Pith. sign in

REVIEW 6 major objections 4 minor 1 cited by

Quake: Adaptive Indexing for Vector Search

T0 review · 6 major / 4 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 Quake keeps vector search fast and at target recall on shifting, skewed workloads by reshaping partitions with a cost model, estimating recall online, and using NUMA-aware parallelism.

desk verdict Real new machinery in an adaptive vector index, but the headline speedups depend on comparing 16-thread Quake to single-thread baselines. read the letter →

arxiv 2506.03437 v2 pith:ZZVGYJ3X submitted 2025-06-03 cs.IR

Jason Mohoney , Devesh Sarda , Mengze Tang , Shihabur Rahman Chowdhury , Anil Pacaci , Ihab F. Ilyas , Theodoros Rekatsinas , Shivaram Venkataraman This is my paper
classification cs.IR
keywords vectorsearchapproximatenearestneighboradaptiveindexingdynamicworkloadsincrementalindexmaintenancerecallestimationNUMA-awareparallelismpartitioned
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

The paper tries to establish that vector search can stay fast and accurate when the data and the queries shift over time, the situation real applications such as retrieval-augmented generation and recommendation systems face. Its system, Quake, is a partitioned index that watches which partitions are hot or large and splits or merges them according to a cost model, so the index reshapes itself as the workload changes. Instead of using a fixed number of partitions per query, Quake estimates online when enough partitions have been scanned to hit a recall target, using a geometric model of where nearest neighbors are likely to lie. Because scan throughput, not CPU, is the bottleneck, Quake also places partitions across NUMA memory nodes and schedules worker threads next to their data. If these claims hold, dynamic vector search can run without per-workload tuning of how many partitions each query scans or repeated recalibration as data drifts.

What carries the argument

Two mechanisms carry the argument. The first is the per-partition cost model $C_{l,j}=A_{l,j}\lambda(s_{l,j})$, where $A_{l,j}$ is the fraction of recent queries that scan partition $(l,j)$ and $\lambda(s)$ is the measured latency of scanning $s$ vectors; the total $C=\sum_{l,j} C_{l,j}$ identifies which partitions hurt latency most. Maintenance actions—split, merge, add level, remove level—are proposed when the estimated change $\Delta' < -\tau$, then re-evaluated on actual sizes and committed only if the measured $\Delta$ still improves cost, which guarantees monotone decrease and convergence to a local minimum. The second is Adaptive Partition Scanning (APS), whose geometric model sets $p_i = \mathrm{Vol}(B(q,\rho)\cap P_i)/\mathrm{Vol}(B(q,\rho))$, the volume fraction of the query hypersphere $B(q,\rho)$ of radius equal to the current $k$-th neighbor distance that intersects partition $P_i$; intersection volumes are approximated by hyperspherical caps, the cap-shaped slices of the sphere cut by the bisecting planes between centroids, and partitions are scanned in descending $p_i$ until the estimated recall reaches target $\tau_R$. NUMA-aware partition placement and affinity scheduling, with Non-Uniform Memory Access (NUMA) nodes holding partitions that local worker threads scan, are what let the system actually spend the available memory bandwidth during those scans.

What would settle it

On a synthetic dataset whose nearest neighbors are concentrated inside one small partition, run a query at a 90% recall target and compare APS's estimated recall against known ground truth; if the estimate crosses 90% before that partition is scanned while true recall is below 90%, the uniform-density volume assumption is falsified.

Watch

Extended reading notes

Core claim

Quake claims that a partitioned vector index can hold low query latency and meet a fixed recall target on dynamic, skewed workloads, provided the index structure is driven by workload feedback rather than by static parameters. The system keeps a multi-level inverted-file index, a hierarchy of centroid directories over vector partitions, and after each batch of operations evaluates a cost model that attributes total query latency to individual partitions through their sizes and access frequencies; partitions whose predicted gain passes an estimate-then-verify test are split or merged, and the hierarchy is shown to converge to a local minimum of the estimated cost. For each query, Adaptive Partition Scanning (APS) estimates when the recall target has been reached from the geometry of partition boundaries and the intermediate top-$k$ results, then stops scanning, so the number of partitions probed adapts as data and index change. On the paper's workloads, a Wikipedia-derived trace with skewed page-view queries and monthly inserts, an Open Images stream with insert/delete churn, and million- and hundred-million-scale subsets of a standard benchmark, Quake reports 1.5–38× lower query latency and 4.5–126× lower update latency than the published baselines, with APS matching an oracle's per-query partition count within 17–29% added latency and no offline tuning.

