Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

VecFlow: A High-Performance Vector Data Management System for Filtered-Search on GPUs

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VecFlow, a GPU vector search system, claims it can answer filtered approximate nearest neighbor queries at 5 million queries per second at 90% recall by indexing labels rather than only vectors.

desk verdict A solid, well-engineered GPU filtered-ANNS system with a genuinely new index design; the headline SIFT result is credible, but the recall metric needs clarification on sparse-label queries before the long-tail results can be trusted. read the letter →

arxiv 2506.00812 v1 pith:7TCDNWXJ submitted 2025-06-01 cs.DB

Jingyi Xi , Chenghao Mo , Benjamin Karsin , Artem Chirkin , Mingqin Li , Minjia Zhang This is my paper
classification cs.DB
keywords vectordatabasefilteredapproximatenearestneighborsearchGPUaccelerationlabel-centricinvertedindexdual-structuredIVFpersistentkernelmulti-labelpredicateslong-taillabeldistribution
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

VecFlow sets out to show that filtered approximate nearest neighbor search—finding the top-K vectors that are both close to a query and carry required metadata labels—can run at web-serving speeds on a GPU instead of on CPUs. Its central claim is that a label-centric inverted index, which groups vectors by their labels rather than only by spatial proximity, turns filtered search into a series of small unconstrained searches, and that this design reaches 5 million queries per second at 90% recall on a standard benchmark while beating Filtered-DiskANN by up to 135 times. The paper also claims that the same index handles multi-label AND/OR queries and small online batches, making GPU filtered search practical where prior GPU baselines plateau near 80% recall. A sympathetic reader would care because filtered vector search is the query pattern behind hybrid search in production AI systems, and the paper argues it no longer has to be the slow step.

What carries the argument

The load-bearing object is the dual-structured label-centric IVF index. For each label the index stores a posting list of data-point IDs; a specificity threshold T splits these lists into a high-specificity partition, searched with per-label CAGRA-style graphs, and a low-specificity partition, searched with brute-force scans. The high-specificity graphs are stitched into one compacted graph with a local-to-global ID mapping so that all labels share a single vector array rather than duplicating vectors; the low-specificity lists use an interleaved vector layout so warps issue coalesced 128-bit loads. A persistent kernel with atomic ring buffers carries small batches without repeated kernel launches, and a predicate function over per-point sorted label arrays verifies AND conditions during graph traversal and before brute-force distance computation.

What would settle it

Recompute the reported QPS-vs-recall curves using only queries whose filtered ground-truth set has at least K = 10 points, and separately report the fraction of queries excluded for having fewer matches. If VecFlow's 5M QPS at 90% recall on SIFT-1M or 2.6M on YFCC-10M depends on including queries with fewer than 10 valid answers, the comparison against Filtered-DiskANN and the GPU baselines is not apples-to-apples.

Watch

Extended reading notes

Core claim

On its own terms, VecFlow's discovery is that the long-tailed label distributions of real filtered-search workloads make a dual index more efficient than any single graph. Labels are classified by specificity, the fraction of the dataset carrying the label. High-specificity labels are searched through per-label GPU graphs that are compacted into one continuous graph and share a single global copy of the vectors, using a local-to-global mapping to eliminate replicated storage; low-specificity labels are searched by brute force over an interleaved memory layout that maximizes coalesced bandwidth. The paper reports 5M QPS at 90% recall on SIFT-1M, 2.6M QPS on YFCC-10M, about 3M QPS on DEEP-50M, and 150K QPS on WIKI-1M with AND predicates, and states that VecFlow can reach above 99% recall where CAGRA-based GPU baselines stop near 80%.

Load-bearing premise

The load-bearing premise is that the reported recall of 90% is well-defined for every query: the recall formula divides by K = 10, which assumes each query's filtered ground truth contains at least 10 matching points, yet the paper never states how queries with fewer than 10 filtered matches are counted or excluded on datasets like YFCC-10M with 200K labels and WIKI-1M AND queries.

