Pith. sign in

REVIEW 3 major objections 4 minor 28 references

SieveIVF: Threshold-Aware IVF Execution for Large-Scale Training Data Deduplication

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read SieveIVF makes each IVF query stop after a run of empty threshold partitions, cutting deduplication search time by 4-8x while keeping top-10 recall loss small.

desk verdict Solid, well-scoped systems paper: a threshold-aware IVF executor with measured 4-8x speedups at small recall cost; the stopping-rule characterization has a real gap, but the empirical claims hold on their own. read the letter →

arxiv 2608.03199 v1 pith:4ZB3UEOY submitted 2026-08-04 cs.DB

classification cs.DB
keywords trainingdatadeduplicationvectorsearchIVFthresholdretrievalearlystoppingcontinuousbatchinglookaheadschedulingapproximatenearestneighbor
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

Fixed-probe inverted-file (IVF) search spends the same partition budget on every query, even when only neighbors above a similarity threshold can be duplicate candidates. SieveIVF is a threshold-aware executor that watches each query's local search results: after $W$ consecutive partitions return no qualifying candidate, it stops that query early. Because qualifying neighbors concentrate near the front of the partition order in the workloads studied, most queries need far fewer than the fixed 90 probes. To keep batched IVF efficient under per-query stopping, the paper contributes continuous batching, which dynamically groups ready queries by partition, and a lookahead scheduler that enqueues only ranks already committed by the stopping rule. The measured result, with $W=8$ and identical indexes and search parameters, is a 4.1-7.6x speedup on four 10M workloads and 6.1-8.4x speedup on two public 100M workloads, at pooled filtered top-10 recall losses of 0.03-1.13 and 1.43-2.29 percentage points.

What carries the argument

The central mechanism is the consecutive-empty window $W$: after IVF partitions are searched in query-centroid distance order, a query stops when $W$ consecutive partitions return no candidate above the similarity threshold $\tau$, and any qualifying hit resets the counter. This converts the measured front-loading of threshold-qualified neighbors into a predictor-free, per-query probe budget. Continuous batching preserves partition-major execution by dynamically regrouping ready queries by partition, while lookahead scheduling admits only ranks already committed by the stopping rule, keeping low-concurrency batches from serializing.

What would settle it

Run SieveIVF with $W=8$ against fixed-probe IVF on a workload whose exact threshold-filtered top-10 neighbors are not front-loaded, for example a workload where the 90th percentile last productive partition rank is far above the values in Figure 1(b) or where the first eight partitions store well under 90% of qualifying neighbors. If pooled filtered top-10 recall loss rises materially beyond the reported 0.03-2.29 percentage points while speedup also falls below the reported 4-8x range, the front-loading premise is falsified.

Watch

Extended reading notes

Core claim

Fixed-probe inverted-file (IVF) search gives every query the same number of partition probes ($n_{\text{probe}}$), even though in training-data deduplication only neighbors above an application similarity threshold $\tau$ can become duplicate edges. SieveIVF replaces that fixed budget with a per-query stopping rule: walk the partitions in the same centroid-distance order, and stop after $W$ consecutive partition searches return no candidate satisfying $\tau$; any qualifying hit resets the counter. The systems contribution is making this rule fast at batch scale without changing the index or the bounded top-$k$ interface: continuous batching keeps queries grouped by the partition they are rea

Load-bearing premise

The load-bearing premise is that threshold-qualified neighbors are front-loaded in the IVF partition order, so a run of $W$ empty partitions is reliable evidence that a query has left the productive region; the paper also relies on exact centroid assignment at build time so the query-time partition order matches storage.