Load-bearing premise

The load-bearing assumption is that vectors inside the query's hypersphere are uniformly distributed across partitions, so that volume fractions translate directly into recall probabilities; real embeddings cluster, and when they do, the estimator can terminate a query too early or scan too much.

Editorial extensions

If this is right

  • On a Wikipedia-derived workload with read and write skew, Quake's multi-threaded search completes in about 1.5 hours versus 12 hours for the strongest graph baseline, while recall stays near the 90% target.
  • APS matches the number of partitions an oracle would scan across 80%, 90%, and 99% recall targets on a million-vector benchmark, with only 17–29% added latency and zero offline tuning.
  • Because every split or merge is committed only when the measured cost change is negative, the index's total estimated query cost decreases monotonically and converges to a local minimum under a fixed workload distribution.
  • NUMA-aware intra-query parallelism gives near-linear speedup and roughly 200 GB/s of scan throughput on a 100 million-vector benchmark, reducing query latency about 20x versus single-threaded scanning.
  • On streams with both insertions and deletions, Quake keeps update-plus-maintenance time orders of magnitude below graph-based indexes, which pay large costs to rewire or consolidate the graph.

Reading between the lines

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

  • Editorial inference: profiling scan latency per storage device and feeding it into the same cost model would let the adaptive split/merge policy place hot partitions on fast memory and cold partitions on slow storage, extending the system to heterogeneous hardware.
  • Editorial inference: replacing the uniform-density volume fraction with a locally learned density estimate, for example from per-partition sample statistics, would likely make the recall target hold on strongly clustered embedding distributions at the cost of periodic statistics updates.
  • Editorial inference: the initial candidate fraction is the main remaining user-set knob; a rule that sets it from the index fan-out and the observed number of partitions actually scanned would remove the last manual search parameter.
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

6 major / 4 minor

Summary. The paper introduces Quake, a partitioned vector-search index that targets dynamic and skewed workloads. Quake has three main mechanisms: a cost-model-driven maintenance procedure that splits and merges partitions based on access frequency and size, an online adaptive partition scanning scheme (APS) that estimates per-query recall from hypersphere/partition volume ratios and terminates once the recall target is reached, and NUMA-aware intra-query parallelism. The evaluation introduces a Wikipedia-derived workload and a configurable workload generator, and compares Quake with Faiss-IVF, DeDrift, LIRE, ScaNN, Faiss-HNSW, DiskANN, and SVS on dynamic and static benchmarks, reporting large query- and update-latency reductions. The paper also includes ablations, a comparison to early-termination methods, and a multi-level recall estimation study.

Significance. If the central claims hold under resource-matched evaluation, Quake would be a useful contribution to dynamic approximate nearest neighbor search: the open-source artifact, the workload generator, and the oracle-nprobe comparison are valuable, and the cost-model maintenance idea is clearly motivated by real skew patterns. The paper is also careful to ablate individual components and to report recall stability. However, the headline results are weakened by asymmetric thread counts in the main table, by estimated or non-conforming baseline runs, and by an unverified uniform-density assumption in APS. These issues affect the strength of the central claim and need correction before the results can be taken at face value.