Editorial extensions

If this is right

  • Filtered approximate nearest neighbor search reaches million-query-per-second throughput on a single GPU for datasets up to tens of millions of vectors, a regime previously reserved for unconstrained search.
  • Multi-label AND queries can be answered by searching the rarest label's posting list and verifying the remaining labels, so cost scales with the smallest matching class rather than the whole dataset.
  • Small online batches become GPU-efficient because the persistent kernel removes per-batch launch overhead, which matters for interactive serving workloads.
  • The redundancy-bypassing layout keeps index memory close to that of a single unfiltered graph even when points carry many labels, removing the main obstacle to label-centric IVF on memory-limited GPUs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not explore it, but if these throughput numbers generalize, filtered search could be fused into retrieval pipelines as a single GPU operator, replacing the common pre-filter-then-search cascade and removing the need to guess candidate list sizes.
  • Not claimed by the paper: the specificity threshold T is effectively a tunable crossover point between graph traversal and brute force, and the same profiling-driven split might extend to other hardware or to dynamic label distributions.
  • A natural stress test the paper does not run is K larger than 10 and streams of label updates, since the recall metric and static indexes are defined for fixed K and offline construction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. VecFlow is a GPU-resident vector data management system for filtered approximate nearest neighbor search. The paper proposes a label-centric inverted-file index that separates labels into high-specificity and low-specificity groups, using a CAGRA-style graph index for the former and a brute-force scan with an interleaved memory layout for the latter. It adds a redundancy-bypassing layout to avoid replicating vectors across label-specific graphs, a persistent-kernel mechanism for small query batches, and specialized handling of multi-label OR/AND predicates. Experiments on SIFT-1M, YFCC-10M, WIKI-1M, and DEEP-50M report up to 5 million QPS at 90% recall on SIFT-1M, with claims of up to 135x speedup over Filtered-DiskANN and better recall than GPU-based CAGRA baselines.

Significance. If the results hold, VecFlow is an important contribution: it is one of the first systems to make filtered ANNS practical at million-query-per-second scale on GPUs, and it directly addresses a workload that CPU-oriented systems such as Filtered-DiskANN handle poorly. The paper is unusually complete in its system-oriented contributions: an open-source implementation is promised, the index layout and kernels are described in sufficient detail to be reproducible, and the evaluation covers multiple datasets, multiple GPUs, small-batch latency, memory footprint, and index construction time. The central algorithmic idea of separating high- and low-specificity labels and choosing graph search versus brute-force scan on that basis is well motivated and supported by the ablation in Fig. 4 and Fig. 15. However, the recall metric used for the headline QPS claims is not well-defined for queries whose filtered ground truth contains fewer than K points, and this ambiguity affects the YFCC and WIKI results that are needed to demonstrate the long-tail regime the paper motivates.

major comments (2)
  1. [§3, Eq. (1); §5.2] Eq. (1) defines recall as |A_topK ∩ GT_topK| / K, which assumes every query's filtered ground-truth set has exactly K elements. The manuscript never states how queries whose label set has fewer than K matching points are handled, although such queries are inevitable in the evaluated workloads: YFCC-10M has 200K labels with mean posting-list size 540 (§4.2.4), and §5.2 explicitly says that WIKI-1M two-label intersections are 'extremely small' and that the CAGRA baselines return 'nearly 0 recall' for them. If such queries are included in the average, the maximum achievable recall is |GT_topK|/K < 1, so a reported 90% recall is mathematically impossible unless the evaluation drops or pads these queries; if it drops or pads them, the query workload changes and the QPS comparison against Filtered-DiskANN, IVF2, FAISS, and CAGRA is not apples-to-apples. The paper must specify the exact policy for degenerate queries (drop, pad, or cap the denominator at |GT_topK|), apply the same policy to all baselines, and report the fraction of affected queries per dataset.
  2. [§5.2, Fig. 15] The headline QPS figures are obtained after selecting the specificity threshold T (e.g., T=2000 in §5.2) and the CAGRA itopk per dataset, but the manuscript does not report the tuning procedure, the grid searched, or any measure of run-to-run variance. Fig. 15 shows that QPS on YFCC varies substantially with T, so the reader cannot tell whether the reported 5M QPS at 90% recall is a best-case value selected after tuning or a representative operating point. Please report the tuning protocol (e.g., grid ranges, selection criterion) and state whether the reported numbers are best-of-N, median, or mean over multiple runs, ideally with error bars on the QPS-vs-recall curves.
minor comments (6)
  1. [§4.2.4] The memory analysis uses 'GB' for what appear to be binary gibibytes (GiB), which makes the example numbers confusing (76.8 GB appears as 71.53). Please use consistent units throughout.
  2. [§5.2, Fig. 9] The curves for CAGRA-Post and CAGRA-Inline are difficult to distinguish in grayscale; consider using distinct markers or line styles to improve readability.
  3. [§5.1, §5.2] The paper states that experiments run on A100 and GH200, and Fig. 17 reports GH200, but it is not stated explicitly that all other figures use the A100 40GB GPU. Please state this in the testbed description.
  4. [§5.2] The WIKI-1M paragraph says FAISS 'achieves high recall by performing BFS for almost every query' while the CAGRA methods 'achieve nearly 0 recall'; providing the actual recall values or curve endpoints would make the comparison more concrete.
  5. [References] Reference [8] lacks a publication year and venue; please complete the citation.
  6. [§4.2.3] The phrase 'avoid synchronization at all costs' should be 'avoid synchronization overhead' to be precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: VecFlow's QPS and recall are measured against brute-force ground truth and external baselines; tunable parameters are standard hyperparameter selection, not fitted predictions.