Editorial extensions

If this is right

  • Deduplication pipelines can adapt per-query probe counts at runtime without choosing one fixed $n_{\text{probe}}$ that must fit both sparse and dense query distributions.
  • The speed/recall tradeoff is controlled by one integer $W$: small $W$ prunes aggressively (up to 15x speedup at $W=2$, with larger recall loss), while large $W$ approaches fixed probing with near-zero loss.
  • Savings do not come only from queries with no qualifying neighbors: even queries with ten threshold-qualified neighbors skip 70-90% of partition visits, because front-loading matters more than neighbor count.
  • The same execution mechanism scales to 100M vectors and works across different embedding dimensions, and it still beats a recall-matched fixed-probe baseline by 1.09-1.22x.
  • Capping the maximum probe budget at $n_{\text{probe}}=24$ still leaves SieveIVF 1.53-2.39x faster with 0.02-0.72 percentage points recall loss.

Reading between the lines

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

  • A natural extension not claimed by the paper: the threshold-empty-run signal should transfer to any search structure with a deterministic partition or segment order, such as coarse quantizers in product-quantization indexes or shard-ordered searches, as a general predicate-aware early-termination pattern.
  • The front-loading property is empirical, and the paper's own dense TABLE workload shows the smallest speedup and largest recall loss; on workloads with diffuse threshold-neighbor placement, $W=8$ would either stop too late, killing the speedup, or stop too early, losing recall.
  • The appendix's in-call calibration selects $W$ per batch and per shard from a sample without exact ground truth; if that holds as data distributions drift, the method becomes self-tuning, but its sample-relative loss target is not a guarantee on full-call exact recall.
  • A deployment could use the reported microbatch-occupancy effect as a diagnostic: when mean queries-per-microbatch falls, lookahead has little room to help, so the main tuning levers become $W$ and threshold selectivity.
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

3 major / 4 minor

Summary. SieveIVF is a threshold-aware IVF executor for embedding-based training-data deduplication. Instead of probing a fixed number of partitions for every query, it stops a query after W consecutive partition searches return no candidate above the similarity threshold, exploiting the empirical observation that qualifying neighbors tend to appear early in the query–centroid probe order. The core systems contribution is preserving partition-major batching under data-dependent per-query stopping: continuous batching dynamically regroups ready queries by partition, and a lookahead scheduler enqueues only the prefix already committed by the stopping rule, with a proof that dispatch timing does not change stopping decisions or returned results. The paper evaluates on four 10M-vector proprietary Hunyuan workloads and two public 100M datasets, reporting 4.1–7.6x and 6.1–8.4x speedups over fixed-probe IVF at W=8 with small pooled filtered top-10 recall losses, plus an ablation separating pruning, batching, lookahead, and an automatic W-calibration mode.

Significance. If the reported results hold, the paper makes a practical contribution: it shows that an application predicate known before search can guide physical IVF work allocation without changing the index format, the within-partition search, or the bounded top-k interface. The evaluation is carefully controlled—same index, same nprobe, k, threshold, and concurrency for both executors—and the use of two public 100M datasets mitigates reliance on the proprietary workloads. The ablation in Section 6.3 cleanly attributes the speedup to pruning, batching, and lookahead, and the lookahead equivalence argument is convincing. The transparent single-knob design (W) and the disclosed exact-centroid-assignment precondition in Section 4.4 are strengths. The main weakness is that the workload characterization used to motivate the stopping rule measures last-hit rank, not the consecutive-empty gap distribution that the rule actually consumes; this is a fixable but load-bearing gap in the mechanism evidence.