major comments (6)
  1. [§7.2, Table 3, Table 4] The headline search-latency comparison is not resource-matched. Section 7.2 states that all search numbers use a single worker thread 'unless otherwise stated' and that Quake-MT uses 16 threads, so Table 3's baseline search columns are single-threaded while Quake-MT is 16-threaded. Table 4 shows that this thread asymmetry alone accounts for a 6.2x latency drop on Wikipedia-12M (3.28 ms Quake-ST vs. 0.53 ms Quake-MT). Table 3 also shows Quake-ST is slower than SVS on MSTuring-10M-RO (2.43 h vs. 0.33 h) and slower than DiskANN on MSTuring-10M-IH (2.12 h vs. 0.81 h), so the claim in the abstract and Section 7.3 that Quake 'achieves the lowest search time across all dynamic workloads' is only true with asymmetric thread allocation. Please add 16-thread search results for all baselines (or otherwise match resources) and state the headline speedups separately for Quake-MT and Quake-ST.
  2. [Table 3 footnotes] Several Table 3 entries are estimated rather than measured. The footnote states that methods marked '+' did not finish within the 24-hour budget and their runtime was estimated from a 10% subsample of search queries, and methods marked '*' did not meet the recall target. Extrapolating a 165.8 h Faiss-IVF search time from a 10% subsample assumes constant throughput over the full workload, which is exactly what Section 2.3 argues is false for unmaintained indexes; and comparing total time for configurations that failed the recall target mixes accuracy and latency. Please report only completed runs, or provide the measured prefix lengths and a validated extrapolation, and clearly separate 'did not meet recall target' results from the main speedup comparison.
  3. [§5, Eq. (7); Table 5] The APS recall estimator assumes uniform density of data points inside the query hypersphere (Eq. 7), and this assumption is load-bearing for the claim that APS meets recall targets without offline tuning. The only dedicated validation (Table 5) is on SIFT 1M, and on the skewed Wikipedia-12M workload Table 4 reports only recall standard deviation, not mean recall or the fraction of queries below the 90% target. If the true data density is clustered, Eqs. (8)-(9) can systematically over- or underestimate the probability that a partition holds a nearest neighbor, causing premature termination below the recall target or excess scanning. Please report the recall distribution on the skewed workloads and include a controlled experiment with synthetic nonuniform densities (e.g., varying cluster concentration) to show the estimator's error as a function of density divergence.
  4. [§7.2, Baselines; Fig. 4] LIRE and DeDrift are reimplemented inside Quake rather than run as original systems. Because both baselines share Quake's code paths for scanning, k-means, and refinement, the comparison in Table 3 and Figure 4 may reflect implementation quality rather than the algorithms' intended performance. Please validate the reimplementations against the original implementations or published numbers, and state explicitly what was reused and what was rewritten. At minimum, the artifact should contain the reimplementation code so reviewers can inspect the equivalence.
  5. [§4.2.3, ref [27]] The monotonic-convergence claim is not supported in the manuscript. Section 4.2.3 states that total cost across all levels monotonically decreases and the hierarchy converges to a stable state under a fixed workload distribution, with the proof attributed to the technical report [27]; however, reference [27] is listed as arXiv:2506.03437, which is the same identifier as the current paper. If the technical report is not a separate public document, this is a missing proof for a stated safety property. Please include the proof in the paper or an appendix, or cite a distinct technical report with a different identifier.
  6. [§8.1, Table 5] The 'requires no offline tuning' claim for APS is narrower than stated. Section 8.1 says fM has the largest impact on performance, is set between 1% and 10%, and benefits from tuning; Table 5 reports zero offline tuning time for APS, but this excludes selecting fM and other fixed parameters such as τρ, α, and τ. Because fM is chosen per workload in the evaluation, the zero-tuning claim in the abstract and Section 7.6 should be scoped to per-recall-target nprobe selection, not to all parameters, or the paper should demonstrate that fixed defaults meet recall targets without workload-specific fM.
minor comments (4)
  1. [Table 4] Table 4 reports recall standard deviation but not mean recall, so the reader cannot tell whether the configurations actually meet the 90% target; please report the mean recall (and ideally the fraction of queries below target) for each configuration.
  2. [Algorithm 1, line 13] The line 'r ← r = ∑_{i=0}^{m-1} p_i' appears to contain a typo; the update should be 'r ← ∑_{i=0}^{m-1} p_i' or similar.
  3. [§7.2 and Table 3] The caption and text should make the threading configuration explicit in Table 3 itself, since the current wording in Section 7.2 is easy to misread and the asymmetry is central to interpreting the results.
  4. [Throughout] The spelling 'SCANN' and 'ScaNN' are used inconsistently; please pick one form and apply it consistently, including in Table 3 and the references.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity in the main empirical claims; only a secondary convergence proof is deferred to a self-cited technical report.

  1. self citation load bearing [Section 4.2.3 (Safety), also Sections 3 and 4.2.2]
    "Safety: Because every level enforces the same Δ < −τ guard, total cost across all levels monotonically decreases and the hierarchy converges to a stable state under a fixed workload distribution (proof in technical report [27])."