full rationale

I found no circular step in VecFlow's derivation chain. The headline QPS@90% recall figures are empirical results computed against exact brute-force ground truth (Eq. 1) and compared with independent baselines (Filtered-DiskANN, FAISS, IVF2, CAGRA variants); recall is not derived from a fitted parameter. The specificity threshold T of Eq. 2 is an auto-tuned routing hyperparameter (Sec. 4.1.2, 'determined through offline profiling ... and auto-tuning'), and Fig. 15 shows performance as T varies; tuning T to maximize throughput is standard hyperparameter selection, not a construction that forces the reported recall. CAGRA and cuVS are used as both building blocks and baselines, but they are external, code-reproducible implementations, and the NVIDIA affiliation of some co-authors does not make the evaluation circular. The one flagged concern is a correctness/comparability issue rather than circularity: Eq. 1 divides by K and assumes every query has exactly K filtered ground-truth neighbors, but the paper never states how YFCC/WIKI queries with fewer than K matches are handled, which can affect whether the reported 90% recall is well-defined; this does not reduce the system's claimed performance to its inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The system's design relies on the label distribution being long-tailed, on the threshold T separating regimes where graphs beat brute force, and on the recall metric being well-defined. None of these are derived; they are empirical or assumed. The main free parameters are the routing threshold T and graph degree R'.

free parameters (3)
  • Specificity threshold T = 2000
    Controls routing between IVF-Graph (HS) and IVF-BFS (LS). Chosen by offline auto-tuning to maximize QPS; varies per dataset (Fig. 15).
  • CAGRA itopk (search width) = varying (e.g., 32 default)
    Graph search beam width that trades recall for throughput; swept in Fig. 9 to produce QPS-recall curves.
  • Graph out-degree R' = 16
    Chosen empirically for HS graphs; paper states 'shows no performance degradation based on empirical observation' (Section 4.2.4).
assumptions (3)
  • domain assumption Label distributions in real filtered-ANNS workloads are long-tailed (few frequent, many rare labels).
    Motivates splitting labels into high/low specificity; supported by Fig. 3 on SIFT/YFCC but assumed to generalize.
  • domain assumption A graph index's traversal overhead (CAGRA) makes it less efficient than brute-force for posting lists below size T.
    Justifies IVF-BFS for low-specificity lists; based on scaling experiments in Fig. 4, not a theorem.
  • ad hoc to paper The filtered ground-truth set for every evaluated query contains at least K=10 points.
    Recall is defined with denominator K (Eq. 1); the paper does not report how queries with fewer than K matching points are handled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VecFlow: A High-Performance Vector Data Management System for Filtered-Search on GPUs." pith.science (2026). https://pith.science/paper/7TCDNWXJ