major comments (3)
  1. [Section 6.1 / 6.2] The stopping rule consumes runs of W consecutive threshold-empty partitions, but the characterization evidence in Figure 1(b) reports only the rank of the last IVF partition containing a threshold-filtered exact top-10 neighbor. These are not the same quantity: a query whose qualifying neighbors lie at ranks 1 and 16 has last rank 16 yet with W=8 will be stopped at rank 9 and miss the rank-16 neighbor; conversely, small last-rank percentiles do not guarantee that gaps between consecutive hits are smaller than W. Section 3.5 therefore overstates what the presented data establish when it says that a run of empty results is 'runtime evidence' of leaving the productive region. The end-to-end recall losses in Section 6.2 empirically bound the damage on the six tested workloads, but the paper should directly measure the gap distribution—e.g., the distribution of distances between consecutive t
  2. [Section 6.2] The primary quality metric in Section 6.1 defines pooled filtered top-k recall as an absolute recall against exact ground truth, but the main results in the abstract and Section 6.2 report 'recall loss' without stating the baseline. Figure 15 later says 'recall loss against pooled filtered exact top-10 ground truth,' while Section 6.5 compares 'relative to each index's own fixed-probe recall.' These are different quantities, and the interpretation of the headline 0.03–1.13pp and 1.43–2.29pp numbers depends on which baseline is used. Given that fixed probing at nprobe=90 may itself have substantial absolute recall loss, the paper should explicitly state the baseline for every reported loss and, ideally, report both absolute recall and loss relative to fixed probing.
  3. [Section 3.5] The characterization in Section 3.3 is based on exact top-10 neighbors, while the stopping rule operates on the approximate RaBitQ search's threshold-empty runs. The paper notes that exact-GT partition placement differs from approximate search recall, but does not show how the approximate search's empty-run distribution relates to the exact-GT front-loading evidence. A direct measurement of the consecutive-empty-run lengths experienced by the actual executor on the evaluation workloads would close this gap and would also strengthen the external-validity argument for the two public 100M datasets.
minor comments (4)
  1. [Section 4.3] The caption of Figure 4 is dense and the notation r1..r8, t1..t6 is not defined in the caption. Please clarify that r_i are partition ranks and t_i are dispatch rounds, or label the figure more explicitly.
  2. [Section 6.4] Figure 10(b) reports 'the first measured point above 1x is τ=0.35 for WEB, 0.90 for TABLE, and 0.40 for STEM and SCENE.' The relationship between these thresholds and the application thresholds used elsewhere could be stated more clearly, since loose thresholds are explicitly outside the intended operating range.
  3. [Section 6.2] The sentence 'No single window is optimal for every workload. We use W=8 in later experiments as a common operating point' could be softened: Figure 5(a) and 5(b) show that W=8 is a reasonable compromise, but the 'optimal' wording may be ambiguous without a defined cost function over speed and recall.
  4. [Appendix A] The calibrated mode uses a 2% sample and a 0.5% sample-relative loss target, and the paper honestly states that this is not a confidence bound. It would be helpful to report the variance of the selected W across repeated samples, since the shard-level variation in Figure 14(b) suggests sensitivity to the sample.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: stopping rule is a disclosed heuristic; speed and recall claims are empirical with external validation.

full rationale

SieveIVF's stopping rule is a transparent heuristic, not a derived quantity: it stops after W consecutive threshold-empty partitions, where W is a single explicit knob. The workload characterization in Section 3.3 motivates the rule by showing exact neighbors are front-loaded, but the rule is not fitted to recall targets; the full W sweep is shown (Figure 5), and the main result uses a fixed W=8 on both the four Hunyuan workloads and two external 100M workloads (LAION, DEEP), providing generalization evidence. The lookahead scheduler's claimed equivalence to the sequential rule is argued from state invariants (stop line = max over hits of j+W, capped by nprobe) rather than from measured outputs, so it is not circular. The automatic calibration in Appendix A is explicitly framed as sample-relative loss estimation and is evaluated separately against exact ground truth, with its limitations stated. The only substantive concern is that the motivating metric in Figure 1(b) (last productive rank) differs from the consecutive-empty gap distribution that actually triggers a W-stop; this is an evidential gap about robustness, not a circular derivation. The paper's recall numbers directly measure the rule's consequences on six workloads, and Section 4.4 discloses the speed/recall tradeoff and the exact-assignment precondition. No self-citation chain, uniqueness import, or fitted-input-renamed-as-prediction is load-bearing. Therefore no circularity is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method introduces no new storage entities, data structures, or index formats; it is a runtime state machine and scheduler layered on an unmodified IVF index. The only execution knob is W, and the application similarity threshold and nprobe are inputs from the existing deduplication policy.

