGE:figures/full_fig_p023_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Dimension pruning via DCO and its impact on recall on IVF [PITH_FULL_IMAGE:figures/full_fig_p023_30.png] view at source ↗
read the original abstract

Distance Comparison Operations (DCOs), which decide whether the distance between a data vector and a query is within a threshold, are a critical performance bottleneck in vector similarity search. Recent DCO methods that avoid full-dimensional distance computations promise significant speedups, but their readiness for production vector database systems remains an open question. To address this, we conduct a comprehensive benchmark of 8 DCO algorithms across 10 datasets (with up to 100M vectors and 12,288 dimensions) and diverse hardware configurations (CPUs with/without SIMD, and GPUs). Our study reveals that these methods are not silver bullets: their efficiency is highly sensitive to data dimensionality, degrades under out-of-distribution queries, and is unstable across hardware. Yet, our evaluation also demonstrates often-overlooked merits: they can accelerate index construction and data updates. Despite these benefits, their unstable performance, which can be slower than a full-dimensional scan, leads us to conclude that recent algorithmic advancements in DCO are not yet ready for production deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript benchmarks eight distance comparison operation (DCO) algorithms across ten datasets (up to 100M vectors and 12,288 dimensions) on CPU (with/without SIMD) and GPU hardware. It reports that DCO performance is highly sensitive to dimensionality, degrades under out-of-distribution queries, exhibits instability that can exceed the cost of full-dimensional scans, yet offers benefits for index construction and data updates; the authors conclude that recent DCO advances are not yet ready for production vector-database deployment.

Significance. If the empirical findings hold, the work is significant for vector-database practitioners and researchers: it supplies large-scale evidence against treating DCO methods as default optimizations and identifies concrete failure modes (dimensionality sensitivity, OOD degradation, hardware instability) that future algorithms must address. The scale of the evaluation (multiple datasets, hardware platforms, index-construction and update scenarios) is a strength that could usefully inform both system design and follow-on research.

major comments (2)
  1. [Section 5] Section 5 (Results and Analysis): the central claim that DCO methods 'can be slower than a full-dimensional scan' and are therefore 'not yet ready for production' is load-bearing; the manuscript should report, for each dataset and hardware configuration, the fraction of queries or index-build phases where DCO exceeds the baseline scan time, together with confidence intervals, so readers can assess how frequently the slowdown occurs rather than only that it is possible.
  2. [Section 4] Section 4 (Experimental Setup): the generalization from the observed instability to a production-readiness verdict rests on the assumption that the ten datasets, query distributions (including OOD), and hardware configurations are representative of real vector-DB workloads; the paper does not provide a justification or sensitivity analysis for this choice, which weakens the deployment conclusion.
minor comments (2)
  1. [Table 2] Table 2 and Figure 4: axis labels and legends should explicitly state whether reported times include or exclude data-transfer overhead on GPU; this affects interpretation of the hardware-stability claims.
  2. [§3] The abstract states 'unstable performance' without defining the metric (e.g., coefficient of variation across runs or across datasets); a short definition in §3 would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our benchmark study. We address each major comment below and have revised the manuscript to incorporate the suggested improvements where feasible.

read point-by-point responses

  1. Referee: [Section 5] Section 5 (Results and Analysis): the central claim that DCO methods 'can be slower than a full-dimensional scan' and are therefore 'not yet ready for production' is load-bearing; the manuscript should report, for each dataset and hardware configuration, the fraction of queries or index-build phases where DCO exceeds the baseline scan time, together with confidence intervals, so readers can assess how frequently the slowdown occurs rather than only that it is possible.

    Authors: We agree that reporting the frequency of slowdowns with statistical support would strengthen the evidence. In the revised Section 5, we have added tables for each dataset and hardware configuration (CPU with/without SIMD, GPU) showing the fraction of queries and index-build phases where DCO runtime exceeds the full scan baseline, accompanied by 95% confidence intervals computed across repeated runs. This quantifies prevalence rather than possibility alone. revision: yes

  2. Referee: [Section 4] Section 4 (Experimental Setup): the generalization from the observed instability to a production-readiness verdict rests on the assumption that the ten datasets, query distributions (including OOD), and hardware configurations are representative of real vector-DB workloads; the paper does not provide a justification or sensitivity analysis for this choice, which weakens the deployment conclusion.