@misc{pith2026250600812,
  author       = {Pith},
  title        = {Pith review of: VecFlow: A High-Performance Vector Data Management System for Filtered-Search on GPUs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7TCDNWXJ}},
  note         = {Machine review of arXiv:2506.00812}
}
read the original abstract

Vector search and database systems have become a keystone component in many AI applications. While many prior research has investigated how to accelerate the performance of generic vector search, emerging AI applications require running more sophisticated vector queries efficiently, such as vector search with attribute filters. Unfortunately, recent filtered-ANNS solutions are primarily designed for CPUs, with few exploration and limited performance of filtered-ANNS that take advantage of the massive parallelism offered by GPUs. In this paper, we present VecFlow, a novel high-performance vector filtered search system that achieves unprecedented high throughput and recall while obtaining low latency for filtered-ANNS on GPUs. We propose a novel label-centric indexing and search algorithm that significantly improves the selectivity of ANNS with filters. In addition to algorithmic level optimization, we provide architectural-aware optimization for VecFlow's functional modules, effectively supporting both small batch and large batch queries, and single-label and multi-label query processing. Experimental results on NVIDIA A100 GPU over several public available datasets validate that VecFlow achieves 5 million QPS for recall 90%, outperforming state-of-the-art CPU-based solutions such as Filtered-DiskANN by up to 135 times. Alternatively, VecFlow can easily extend its support to high recall 99% regime, whereas strong GPU-based baselines plateau at around 80% recall. The source code is available at https://github.com/Supercomputing-System-AI-Lab/VecFlow.

Figures

Figures reproduced from arXiv: 2506.00812 by the authors.

Figure 1
Figure 1. System overview of VecFlow. The architecture [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Three common approaches to enable filtered-ANNS. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Label specificity distribution in different datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: (a) Scaling efficiency of IVF-Graph and IVF-BFS [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The data layout used in VecFlow ’s redundancy [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Interleaved memory layout for VecFlow’s IVF-BFS [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: shows our persistent kernel design, which uses a job queue system with atomic ring buffers to manage incoming queries and a worker queue to track available GPU thread blocks. When a query arrives, it is assigned a job ID and mapped to an available worker. The persisten…
Figure 8
Figure 8. Figure 8: Multi-label query processing in VecFlow. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: The QPS vs. recall comparison results on (a) SIFT, (b) YFCC, (c) WIKI and (d) DEEP respectively. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 11
Figure 11. Figure 11: Comparison results between VecFlow with Redundancy-bypassing and VecFlow without Redundancy￾bypassing [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 10
Figure 10. Figure 10: Memory efficiency of redundancy-bypassing. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 13
Figure 13. Figure 13: The comparison of multi-label search policy in [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 17
Figure 17. Figure 17: Performance of VecFlow on NVidia GH200 GPU. How does VecFlow perform on real-world production workloads? We evaluated VecFlow on a real-world recommendation dataset containing 4 million vectors (D=64) with over 10,000 labels show￾ing a long-tailed distribution. Using …
Figure 14
Figure 14. Figure 14: Persistent kernel results on SIFT and YFCC [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 18
Figure 18. Figure 18: Index construction time comparison. 6 Conclusion Modern AI-based applications require ANNS with filtered search. However, indexing and search algorithms with filters on GPUs must be re-designed to achieve high performance potential. We present a new GPU-based indexing…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ANNLib: A Development Framework for Efficient Approximate Nearest Neighbor Search

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A modular ANNS framework decouples search algorithms from graph storage, delivering filtered search, dynamic updates, and snapshot queries at performance close to that of specialized systems.

Reference graph

Works this paper leans on

76 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    pgvector: Open-source vector similarity search for Postgres

    Accessed: 04-13-2025. pgvector: Open-source vector similarity search for Postgres. https://github.com/pgvector/pgvector

    work page 2025

  2. [2]

    RAPIDS AI. 2025. cuVS. https://github.com/rapidsai/cuvs. Accessed: 2025-01-18

    work page 2025

  3. [3]

    Bruce Croft

    Mohammad Aliannejadi, Hamed Zamani, Fabio Crestani, and W. Bruce Croft

  4. [4]