free parameters (1)
  • W (consecutive-empty window) = 8 in main experiments; swept 2 to 16; automatic calibration in Appendix A
    Controls how many consecutive threshold-empty partition searches trigger stopping. It is chosen by hand as a common operating point in the paper, not learned from the data. The calibration appendix selects W per call using a fixed-probe sample, which is a per-call fitting of the knob.
assumptions (3)
  • domain assumption Threshold-qualified neighbors are front-loaded in the query-centroid probe order.
    Measured on the evaluation workloads in Section 3.3 (Figure 1b). If this property fails, the consecutive-empty stopping rule either prunes too aggressively and loses recall, or does not prune enough and yields little speedup.
  • domain assumption Vectors are assigned to their exact nearest centroid at build time.
    Stated in Section 4.4 as a precondition. Section 6.5 shows that HNSW-assisted approximate assignment increases recall loss, confirming that the stopping signal relies on the search-time centroid order matching the build-time assignment.
  • domain assumption Each local partition search has a deterministic hit-or-empty outcome independent of dispatch time.
    Required for the lookahead equivalence argument in Section 4.3. The index is static and the search is deterministic, so this is a reasonable but unstated standard assumption for the algorithm's correctness proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SieveIVF: Threshold-Aware IVF Execution for Large-Scale Training Data Deduplication." pith.science (2026). https://pith.science/paper/4ZB3UEOY

@misc{pith2026260803199,
  author       = {Pith},
  title        = {Pith review of: SieveIVF: Threshold-Aware IVF Execution for Large-Scale Training Data Deduplication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZB3UEOY}},
  note         = {Machine review of arXiv:2608.03199}
}
abstract

Embedding-based training data deduplication retrieves candidate duplicate edges above an application similarity threshold, but fixed-probe inverted-file (IVF) search ignores this predicate when giving every query the same partition budget. Across four Hunyuan workloads, qualifying neighbors appear early despite sharply varying search depths. We present SieveIVF, a threshold-aware IVF executor that stops after $W$ consecutive searches find no qualifying candidate. The systems challenge is to preserve partition-major batching when each query's remaining work depends on prior results. Continuous batching groups ready queries by partition. A lookahead scheduler layers on top, exposing only work committed by the stopping rule to increase concurrency without changing stopping decisions or returned results. We implement SieveIVF in Lance. At $W=8$, SieveIVF is $4.1$--$7.6\times$ faster than fixed-probe IVF on four 10M Hunyuan workloads and $6.1$--$8.4\times$ faster on two public 100M workloads under the same index and search parameters, with pooled filtered top-10 recall losses of $0.03$--$1.13$ percentage points on Hunyuan and $1.43$--$2.29$ percentage points on the public workloads. These results show how an application predicate can guide IVF work allocation without changing the index or bounded top-$k$ interface.

Figures

Figures reproduced from arXiv: 2608.03199 by the authors.

Figure 1
Figure 1. Exact-GT characterization of Hunyuan workloads. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Threshold-aware stopping over IVF partitions. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Continuous batching groups queries by partition. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Lookahead advances dispatch within the committed [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Speed/recall tradeoff across empty windows. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Scale-out results on two public 100M workloads. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Probe-budget sensitivity and tuned fixed probing. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Execution-mechanism ablation at 𝑊 = 8. Wall time is uninstrumented and counters are diagnostic. 0 1 2-3 4-5 6-9 10 Filtered exact top-10 GT neighbors WEB TABLE STEM SCENE 91.1 89.7 89.5 89.3 89.2 87.9 91.1 89.0 87.6 86.2 84.8 70.0 91.1 89.9 89.7 89.7 89.6 89.3 91.1 89.…
Figure 9
Figure 9. Figure 9: Pruned probes by threshold-neighbor count. [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 12
Figure 12. Figure 12: HNSW-assisted assignment at build time. Bar labels show nearest-centroid assignment mismatch. measured point above 1× is 𝜏 = 0.35 for WEB, 0.90 for TABLE, and 0.40 for STEM and SCENE. In practice, semantic deduplication commonly uses high similarity thresholds [1], so…
Figure 13
Figure 13. Figure 13: Lance and DiskJoin as different deduplication [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]
Figure 15
Figure 15. Figure 15: Performance and recall under fixed and calibrated [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 14
Figure 14. Figure 14: Online window selection across workloads and [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 14 canonical work pages