    Authors: We acknowledge the value of explicit justification. The revised Section 4 now includes a new subsection explaining the dataset selection criteria (standard benchmarks from prior literature covering diverse dimensionalities up to 12k, sizes up to 100M, and both ID/OOD queries) and hardware choices as representative of common production environments. We also added a sensitivity analysis varying query distributions, confirming that instability patterns persist. While exhaustive coverage of all workloads is impossible, these additions better ground our conclusions. revision: yes

Circularity Check

0 steps flagged

Pure empirical benchmark with direct measurements; no derivations or self-referential predictions

full rationale

The paper is a benchmark study that reports measured runtimes, speedups, and stability of 8 DCO algorithms across 10 datasets (up to 100M vectors, 12k dimensions) and CPU/GPU hardware. All load-bearing claims (unstable performance, occasional slowdowns vs. full scan, sensitivity to OOD queries) are direct empirical observations rather than outputs of any derivation, fitted parameter, or equation. No self-citation chain or ansatz is invoked to justify the central conclusion; the results stand on the reported measurements themselves. This is the expected non-circular outcome for a pure empirical evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities; the work rests on standard empirical benchmarking of existing algorithms.

pith-pipeline@v0.9.0 · 5497 in / 986 out tokens · 36693 ms · 2026-05-13T18:55:36.621774+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages · 2 internal anchors

  1. [1]

    High-dimensional approximate nearest neighbor search: with reliable and efficient distance comparison operations,

    J. Gao and C. Long, “High-dimensional approximate nearest neighbor search: with reliable and efficient distance comparison operations,”Proc. ACM Manag. Data, vol. 1, no. 2, pp. 137:1–137:27, 2023

    work page 2023

  2. [2]

    Effi- cient data-aware distance comparison operations for high-dimensional approximate nearest neighbor search,

    L. Deng, P. Chen, X. Zeng, T. Wang, Y . Zhao, and K. Zheng, “Effi- cient data-aware distance comparison operations for high-dimensional approximate nearest neighbor search,”PVLDB, vol. 18, no. 3, pp. 812– 821, 2024

    work page 2024

  3. [3]

    Effective and general distance computation for approximate nearest neighbor search,

    M. Yang, W. Li, J. Jin, X. Zhong, X. Wang, Z. Shen, W. Jia, and W. Wang, “Effective and general distance computation for approximate nearest neighbor search,” inICDE, 2025, pp. 1098–1110

    work page 2025

  4. [4]

    Dimensionality-reduction techniques for approximate nearest neighbor search: A survey and evaluation,

    Z. Wang, H. Xiong, Q. Wang, Z. He, P. Wang, T. Palpanas, and W. Wang, “Dimensionality-reduction techniques for approximate nearest neighbor search: A survey and evaluation,”IEEE Data Eng. Bull., vol. 48, no. 3, pp. 63–80, 2024

    work page 2024

  5. [5]

    ANSMET: approximate nearest neighbor search with near-memory processing and hybrid early termination,

    Y . Li, Y . Jin, B. Tian, H. Zhang, and M. Gao, “ANSMET: approximate nearest neighbor search with near-memory processing and hybrid early termination,” inISCA, 2025, pp. 1093–1107

    work page 2025

  6. [6]

    Dforest: A minimal dimensionality-aware indexing for high-dimensional exact similarity search,

    L. Li, W. Sun, and B. Wu, “Dforest: A minimal dimensionality-aware indexing for high-dimensional exact similarity search,”IEEE Trans. Knowl. Data Eng., vol. 36, no. 10, pp. 5092–5105, 2024

    work page 2024

  7. [7]

    Milvus: A purpose-built vector data management system,

    J. Wang, X. Yi, R. Guo, H. Jin, P. Xu, S. Li, X. Wang, X. Guo, C. Li, X. Xu, K. Yu, Y . Yuan, Y . Zou, J. Long, Y . Cai, Z. Li, Z. Zhang, Y . Mo, J. Gu, R. Jiang, Y . Wei, and C. Xie, “Milvus: A purpose-built vector data management system,” inSIGMOD, 2021, pp. 2614–2627

    work page 2021

  8. [8]

    Weaviate

    “Weaviate.” [Online]. Available: https://github.com/weaviate/weaviate

    work page

  9. [9]

    Available: https://github.com/qdrant/qdrant

    “Qdrant.” [Online]. Available: https://github.com/qdrant/qdrant

    work page

  10. [10]

    Source code and datasets in the benchmark

    “Source code and datasets in the benchmark.” [Online]. Available: https://github.com/zzlin237/dco-benchmark.git

    work page

  11. [11]