    Alexandr Andoni, Piotr Indyk, and Ilya Razenshteyn. 2018. Approximate Nearest Neighbor Search in High Dimensions. arXiv preprint arXiv:1806.09823 (2018)

    arXiv 2018

  5. [5]

    Norbert Beckmann, Hans-Peter Kriegel, Ralf Schneider, and Bernhard Seeger

  6. [6]

    Ben Landrum and Magdalen Dobson Manohar and Mazin Karjikar and Lax- man Dhulipala . 2024. IVF2: Fusing Classic and Spatial Inverted Indices for Fast Filtered ANNS. https://big-ann-benchmarks .com/neurips23_slides/ IVF_2_filter_Ben.pdf

    work page 2024

  7. [7]

    Jon Louis Bentley. 1975. Multidimensional Binary Search Trees Used for Associa- tive Searching. Commun. ACM 18, 9 (Sept. 1975), 509–517

    work page 1975

  8. [8]

    Philip A Bernstein, Siddharth Gollapudi, Suryansh Gupta, Ravishankar Krish- naswamy, Sepideh Mahabadi, Sandeep Silwal, Gopal R Srinivasa, Varun Suriya- narayana, Jakub Tarnawski, Haiyang Xu, et al. [n. d.]. Graph-based algorithms for nearest neighbor search with multiple filters. ([n. d.])

    work page

Show all 76 references
  1. [9]

    Big-ANN. [n. d.]. NeurIPS’23 Competition Track: Big-ANN. https://big-ann- benchmarks.com/neurips23.html. Accessed: 2024

    2024

  2. [10]

    Cheng Chen, Chenzhe Jin, Yunan Zhang, Sasha Podolsky, Chun Wu, Szu-Po Wang, Eric Hanson, Zhou Sun, Robert Walzer, and Jianguo Wang. 2024. SingleStore-V: An Integrated Vector Database System in SingleStore. Proc. VLDB Endow. 17, 12 (Aug. 2024), 3772–3785. https://doi .org/10.14...

    2024

  3. [11]

    Qi Chen, Haidong Wang, Mingqin Li, Gang Ren, Scarlett Li, Jeffery Zhu, Jason Li, Chuanjie Liu, Lintao Zhang, and Jingdong Wang. 2018. SPTAG: A library for fast approximate nearest neighbor search . https://github .com/Microsoft/SPTAG

    2018

  4. [12]

    Wei Chen, Jincai Chen, Fuhao Zou, Yuan-Fang Li, Ping Lu, Qiang Wang, and Wei Zhao. 2019. Vector and line quantization for billion-scale similarity search on GPUs. Future Gener. Comput. Syst. 99 (2019), 295–307

    2019

  5. [13]

    Muthukrishnan

    Graham Cormode and S. Muthukrishnan. 2005. An improved data stream sum- mary: the count-min sketch and its applications. Journal of Algorithms 55, 1 (2005), 58–75. https://doi .org/10.1016/j.jalgor.2003.12.001

    2005 doi

  6. [14]

    Bruce Croft

    Mostafa Dehghani, Hamed Zamani, Aliaksei Severyn, Jaap Kamps, and W. Bruce Croft. 2017. Neural Ranking Models with Weak Supervision. InSIGIR 2017. 65–74

    2017

  7. [15]

    Wei Dong, Charikar Moses, and Kai Li. 2011. Efficient k-nearest neighbor graph construction for generic similarity measures. In Proceedings of the 20th International Conference on World Wide Web (Hyderabad, India) (WWW ’11) . Association for Computing Machinery, New York, NY, ...

    2011

  8. [16]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library.arXiv preprint arXiv:2401.08281 (2024). https://doi.org/10.48550/ arXiv.2401.08281

    2024 doi arXiv

  9. [17]

    Jian Fang, Yvo TB Mulder, Jan Hidders, Jinho Lee, and H Peter Hofstee. 2020. In-memory database acceleration on FPGAs: a survey. The VLDB Journal 29 (2020), 33–59

    2020

  10. [18]

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. 2019. Fast Approximate Nearest Neighbor Search with the Navigating Spreading-out Graph. In VLDB’19

    2019

  11. [19]