    The paper's theoretical claim that maintenance converges to a stable state (and, per the introduction, 'converges to a local minimum of the cost model') is not proven in the manuscript; the proof is deferred to reference [27], which is the authors' own technical report with the same arXiv identifier and author list as the present paper. The monotonic-decrease half of the claim is indeed enforced by the commit/reject guard in Stage 3, but the convergence and local-minimum assertions rest on an unverified self-citation. This is load-bearing for the theoretical contribution, although the paper's headline latency/recall results are empirical measurements and do not depend on that proof.

full rationale

The central claims of the paper are empirical: measured query and update latencies, recall, and throughput on standard and custom workloads. These results are not derived by fitting a model to the reported outputs, so the main evaluation is not circular. The APS recall model (Eq. 7) is an explicit geometric assumption (uniform density inside the query hypersphere); it is not calibrated to the ground-truth recall, and the paper reports actual recall separately, so the recall estimates are not self-definitional. Similarly, the cost model (Eqs. 1-2) is a proposed estimator used to guide maintenance, with effectiveness demonstrated by measured latency rather than by the model's own predictions. The split/merge deltas are approximate and use tuned parameters (α, τ, fM), but these are not 'predictions' that reduce by construction to the inputs; they are heuristics evaluated empirically. The only circular-adjacent element is the deferral of the convergence proof and full derivations to reference [27], which is the same paper's technical report. This self-citation is not central to the measured performance comparisons, so the overall circularity score is low. Concerns about the Quake-MT versus single-threaded baseline comparison are about resource-matched evaluation, not circularity, and therefore do not raise the circularity score under the review rules.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a set of modeling assumptions and tuned constants rather than on a small set of derived equations. The cost model assumes additive per-partition latency, the APS recall estimate assumes uniform density and half-space partition boundaries, and maintenance delta estimates assume balanced splits and proportional access scaling. Most constants are fixed across workloads, but fM, tau, and alpha are chosen on the benchmarks, which is a nontrivial load on the no-tuning claim. No invented entities are required.

free parameters (7)
  • Maintenance threshold tau = 250 ns
    Minimum predicted latency decrease for a split or merge to be committed; globally fixed across workloads.
  • Split access scaling alpha = 0.9
    Assumed fraction of a parent partition's access frequency inherited by each child in split cost estimates; chosen because it worked well across all benchmarks.
  • Initial candidate fraction fM = 1%-10% (workload-dependent)
    APS starts by considering fM times the nearest centroids; the authors call it the highest-impact search parameter and plan to remove it.
  • APS recompute threshold tau_rho = 1%
    Recompute partition probabilities only when the query radius shrinks by more than this relative amount.
  • Refinement radius r_f = 50 partitions, 1 iteration
    Number of nearby partitions considered for k-means refinement after a split.
  • Access-frequency sliding window W = Equal to maintenance interval (e.g., 100,000 queries)
    Determines the access frequency Al,j used in the cost model.
  • Upper-level recall target tau_r(1) = 99%
    Fixed target for non-base levels to prevent recall error propagation in multi-level search.
assumptions (5)
  • domain assumption Uniform density assumption in Eq. 7: the probability that a partition contains a nearest neighbor equals the volume fraction of the query hypersphere intersecting that partition.
    APS uses this to estimate recall online; real embedding distributions are clustered, so this is an approximation.
  • domain assumption Half-space approximation of Voronoi partition boundaries in the intersection volume calculation.
    Needed to give a closed-form hyperspherical cap volume; the paper states exact Voronoi intersection volumes are infeasible.
  • domain assumption Additive cost model in Eq. 2: total query latency is the sum over partitions of access frequency times scan latency, with no interaction between partitions.
    The maintenance loop optimizes this additive surrogate, not measured end-to-end latency.
  • ad hoc to paper Balanced split and proportional access scaling in the split cost estimate Eq. 6.
    The estimate assumes children are half the parent size and inherit a fixed fraction alpha of parent access frequency; the verify step checks sizes but keeps the frequency assumption.
  • ad hoc to paper Monotonic convergence of the maintenance procedure under a fixed workload distribution.
    The proof is deferred to the technical report, but the claim is used to justify maintenance stability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quake: Adaptive Indexing for Vector Search." pith.science (2026). https://pith.science/paper/ZZVGYJ3X

@misc{pith2026250603437,
  author       = {Pith},
  title        = {Pith review of: Quake: Adaptive Indexing for Vector Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZVGYJ3X}},
  note         = {Machine review of arXiv:2506.03437}
}
read the original abstract