    Gaussdb- vector: A large-scale persistent real-time vector database for LLM applications,

    G. Li, J. Sun, J. Pan, J. Wang, Y . Xie, R. Liu, and W. Nie, “Gaussdb- vector: A large-scale persistent real-time vector database for LLM applications,”PVLDB, vol. 18, no. 12, pp. 4951–4963, 2025

    work page 2025

  12. [12]

    A comprehensive survey on vector database: Storage and retrieval technique, challenge.Computing Research Repository, abs/2310.11703, 2023

    Y . Han, C. Liu, and P. Wang, “A comprehensive survey on vec- tor database: Storage and retrieval technique, challenge,”CoRR, vol. abs/2310.11703, 2023

    work page arXiv 2023

  13. [13]

    Casella and R

    G. Casella and R. Berger,Statistical inference. Chapman and Hall/CRC, 2024

    work page 2024

  14. [14]

    James, D

    G. James, D. Witten, T. Hastie, and R. Tibshirani,An introduction to statistical learning: with applications in R. Springer, 2013, vol. 103

    work page 2013

  15. [15]

    Szeliski,Computer vision: algorithms and applications

    R. Szeliski,Computer vision: algorithms and applications. Springer Nature, 2022

    work page 2022

  16. [16]

    Manning and H

    C. Manning and H. Schutze,Foundations of statistical natural language processing. MIT press, 1999

    work page 1999

  17. [17]

    Recent advances in natural language processing via large pre-trained language models: A survey,

    B. Min, H. Ross, E. Sulem, A. P. B. Veyseh, T. H. Nguyen, O. Sainz, E. Agirre, I. Heintz, and D. Roth, “Recent advances in natural language processing via large pre-trained language models: A survey,”ACM Comput. Surv., vol. 56, no. 2, pp. 30:1–30:40, 2024

    work page 2024

  18. [18]

    A survey on stability of learning with limited labelled data and its sensitivity to the effects of random- ness,

    B. Pecher, I. Srba, and M. Bielikov ´a, “A survey on stability of learning with limited labelled data and its sensitivity to the effects of random- ness,”ACM Comput. Surv., vol. 57, no. 1, pp. 19:1–19:40, 2025

    work page 2025

  19. [19]

    C. D. Toth, J. O’Rourke, and J. E. Goodman,Handbook of discrete and computational geometry. CRC press, 2017

    work page 2017

  20. [20]

    A comprehensive survey and experimental comparison of graph-based approximate nearest neighbor search,

    M. Wang, X. Xu, Q. Yue, and Y . Wang, “A comprehensive survey and experimental comparison of graph-based approximate nearest neighbor search,”PVLDB, vol. 14, no. 11, pp. 1964–1978, 2021

    work page 1964

  21. [21]

    Survey of vector database management systems,

    J. J. Pan, J. Wang, and G. Li, “Survey of vector database management systems,”VLDB J., vol. 33, no. 5, pp. 1591–1615, 2024

    work page 2024

  22. [22]

    Billion-scale similarity search with gpus,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with gpus,”IEEE Trans. Big Data, vol. 7, no. 3, pp. 535–547, 2021

    work page 2021

  23. [23]

    The Faiss library

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P. Mazar ´e, M. Lomeli, L. Hosseini, and H. J ´egou, “The faiss library,”CoRR, vol. abs/2401.08281, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    FINGER: fast inference for graph-based approximate nearest neighbor search,

    P. H. Chen, W. Chang, J. Jiang, H. Yu, I. S. Dhillon, and C. Hsieh, “FINGER: fast inference for graph-based approximate nearest neighbor search,” inWWW, 2023, pp. 3225–3235

    work page 2023

  25. [25]

    Fast approximate nearest neighbors with automatic algorithm configuration,

    M. Muja and D. G. Lowe, “Fast approximate nearest neighbors with automatic algorithm configuration,” inVISAPP, 2009, pp. 331–340

    work page 2009

  26. [26]

    Principal component analysis,

    H. Abdi and L. J. Williams, “Principal component analysis,”WIRES COMPUT STAT, vol. 2, no. 4, pp. 433–459, 2010

    work page 2010

  27. [27]

    Efficient and robust approxi- mate nearest neighbor search using hierarchical navigable small world graphs,

    Y . A. Malkov and D. A. Yashunin, “Efficient and robust approxi- mate nearest neighbor search using hierarchical navigable small world graphs,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 42, no. 4, pp. 824–836, 2020

    work page 2020

  28. [28]