    Gibbons, Yossi Matias, and Avi Silberschatz

    Sumit Ganguly, Phillip B. Gibbons, Yossi Matias, and Avi Silberschatz. 1996. Bifocal sampling for skew-resistant join size estimation. SIGMOD Rec. 25, 2 (June 1996), 271–281. https://doi .org/10.1145/235968.233340

    1996

  12. [20]

    Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized Product Quantization for Approximate Nearest Neighbor Search. In CVPR 2013

    2013

  13. [21]

    Aristides Gionis, Piotr Indyk, and Rajeev Motwani. 1999. Similarity Search in High Dimensions via Hashing. In VLDB’99. 518–529

    1999

  14. [22]

    Siddharth Gollapudi, Neel Karia, Varun Sivashankar, Ravishankar Krishnaswamy, Nikit Begwani, Swapnil Raz, Yiyong Lin, Yin Zhang, Neelam Mahapatro, Premku- mar Srinivasan, et al. 2023. Filtered-diskann: Graph algorithms for approximate nearest neighbor search with filters. In P...

    2023

  15. [23]

    Google. 2022. Go beyond the search box: Introducing multisearch. https:// blog.google/products/search/multisearch/. Accessed: 2025

    2022

  16. [24]

    Fabian Groh, Lukas Ruppert, Patrick Wieschollek, and Hendrik P. A. Lensch

  17. [25]

    Bruce Croft

    Jiafeng Guo, Yixing Fan, Qingyao Ai, and W. Bruce Croft. 2016. A Deep Relevance Matching Model for Ad-hoc Retrieval. In CIKM 2016. 55–64

    2016

  18. [26]

    Neha Gupta. 2021. Introduction to hardware accelerator systems for artificial intelligence and machine learning. In Advances in Computers. Vol. 122. Elsevier, 1–21

    2021

  19. [27]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In CIKM ’13. 2333–2338

    2013

  20. [28]

    HuggingFace. 2025. WikiANN dataset. https://huggingface.co/2024annonymous/ wiki-ann

    2025

  21. [29]

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnawamy, and Rohan Kadekodi. 2019. Diskann: Fast accurate billion-point nearest neighbor search on a single node. Advances in Neural Information Pro- cessing Systems 32 (2019)

    2019

  22. [30]

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2011. In Product Quantization for Nearest Neighbor Search. TPAMI 2011

    2011

  23. [31]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2017. Billion-scale similarity search with GPUs. CoRR abs/1702.08734 (2017). arXiv: http://arxiv .org/abs/ 1702.08734

    2017 arXiv

  24. [32]

    Avrithis

    Yannis Kalantidis and Yannis S. Avrithis. 2014. Locally Optimized Product Quan- tization for Approximate Nearest Neighbor Search. In CVPR 2014. 2329–2336

    2014

  25. [33]

    Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. 2020. Learning and Evaluating Contextual Embedding of Source Code. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event (Proceedings of Machine Lea...

    2020

  26. [34]

    Victor Lempitsky. 2012. The Inverted Multi-index. In CVPR ’12. 3069–3076

    2012

  27. [35]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. InProceedings o...

    2020

  28. [36]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022. BLIP: Bootstrap- ping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In International Conference on Machine Learning, ICML 2022, 17- 23 July 2022, Baltimore, Maryland, USA (...

    2022

  29. [37]

    Wen Li, Ying Zhang, Yifang Sun, Wei Wang, Mingjie Li, Wenjie Zhang, and Xuemin Lin. 2020. Approximate Nearest Neighbor Search on High Dimensional Data – Experiments, Analyses, and Improvement. IEEE Transactions on Knowl- edge and Data Engineering 32, 8 (2020), 1475–1488. https...