  1. [1]

    Mor- cos

    Amro Abbas, Kushal Tirumala, Dániel Simig, Surya Ganguli, and Ari S. Mor- cos. 2023. SemDeDup: Data-efficient learning at web-scale through semantic deduplication. arXiv:2303.09540 [cs.LG] https://arxiv.org/abs/2303.09540

  2. [2]

    Fabien André, Anne-Marie Kermarrec, and Nicolas Le Scouarnec. 2015. Cache locality is not enough: high-performance nearest neighbor search with product quantization fast scan.Proc. VLDB Endow.9, 4 (Dec. 2015), 288–299. https: //doi.org/10.14778/2856318.2856324

  3. [3]

    Hellerstein

    Ron Avnur and Joseph M. Hellerstein. 2000. Eddies: continuously adaptive query processing. InProceedings of the 2000 ACM SIGMOD International Conference on Management of Data(Dallas, Texas, USA)(SIGMOD ’00). Association for Computing Machinery, New York, NY, USA, 261–272. https://doi.org/10.1145/ 342009.335420

  4. [4]

    A.Z. Broder. 1997. On the resemblance and containment of documents. InPro- ceedings. Compression and Complexity of SEQUENCES 1997 (Cat. No.97TB100171). 21–29. https://doi.org/10.1109/SEQUEN.1997.666900

  5. [5]

    Charikar

    Moses S. Charikar. 2002. Similarity estimation techniques from rounding algo- rithms. InProceedings of the Thiry-Fourth Annual ACM Symposium on Theory of SieveIVF: Threshold-Aware IVF Execution for Large-Scale Training Data Deduplication Computing(Montreal, Quebec, Canada)(STOC ’02). Association for Computing Machinery, New York, NY, USA, 380–388. https:/...

  6. [6]

    Manos Chatzakis, Yannis Papakonstantinou, and Themis Palpanas. 2025. DARTH: Declarative Recall Through Early Termination for Approximate Nearest Neighbor Search.Proc. ACM Manag. Data3, 4, Article 242 (Sept. 2025), 26 pages. https: //doi.org/10.1145/3749160

  7. [7]

    Qi Chen, Bing Zhao, Haidong Wang, Mingqin Li, Chuanjie Liu, Zengzhong Li, Mao Yang, and Jingdong Wang. 2021. SPANN: highly-efficient billion-scale approximate nearest neighbor search. InProceedings of the 35th International Conference on Neural Information Processing Systems (NIPS ’21). Curran Associates Inc., Red Hook, NY, USA, Article 398, 14 pages

  8. [8]

    Yanqi Chen, Xiao Yan, Alexandra Meliou, and Eric Lo. 2025. DiskJoin: Large-scale Vector Similarity Join with SSD.Proc. ACM Manag. Data3, 6, Article 315 (Dec. 2025), 27 pages. https://doi.org/10.1145/3769780

Show all 28 references
  1. [9]

    Ronald Fagin, Amnon Lotem, and Moni Naor. 2001. Optimal aggregation algo- rithms for middleware. InProceedings of the Twentieth ACM SIGMOD-SIGACT- SIGART Symposium on Principles of Database Systems(Santa Barbara, California, USA)(PODS ’01). Association for Computing Machinery,...

  2. [10]

    Jianyang Gao and Cheng Long. 2024. RaBitQ: Quantizing High-Dimensional Vectors with a Theoretical Error Bound for Approximate Nearest Neighbor Search.Proc. ACM Manag. Data2, 3, Article 167 (May 2024), 27 pages. https: //doi.org/10.1145/3654970

  3. [11]

    Siddharth Gollapudi, Neel Karia, Varun Sivashankar, Ravishankar Krishnaswamy, Nikit Begwani, Swapnil Raz, Yiyong Lin, Yin Zhang, Neelam Mahapatro, Premku- mar Srinivasan, Amit Singh, and Harsha Vardhan Simhadri. 2023. Filtered- DiskANN: Graph Algorithms for Approximate Nearest...