    The inverted multi-index,

    A. Babenko and V . S. Lempitsky, “The inverted multi-index,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 37, no. 6, pp. 1247–1260, 2015

    work page 2015

  29. [29]

    Big ANN Benchmark,

    “Big ANN Benchmark,” 2024. [Online]. Available: https: //big-ann-benchmarks.com/

    work page 2024

  30. [30]

    GloVe: Global Vectors for Word Representation

    “GloVe: Global Vectors for Word Representation.” [Online]. Available: https://nlp.stanford.edu/projects/glove/

    work page

  31. [31]

    LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki, “LAION-400M: open dataset of clip-filtered 400 million image-text pairs,”CoRR, vol. abs/2111.02114, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  32. [32]

    MS MARCO: A human generated machine reading comprehension dataset,

    T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder, and L. Deng, “MS MARCO: A human generated machine reading comprehension dataset,” inCoCo@NIPS, vol. 1773, 2016

    work page 2016

  33. [33]

    New and improved embedding model,

    “New and improved embedding model,” https://openai.com/index/ new-and-improved-embedding-model/

    work page

  34. [34]

    Distance comparison operations are not silver bullets in vector similarity search: An experimental study on their merits and limits (full paper)

    “Distance comparison operations are not silver bullets in vector similarity search: An experimental study on their merits and limits (full paper).” [Online]. Available: https://github.com/zzlin237/ dco-benchmark.git

    work page

  35. [35]

    Diskann: Fast accurate billion-point nearest neighbor search on a single node,

    S. Jayaram Subramanya, F. Devvrit, H. V . Simhadri, R. Krishnawamy, and R. Kadekodi, “Diskann: Fast accurate billion-point nearest neighbor search on a single node,”NeurIPS, vol. 32, 2019

    work page 2019

  36. [36]

    Model-enhanced vector index,

    H. Zhang, Y . Wang, Q. Chen, R. Chang, T. Zhang, Z. Miao, Y . Hou, Y . Ding, X. Miao, H. Wang, B. Pang, Y . Zhan, H. Sun, W. Deng, Q. Zhang, F. Yang, X. Xie, M. Yang, and B. Cui, “Model-enhanced vector index,” inNeurIPS, 2023

    work page 2023

  37. [37]

    Intel® 64 and ia-32 architectures software developer’s man- ual,

    P. Guide, “Intel® 64 and ia-32 architectures software developer’s man- ual,”Volume 3B: system programming guide, Part, vol. 2, no. 11, pp. 0–40, 2011

    work page 2011

  38. [38]

    Approximate nearest neighbor search on high dimensional data - experiments, analyses, and improvement,

    W. Li, Y . Zhang, Y . Sun, W. Wang, M. Li, W. Zhang, and X. Lin, “Approximate nearest neighbor search on high dimensional data - experiments, analyses, and improvement,”IEEE Trans. Knowl. Data Eng., vol. 32, no. 8, pp. 1475–1488, 2020

    work page 2020

  39. [39]

    Similarity query processing for high-dimensional data,

    J. Qin, W. Wang, C. Xiao, and Y . Zhang, “Similarity query processing for high-dimensional data,”PVLDB, vol. 13, no. 12, pp. 3437–3440, 2020

    work page 2020

  40. [40]

    High-dimensional similarity query processing for data science,

    J. Qin, W. Wang, C. Xiao, Y . Zhang, and Y . Wang, “High-dimensional similarity query processing for data science,” inKDD, 2021, pp. 4062– 4063

    work page 2021

  41. [41]

    HVS: hierarchical graph structure based on voronoi diagrams for solving approximate nearest neighbor search,

    K. Lu, M. Kudo, C. Xiao, and Y . Ishikawa, “HVS: hierarchical graph structure based on voronoi diagrams for solving approximate nearest neighbor search,”PVLDB, vol. 15, no. 2, pp. 246–258, 2021

    work page 2021

  42. [42]

    Accelerating approximate nearest neighbor search in hierarchical graphs: Efficient level navigation with shortcuts,

    Z. Gong, Y . Zeng, and L. Chen, “Accelerating approximate nearest neighbor search in hierarchical graphs: Efficient level navigation with shortcuts,”PVLDB, vol. 18, no. 10, pp. 3518–3530, 2025

    work page 2025

  43. [43]