    2020

  30. [38]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR abs/1907.11692 (2019)

    2019 arXiv

  31. [39]

    LongChain. [n. d.]. LongChain: Build context-aware reasoning applications. https://github.com/langchain-ai/langchain. Accessed: 2025

    2025

  32. [40]

    Malkov and D

    Yury A. Malkov and D. A. Yashunin. 2016. Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs. CoRR arXiv preprint abs/1603.09320 (2016)

    2016 arXiv

  33. [41]

    Corrado, and Jeffrey Dean

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean

  34. [42]

    Milvus-io. 2022. Milvus-docs: Conduct a hybrid search. https://github .com/ milvus-io/milvus-docs/blob/v2.1.x/site/en/userGuide/search/hybridsearch.md. Accessed: 2025

    2022

  35. [43]

    Bhaskar Mitra, Fernando Diaz, and Nick Craswell. 2017. Learning to Match using Local and Distributed Representations of Text for Web Search. InWWW 2017

    2017

  36. [44]

    Marius Muja and David G. Lowe. 2014. Scalable Nearest Neighbor Algorithms for High Dimensional Data. TPAMI 2014 36, 11 (2014), 2227–2240

    2014

  37. [45]

    Mohammad Norouzi and David J. Fleet. 2013. Cartesian K-Means. In CVPR 2013

    2013

  38. [46]

    Nolet, Ray Wang, Tamas B

    Hiroyuki Ootomo, Akira Naruse, Corey J. Nolet, Ray Wang, Tamas B. Fehér, and Y. Wang. 2023. CAGRA: Highly Parallel Graph Construction and Approximate Nearest Neighbor Search for GPUs. 2024 IEEE 40th International Conference on Data Engineering (ICDE) (2023), 4236–4247

    2023

  39. [47]

    James Jie Pan, Jianguo Wang, and Guoliang Li. 2024. Survey of vector database management systems. VLDB J. 33, 5 (2024), 1591–1615

    2024

  40. [48]

    Liana Patel, Peter Kraft, Carlos Guestrin, and Matei Zaharia. 2024. ACORN: Per- formant and Predicate-Agnostic Search Over Vector Embeddings and Structured Data. Proc. ACM Manag. Data 2, 3 (2024), 120

    2024

  41. [49]

    Pinecone Systems

    Inc. Pinecone Systems. 2024. Overview. https://docs.pinecone.io/docs/overview. Accessed: 2025

    2024

  42. [50]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings...

    2021

  43. [51]

    Jie Ren, Minjia Zhang, and Dong Li. 2020. HM-ANN: Efficient Billion-Point Nearest Neighbor Search on Heterogeneous Memory. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 202...

    2020

  44. [52]

    Harshit Sharma and Anmol Sharma. 2024. A Comprehensive Overview of GPU Accelerated Databases. arXiv preprint arXiv:2406.13831 (2024)

    2024 arXiv

  45. [53]

    Harsha Simhadri. 2025. Big ANN Benchmarks. https://github .com/harsha- simhadri/big-ann-benchmarks. Accessed: 2025-01-18

    2025

  46. [54]

    Harsha Vardhan Simhadri, Martin Aumüller, Amir Ingber, Matthijs Douze, George Williams, Magdalen Dobson Manohar, Dmitry Baranchuk, Edo Liberty, Frank Liu, Ben Landrum, et al. 2024. Results of the Big ANN: NeurIPS’23 competition. arXiv preprint arXiv:2409.17424 (2024)

    2024 arXiv

  47. [55]

    Aditi Singh, Suhas Jayaram Subramanya, Ravishankar Krishnaswamy, and Har- sha Vardhan Simhadri. 2021. FreshDiskANN: A Fast and Accurate Graph-Based ANN Index for Streaming Similarity Search. CoRR abs/2105.09613 (2021)

    2021 arXiv

  48. [56]

    Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li

    Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. 2016. YFCC100M: the new data in multimedia research. Commun. ACM 59, 2 (2016), 64–73

    2016

  49. [57]

    Karthik V., Saim Khan, Somesh Singh, Harsha Vardhan Simhadri, and Jyothi Vedurada. 2024. BANG: Billion-Scale Approximate Nearest Neighbor Search using a Single GPU. arXiv: [cs.DC] https://arxiv .org/abs/2401.11324

    2024 arXiv

  50. [58]

    Christophe Van Gysel, Maarten de Rijke, and Evangelos Kanoulas. 2016. Learning Latent Vector Spaces for Product Search. In CIKM ’16. 165–174

    2016

  51. [59]

    Mengzhao Wang, Lingwei Lv, Xiaoliang Xu, Yuxiang Wang, Qiang Yue, and Jiongkang Ni. 2023. An Efficient and Robust Framework for Approximate Nearest Neighbor Search with Attribute Constraint. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural I...