  4. [12]

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2011. Product Quantization for Nearest Neighbor Search.IEEE Trans. Pattern Anal. Mach. Intell.33, 1 (Jan. 2011), 117–128. https://doi.org/10.1109/TPAMI.2010.57

  5. [13]

    Jianfeng Jiang, Shen Xu, and Ying Gao. 2025. A uniformed adaptive early ter- mination model through probabilistic feature to speed up quantization-based search.International Journal of Computers and Applications47, 1 (2025), 17–28. https://doi.org/10.1080/1206212X.2024.2431229

  6. [14]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2021. Billion-Scale Similarity Search with GPUs.IEEE Transactions on Big Data7, 3 (2021), 535–547

  7. [15]

    Kyoungmin Kim, Lennart Roth, Liang Liang, and Anastasia Ailamaki

  8. [16]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. InProceedings of the 29th Symposium on Operating Systems Principl...

  9. [17]

    Lance Format. 2026. Lance: The Open Lakehouse Format for Multimodal AI. https://github.com/lance-format/lance

  10. [18]

    Andersen, and Yuxiong He

    Conglong Li, Minjia Zhang, David G. Andersen, and Yuxiong He. 2020. Im- proving Approximate Nearest Neighbor Search through Learned Adaptive Early Termination. InProceedings of the 2020 ACM SIGMOD International Conference on Management of Data(Portland, OR, USA)(SIGMOD ’20). A...

  11. [19]

    Malkov and D

    Yu A. Malkov and D. A. Yashunin. 2020. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence42, 4 (2020), 824–

  12. [20]

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Lud- wig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. 202...

  13. [21]

    Yitong Song, Pengcheng Zhang, Chao Gao, Bin Yao, Kai Wang, Zongyuan Wu, and Lin Qu. 2025. TRIM: Accelerating High-Dimensional Vector Similarity Search with Enhanced Triangle-Inequality-Based Pruning.Proc. ACM Manag. Data3, 6, Article 373 (Dec. 2025), 26 pages. https://doi.org/...

  14. [22]

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, Kun Yu, Yuxing Yuan, Yinghao Zou, Jiquan Long, Yudong Cai, Zhenxiang Li, Zhifeng Zhang, Yihua Mo, Jun Gu, Ruiyi Jiang, Yi Wei, and Charles Xie. 2021. M...

  15. [23]

    Qian Xu, Juan Yang, Feng Zhang, Junda Pan, Kang Chen, Youren Shen, Amelie Chi Zhou, and Xiaoyong Du. 2025. Tribase: A Vector Data Query Engine for Reliable and Lossless Pruning Compression using Triangle Inequalities.Proc. ACM Manag. Data3, 1, Article 82 (Feb. 2025), 28 pages....

  16. [24]

    Artem Babenko Yandex and Victor Lempitsky. 2016. Efficient Indexing of Billion- Scale Datasets of Deep Descriptors. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2055–2063. https://doi.org/10.1109/CVPR.2016. 226

  17. [25]

    Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung- Gon Chun. 2022. Orca: A Distributed Serving System for Transformer-Based Generative Models. In16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22). USENIX Association, Carlsbad, CA, 521–538

  18. [26]

    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. 2023. VBASE: Unifying Online Vector Similarity Search and Relational Queries via Relaxed Monotonicity. In17th USENIX Symposium on ...

  19. [836]

    https://doi.org/10.1109/TPAMI.2018.2889473

  20. [2026]

    arXiv:2603.16360 [cs.DB] https://arxiv.org/abs/2603.16360

    Fast Approximate Vector Joins via Offline-Online Co-Design. arXiv:2603.16360 [cs.DB] https://arxiv.org/abs/2603.16360

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.