    Incremental ivf index maintenance for streaming vector search,

    J. Mohoney, A. Pacaci, S. R. Chowdhury, U. F. Minhas, J. Pound, C. Renggli, N. Reyhani, I. F. Ilyas, T. Rekatsinas, and S. Venkataraman, “Incremental ivf index maintenance for streaming vector search,”arXiv preprint arXiv:2411.00970, 2024

    work page arXiv 2024

  44. [44]

    PM-LSH: a fast and accurate in-memory framework for high- dimensional approximate NN and closest pair search,

    B. Zheng, X. Zhao, L. Weng, Q. V . H. Nguyen, H. Liu, and C. S. Jensen, “PM-LSH: a fast and accurate in-memory framework for high- dimensional approximate NN and closest pair search,”VLDB J., vol. 31, no. 6, pp. 1339–1363, 2022

    work page 2022

  45. [45]

    Towards efficient index construction and approximate nearest neighbor search in high- dimensional spaces,

    X. Zhao, Y . Tian, K. Huang, B. Zheng, and X. Zhou, “Towards efficient index construction and approximate nearest neighbor search in high- dimensional spaces,”PVLDB, vol. 16, no. 8, pp. 1979–1991, 2023

    work page 1979

  46. [46]

    Ap- proximate nearest neighbor search in high dimensional vector databases: Current research and future directions,

    Y . Tian, Z. Yue, R. Zhang, X. Zhao, B. Zheng, and X. Zhou, “Ap- proximate nearest neighbor search in high dimensional vector databases: Current research and future directions,”IEEE Data Eng. Bull., vol. 47, no. 3, pp. 39–54, 2023

    work page 2023

  47. [47]

    Indexing metric spaces for exact similarity search,

    L. Chen, Y . Gao, X. Song, Z. Li, Y . Zhu, X. Miao, and C. S. Jensen, “Indexing metric spaces for exact similarity search,”ACM Comput. Surv., vol. 55, no. 6, pp. 128:1–128:39, 2023. 13

    work page 2023

  48. [48]

    Optimizing the number of clusters for billion-scale quantization-based nearest neighbor search,

    Y . Fu, C. Chen, X. Chen, W. Wong, and B. He, “Optimizing the number of clusters for billion-scale quantization-based nearest neighbor search,” IEEE Trans. Knowl. Data Eng., vol. 36, no. 11, pp. 6786–6800, 2024

    work page 2024

  49. [49]

    Product quantization for nearest neighbor search,

    H. J ´egou, M. Douze, and C. Schmid, “Product quantization for nearest neighbor search,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 1, pp. 117–128, 2011

    work page 2011

  50. [50]

    An efficient and robust framework for approximate nearest neighbor search with attribute constraint,

    M. Wang, L. Lv, X. Xu, Y . Wang, Q. Yue, and J. Ni, “An efficient and robust framework for approximate nearest neighbor search with attribute constraint,” inNeurIPS, 2023

    work page 2023

  51. [51]

    Rwalks: Random walks as attribute diffusers for filtered vector search,

    A. A. Aomar, K. Echihabi, M. Arnaboldi, I. Alagiannis, D. Hilloulin, and M. Cherkaoui, “Rwalks: Random walks as attribute diffusers for filtered vector search,”Proc. ACM Manag. Data, vol. 3, no. 3, pp. 212:1–212:26, 2025

    work page 2025

  52. [52]

    PASE: postgresql ultra- high-dimensional approximate nearest neighbor search extension,

    W. Yang, T. Li, G. Fang, and H. Wei, “PASE: postgresql ultra- high-dimensional approximate nearest neighbor search extension,” in SIGMOD, 2020, pp. 2241–2253

    work page 2020

  53. [53]

    The metric spaces library maintained by the sisap initiative,

    K. Figueroa, G. Navarro, and E. Chavez, “The metric spaces library maintained by the sisap initiative,” 2017. [Online]. Available: https://github.com/kaarinita/metricSpaces

    work page 2017

  54. [54]

    Hnswlib: fast approximate nearest neighbor search,

    “Hnswlib: fast approximate nearest neighbor search,” 2025. [Online]. Available: https://github.com/nmslib/hnswlib 14 32 64 96 128 160 192 224 0 600 800 1000 1200QPS FDScanning PDScanning PDScanning+ ADSampling DADE DDCres DDCpca /uni0000001b/uni00000014/uni00000019/uni00000016/uni00000015/uni00000019/uni00000017/uni0000001c/uni00000019/uni00000014/uni0000...

    work page 2025