    2023

  52. [60]

    Mengzhao Wang, Xiaoliang Xu, Qiang Yue, and Yuxiang Wang. 2021. A com- prehensive survey and experimental comparison of graph-based approximate nearest neighbor search. arXiv preprint arXiv:2101.12631 (2021)

    2021 arXiv

  53. [61]

    Weaviate. 2022. Weaviate Documentation: Filters. https://weaviate.io/developers/ weaviate/current/graphql-references/filters.html. Accessed: 2025

    2022

  54. [62]

    Chuangxian Wei, Bin Wu, Sheng Wang, Renjie Lou, Chaoqun Zhan, Feifei Li, and Yuanzhe Cai. 2020. AnalyticDB-V: a hybrid analytical engine towards query fusion for structured and unstructured data.Proc. VLDB Endow. 13, 12 (Aug. 2020), 3152–3165. https://doi .org/10.14778/3415478.3415541

    2020

  55. [63]

    Wen Yang, Tao Li, Gai Fang, and Hong Wei. 2020. PASE: PostgreSQL Ultra-High- Dimensional Approximate Nearest Neighbor Search Extension. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data (Portland, OR, USA) (SIGMOD ’20). Association for Comput...

    2020

  56. [64]

    Lei Yu, Karl Moritz Hermann, Phil Blunsom, and Stephen Pulman. 2014. Deep Learning for Answer Sentence Selection. CoRR abs/1412.1632 (2014)

    2014 arXiv

  57. [65]

    Yuanhang Yu, Dong Wen, Ying Zhang, Lu Qin, Wenjie Zhang, and Xuemin Lin

  58. [66]

    Hamed Zamani, Bhaskar Mitra, Xia Song, Nick Craswell, and Saurabh Tiwary

  59. [67]

    Qianxi Zhang, Shuotao Xu, Qi Chen, Guoxin Sui, Jiadong Xie, Zhizhen Cai, Yaoqi Chen, Yinxuan He, Yuqing Yang, Fan Yang, Mao Yang, and Lidong Zhou

  60. [68]

    Zili Zhang, Fangyue Liu, Gang Huang, Xuanzhe Liu, and Xin Jin. 2024. Fast Vector Query Processing for Large Datasets Beyond{GPU} Memory with Re- ordered Pipelining. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). 23–40

    2024

  61. [69]

    Weijie Zhao, Shulong Tan, and Ping Li. 2020. SONG: Approximate Nearest Neigh- bor Search on GPU. In 36th IEEE International Conference on Data Engineering, ICDE 2020, Dallas, TX, USA, April 20-24, 2020 . IEEE, 1033–1044

    2020

  62. [72]

    In WSDM ’18

    Neural Ranking Models with Multiple Document Fields. In WSDM ’18

  63. [74]

    In 17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23)

    VBASE: Unifying Online Vector Similarity Search and Relational Queries via Relaxed Monotonicity. In 17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23). USENIX Association, Boston, MA, 377–395. https://www.usenix.org/conference/osdi23/presentation/zh...

  64. [1990]

    In SIGMOD 1990

    The R*-Tree: An Efficient and Robust Access Method for Points and Rect- angles. In SIGMOD 1990. 322–331

    1990

  65. [2013]

    In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013

    Distributed Representations of Words and Phrases and their Composi- tionality. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013 . 3111–3119

    2013

  66. [2018]

    In SIGIR 2018

    Target Apps Selection: Towards a Unified Search Framework for Mobile Devices. In SIGIR 2018. 215–224

    2018

  67. [2022]

    In 38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022

    GPU-accelerated Proximity Graph Approximate Nearest Neighbor Search and Construction. In 38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022 . IEEE, 552–564

    2022

  68. [2023]

    IEEE Transactions on Big Data 9, 1 (2023), 267–279

    GGNN: Graph-Based GPU Nearest Neighbor Search. IEEE Transactions on Big Data 9, 1 (2023), 267–279. https://doi .org/10.1109/TBDATA.2022.3161156

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.