Vector search, the task of finding the k-nearest neighbors of a query vector against a database of high-dimensional vectors, underpins many machine learning applications, including retrieval-augmented generation, recommendation systems, and information retrieval. However, existing approximate nearest neighbor (ANN) methods perform poorly under dynamic and skewed workloads where data distributions evolve. We introduce Quake, an adaptive indexing system that maintains low latency and high recall in such environments. Quake employs a multi-level partitioning scheme that adjusts to updates and changing access patterns, guided by a cost model that predicts query latency based on partition sizes and access frequencies. Quake also dynamically sets query execution parameters to meet recall targets using a novel recall estimation model. Furthermore, Quake utilizes NUMA-aware intra-query parallelism for improved memory bandwidth utilization during search. To evaluate Quake, we prepare a Wikipedia vector search workload and develop a workload generator to create vector search workloads with configurable access patterns. Our evaluation shows that on dynamic workloads, Quake achieves query latency reductions of 1.5-38x and update latency reductions of 4.5-126x compared to state-of-the-art indexes such as SVS, DiskANN, HNSW, and SCANN.

Figures

Figures reproduced from arXiv: 2506.03437 by the authors.

Figure 1
Figure 1. Skewed access patterns of Faiss-IVF index partitions [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Quake Architecture Diagram. Search queries use [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The query hypersphere (centered at q with radius ρ) intersecting partition boundaries. The intersection volumes v1 and v2 correspond to the probability of finding a nearest neighbor in partitions P1 and P2, respectively. 5.1 APS algorithm Algorithm 1 details the APS procedure. Given query q, recall target τR, and the initial candidate fraction fM: 1. Scan partition P0, initializing the query radius ρ. 2. Compute pro… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of single-threaded search latency, re [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Multi-query evaluation on WIKIPEDIA-12M with 10,000 search queries. QPS @ recall=90% is measured for all baselines while varying the batch size. All methods use 16 threads to process queries. ing to Wikipedia page views from December 2021. For Quake, FaissIVF, and SCAN…

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. When to Repair a Graph ANN Index: A Matched-Budget Negative Result, and the Interpolated-Baseline Trap That Hid It

    cs.DB 2026-07 unverdicted novelty 6.0 of 10

    Signal-triggered local repair in graph ANN indexes improves minimum recall@10 by 0.014-0.050 under bursty churn versus fixed-cadence repair at matched budget on SIFT-128 and Fashion-MNIST-784.

Reference graph

Works this paper leans on

48 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [27]

    Ilyas, Theodoros Rekatsinas, and Shivaram Venkataraman

    Jason Mohoney, Devesh Sarda, Mengze Tang, Shi- habur Rahman Chowdhury, Anil Pacaci, Ihab F. Ilyas, Theodoros Rekatsinas, and Shivaram Venkataraman. Quake: Adaptive indexing for vector search (technical report). arXiv preprint arXiv:2506.03437, 2025

    arXiv 2025

  2. [1]

    https://qdrant.tech/

    Qdrant - Vector Database. https://qdrant.tech/

    work page

  3. [2]

    https://big-ann- benchmarks.com/, 2021

    Billion-scale approximate nearest neighbor search chal- lenge: Neurips’21 competition track. https://big-ann- benchmarks.com/, 2021

    work page 2021

  4. [3]

    https://www.pinecone.io, 2024

    Vector database for vector search | pinecone. https://www.pinecone.io, 2024. Accessed on De- cember 4, 2023

    work page 2024

  5. [4]

    https://en.wikipedia.org/wiki/Wikipedia:Pageview_statistics, 2024

    Wikipedia:pageview statistics. https://en.wikipedia.org/wiki/Wikipedia:Pageview_statistics, 2024

    work page 2024

  6. [5]

    Locally- adaptive quantization for streaming vector search

    Cecilia Aguerrebere, Mark Hildebrand, Ishwar Singh Bhati, Theodore Willke, and Mariano Tepper. Locally- adaptive quantization for streaming vector search. arXiv preprint arXiv:2402.02044, 2024

    arXiv 2024

  7. [6]

    DeDrift: Robust Similarity Search under Content Drift

    Dmitry Baranchuk, Matthijs Douze, Yash Upadhyay, and I. Zeki Yalniz. DeDrift: Robust Similarity Search under Content Drift, August 2023. arXiv:2308.02752 [cs]

    work page Pith review arXiv 2023

  8. [7]

    SPANN: Highly-efficient Billion-scale Approxi- mate Nearest Neighbor Search

    Qi Chen, Bing Zhao, Haidong Wang, Mingqin Li, Chuanjie Liu, Zengzhong Li, Mao Yang, and Jingdong Wang. SPANN: Highly-efficient Billion-scale Approxi- mate Nearest Neighbor Search

    work page

Show all 48 references
  1. [8]

    The faiss library, 2024

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The faiss library, 2024

    2024

  2. [9]

    Real-time person- alization using embeddings for search ranking at airbnb

    Mihajlo Grbovic and Haibin Cheng. Real-time person- alization using embeddings for search ranking at airbnb. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 311–320, 2018

    2018

  3. [10]

    Accel- erating Large-Scale Inference with Anisotropic Vector Quantization

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. Accel- erating Large-Scale Inference with Anisotropic Vector Quantization. In Proceedings of the 37th International Conference on Machine Learning , pages 3887–3896. PMLR, November 20...

    2020

  4. [11]

    Applying deep learning to airbnb search

    Malay Haldar, Mustafa Abdool, Prashant Ramanathan, Tao Xu, Shulin Yang, Huizhong Duan, Qing Zhang, Nick Barrow-Williams, Bradley C Turnbull, Brendan M Collins, et al. Applying deep learning to airbnb search. In Proceedings of the 25th ACM SIGKDD International Conference on Kno...

    1927

  5. [12]

    Neural instant search for music and podcast

    Helia Hashemi, Aasish Pappu, Mi Tian, Praveen Chan- dar, Mounia Lalmas, and Benjamin Carterette. Neural instant search for music and podcast. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 2984–2992, 2021

    2021

  6. [13]

    Product quantization for nearest neighbor search

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelli- gence, 33(1):117–128, 2010

    2010

  7. [14]

    Product Quantization for Nearest Neighbor Search

    Herve Jégou, Matthijs Douze, and Cordelia Schmid. Product Quantization for Nearest Neighbor Search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):117–128, January 2011. Conference Name: IEEE Transactions on Pattern Analysis and Ma- chine Intelligence

    2011

  8. [15]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visu...

    1956

  9. [16]

    Concise formulas for the surface area of the intersection of two hyperspherical caps

    Yongjae Lee and Woo Chang Kim. Concise formulas for the surface area of the intersection of two hyperspherical caps. KAIST Technical Report, 2014

    2014

  10. [17]

    Morsel-driven parallelism: a numa-aware query evaluation framework for the many-core age

    Viktor Leis, Peter Boncz, Alfons Kemper, and Thomas Neumann. Morsel-driven parallelism: a numa-aware query evaluation framework for the many-core age. In Proceedings of the 2014 ACM SIGMOD International Conference on Management of Data , SIGMOD ’14, page 743–754, New York, NY ...

    2014

  11. [18]

    Improving approximate nearest neighbor search through learned adaptive early termination

    Conglong Li, Minjia Zhang, David G Andersen, and Yuxiong He. Improving approximate nearest neighbor search through learned adaptive early termination. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data, pages 2539–2554, 2020

    2020

  12. [19]

    Concise formulas for the area and volume of a hyperspherical cap

    Shengqiao Li. Concise formulas for the area and volume of a hyperspherical cap. Asian Journal of Mathematics & Statistics, 4(1):66–70, 2010

    2010

  13. [20]

    https://pytorch.org/cppdocs

    LibTorch: PyTorch C++ API. https://pytorch.org/cppdocs

  14. [21]

    Related pins at pinterest: The evolution of a real-world recommender system

    David C Liu, Stephanie Rogers, Raymond Shiau, Dmitry Kislyuk, Kevin C Ma, Zhigang Zhong, Jenny Liu, and Yushi Jing. Related pins at pinterest: The evolution of a real-world recommender system. In Proceedings of the 26th international conference on world wide web companion, pag...

    2017

  15. [22]

    Monolith: Real time recommendation system with collisionless embedding table

    Zhuoran Liu, Leqi Zou, Xuan Zou, Caihua Wang, Biao Zhang, Da Tang, Bolin Zhu, Yijie Zhu, Peng Wu, Ke Wang, and Youlong Cheng. Monolith: Real time recommendation system with collisionless embedding table. In 5th Workshop on Online Recommender Systems and User Modeling (ORSUM202...

    2022

  16. [23]

    Cracking vector search indexes

    Vasilis Mageirakos, Bowen Wu, and Gustavo Alonso. Cracking vector search indexes. arXiv preprint arXiv:2503.01823, 2025

    2025

  17. [24]

    Malkov and D

    Yu A. Malkov and D. A. Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs.IEEE Trans. Pattern Anal. Mach. Intell., 42(4):824–836, April 2020

    2020

  18. [25]

    Incremental ivf index maintenance for streaming vector search

    Jason Mohoney, Anil Pacaci, Shihabur Rahman Chowd- hury, Umar Farooq Minhas, Jeffery Pound, Cedric Reng- gli, Nima Reyhani, Ihab F Ilyas, Theodoros Rekatsinas, and Shivaram Venkataraman. Incremental ivf index maintenance for streaming vector search. arXiv preprint arXiv:2411.0...

    2024 arXiv

  19. [26]

    Ilyas, Umar Farooq Min- has, Jeffrey Pound, and Theodoros Rekatsinas

    Jason Mohoney, Anil Pacaci, Shihabur Rahman Chowd- hury, Ali Mousavi, Ihab F. Ilyas, Umar Farooq Min- has, Jeffrey Pound, and Theodoros Rekatsinas. High- Throughput Vector Similarity Search in Knowledge Graphs. Proceedings of the ACM on Management of Data, 1(2):1–25, June 2023

    2023

  20. [28]

    Marius: Learning massive graph embeddings on a single ma- chine

    Jason Mohoney, Roger Waleffe, Henry Xu, Theodoros Rekatsinas, and Shivaram Venkataraman. Marius: Learning massive graph embeddings on a single ma- chine. In 15th {USENIX} Symposium on Operating Sys- tems Design and Implementation ({OSDI} 21), pages 533–549, 2021

    2021

  21. [29]

    https://github.com/cameron314/concurrentqueue

    moodycamel::ConcurrentQueue. https://github.com/cameron314/concurrentqueue

  22. [30]

    https://github.com/ashvardanian/SimSIMD

    SimSIMD. https://github.com/ashvardanian/SimSIMD

  23. [31]

    DiskANN++: Efficient Page-based Search over Isomorphic Mapped Graph Index using Query-sensitivity Entry Vertex, November 2023

    Jiongkang Ni, Xiaoliang Xu, Yuxiang Wang, Can Li, Jia- jie Yao, Shihai Xiao, and Xuecang Zhang. DiskANN++: Efficient Page-based Search over Isomorphic Mapped Graph Index using Query-sensitivity Entry Vertex, November 2023. arXiv:2310.00402 [cs]

    2023 arXiv

  24. [32]

    Embedding-based news recommenda- tion for millions of users

    Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. Embedding-based news recommenda- tion for millions of users. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1933–1942, 2017

    1933

  25. [33]

    Pinner- sage: Multi-modal user embedding framework for rec- ommendations at pinterest

    Aditya Pal, Chantat Eksombatchai, Yitong Zhou, Bo Zhao, Charles Rosenberg, and Jure Leskovec. Pinner- sage: Multi-modal user embedding framework for rec- ommendations at pinterest. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Min...

    2020

  26. [34]

    Micronn: An on-device disk-resident updatable vector database

    Jeffrey Pound, Floris Chabert, Arjun Bhushan, Ankur Goswami, Anil Pacaci, and Shihabur Rahman Chowd- hury. Micronn: An on-device disk-resident updatable vector database. arXiv preprint arXiv:2504.05573, 2025

    2025 arXiv

  27. [35]

    Adaptive numa-aware data placement and task scheduling for ana- lytical workloads in main-memory column-stores

    Iraklis Psaroudakis, Tobias Scheuer, Norman May, Ab- delkader Sellami, and Anastasia Ailamaki. Adaptive numa-aware data placement and task scheduling for ana- lytical workloads in main-memory column-stores. Proc. VLDB Endow., 10(2):37–48, October 2016

    2016

  28. [36]

    Mixer: efficiently understanding and retrieving visual content at web-scale

    An Qin, Mengbai Xiao, Yongwei Wu, Xinjie Huang, and Xiaodong Zhang. Mixer: efficiently understanding and retrieving visual content at web-scale. Proceedings of the VLDB Endowment, 14(12):2906–2917, 2021

    2021

  29. [37]

    Learning transferable vi- sual models from natural language supervision, 2021

    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. Learning transferable vi- sual models from natural language supervision, 2021

    2021

  30. [38]

    Freshdiskann: A fast and accurate graph-based ann in- dex for streaming similarity search

    Aditi Singh, Suhas Jayaram Subramanya, Ravis- hankar Krishnaswamy, and Harsha Vardhan Simhadri. Freshdiskann: A fast and accurate graph-based ann in- dex for streaming similarity search. arXiv preprint arXiv:2105.09613, 2021

    2021 arXiv

  31. [39]

    DiskANN: fast accurate billion-point nearest neighbor search on a single node

    Suhas Jayaram Subramanya, Devvrit, Rohan Kadekodi, Ravishankar Krishaswamy, and Harsha Vardhan Simhadri. DiskANN: fast accurate billion-point nearest neighbor search on a single node. Curran Associates Inc., Red Hook, NY , USA, 2019

    2019

  32. [40]

    Soar: Improved indexing for approx- imate nearest neighbor search

    Philip Sun, David Simcha, Dave Dopson, Ruiqi Guo, and Sanjiv Kumar. Soar: Improved indexing for approx- imate nearest neighbor search. In Neural Information Processing Systems, 2023

    2023

  33. [41]

    Mariusgnn: Resource- efficient out-of-core training of graph neural networks

    Roger Waleffe, Jason Mohoney, Theodoros Rekatsinas, and Shivaram Venkataraman. Mariusgnn: Resource- efficient out-of-core training of graph neural networks. In ACM SIGOPS European Conference on Computer Systems (EuroSys), 2023

    2023

  34. [42]

    Milvus: A purpose- built vector data management system

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, et al. Milvus: A purpose- built vector data management system. In Proceedings of the 2021 International Conference on Management of Data, pages 2614–2627, 2021

    2021

  35. [43]

    Billion-scale commodity embedding for e-commerce recommendation in alibaba

    Jizhe Wang, Pipei Huang, Huan Zhao, Zhibo Zhang, Bin- qiang Zhao, and Dik Lun Lee. Billion-scale commodity embedding for e-commerce recommendation in alibaba. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 839–848, 2018

    2018

  36. [44]

    Analyticdb- v: A hybrid analytical engine towards query fusion for structured and unstructured data

    Chuangxian Wei, Bin Wu, Sheng Wang, Renjie Lou, Chaoqun Zhan, Feifei Li, and Yuanzhe Cai. Analyticdb- v: A hybrid analytical engine towards query fusion for structured and unstructured data. Proceedings of the VLDB Endowment, 13(12):3152–3165, 2020

    2020

  37. [45]

    SPFresh: Incremen- tal In-Place Update for Billion-Scale Vector Search

    Yuming Xu, Hengyu Liang, Jin Li, Shuotao Xu, Qi Chen, Qianxi Zhang, Cheng Li, Ziyue Yang, Fan Yang, Yuqing Yang, Peng Cheng, and Mao Yang. SPFresh: Incremen- tal In-Place Update for Billion-Scale Vector Search. In Proceedings of the 29th Symposium on Operating Sys- tems Princi...

    2023

  38. [46]

    Embedding entities and relations for learn- ing and inference in knowledge bases

    Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learn- ing and inference in knowledge bases. arXiv preprint arXiv:1412.6575, 2014

    2014 arXiv

  39. [47]

    {VBASE}: Unifying online vector similarity search and relational queries via relaxed monotonicity

    Qianxi Zhang, Shuotao Xu, Qi Chen, Guoxin Sui, Ji- adong Xie, Zhizhen Cai, Yaoqi Chen, Yinxuan He, Yuqing Yang, Fan Yang, et al. {VBASE}: Unifying online vector similarity search and relational queries via relaxed monotonicity. In 17th USENIX Symposium on Operating Systems Des...

    2023

  40. [48]

    Fast, approximate vector queries on very large unstructured datasets

    Zili Zhang, Chao Jin, Linpeng Tang, Xuanzhe Liu, and Xin Jin. Fast, approximate vector queries on very large unstructured datasets. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), pages 995–1011, 2023. A Artifact Appendix Abstract This artifa...

    2023

Pith tools

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