Pith. sign in

REVIEW 3 major objections 6 minor 78 references

ANNLib decouples graph-based ANNS algorithms from their data structures, and the paper claims this modularity costs no throughput: assembled applications match or beat specialized systems.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:34 UTC pith:YATDIVZX

load-bearing objection The chrono prefix array and edge-agent decoupling are real contributions; the snapshot claim is overstated until stale-edge effects on query quality are measured. the 3 major comments →

arxiv 2607.17582 v1 pith:YATDIVZX submitted 2026-07-20 cs.LG cs.IR

ANNLib: A Development Framework for Efficient Approximate Nearest Neighbor Search

Zheqi Shen , Jingbo Su , Zijin Wan , Yan Gu , Yihan Sun This is my paper
classification cs.LG cs.IR
keywords approximate nearest neighbor searchgraph-based ANNSmodular frameworkbeam searchdynamic updatesfiltered searchsnapshotsprefix-shared array
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

ANNLib is a programming framework for graph-based approximate nearest neighbor search built on the claim that the algorithm and the data structure of an ANNS system can be cleanly separated without sacrificing performance. It provides a small set of primitives — collection, pruning, and an edge-agent interface — and shows that four distinct applications (plain search, fully dynamic updates, filtered search, and snapshot queries) can be assembled from them in minimal code. The paper's experiments report that these assembled systems are comparable to or faster than monolithic baselines built specifically for each application, and that a new prefix-sharing container cuts snapshot memory usage by up to 70.4%. The value of the claim, if true, is that ANNS development stops being a choice between feature-rich and fast.

Core claim

The central claim is that ANNS algorithm logic and graph storage can be decoupled and independently optimized without a performance penalty. Concretely, ANNLib exposes collection and pruning as parameterized primitives (f_nbhs, f_dist, pred) and a graph-container interface (edge agent) so that the same code path runs across a static nested array, a functional tree with copy-on-write snapshots, and a new chrono prefix array. The paper argues this decoupling is what lets four advanced applications be implemented with minimal code, and its experiments support the claim that the modular versions match or exceed specialist baselines in construction time, query throughput, and recall.

What carries the argument

The load-bearing mechanism is the edge agent: an intermediate view of a vertex's neighbor list that lets each graph container apply its own optimizations behind a uniform interface, so algorithm code is container-agnostic. Around it sit the collect/prune primitives driven by the callables f_nbhs, f_dist, and pred, which encode beam search and robust pruning in a reusable form. For snapshots, the new chrono prefix array stores each vertex's edge versions as cursors over a shared, reference-counted buffer, appending only the increment of new edges per version; the resulting prefix-shared array compresses consecutive versions by trading exactness for compactness.

Load-bearing premise

The prefix-shared array keeps stale edges in a vertex's list across versions, and the paper assumes these extra edges are rare enough and harmless enough in beam search that recall stays intact; if updates make stale edges accumulate or hurt recall, the memory savings come at the cost of search quality.

What would settle it

Run an experiment that updates a single vertex's neighborhood many times with disjoint edge sets so each new version invalidates most old edges, then query an early snapshot: if recall drops noticeably or the prefix buffer grows without bound, the safety argument fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • ANNLib's four demo applications — plain search, fully dynamic updates with deletion, filtered search, and snapshot queries — are each implemented in minimal code on top of shared primitives, so new applications can be assembled the same way.
  • The chrono prefix array reduces memory for historical snapshots by up to 70.4% compared with a chronological-list scheme, at modest build-time cost.
  • The modular Vamana build is on average 1.58x faster than the fastest prior parallel implementation, and HNSW is comparable, showing decoupling doesn't sacrifice insertion throughput.
  • Users can combine components in ways the paper didn't fully evaluate, such as filtered search on snapshots of a dynamically updated index.

Where Pith is reading between the lines

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

  • If the decoupling claim generalizes, ANNS frameworks can become interoperable building blocks: an algorithm module from one library and a storage module from another could be mixed without rewriting search logic, accelerating research prototyping.
  • The prefix-shared array's memory savings rest on the 'extra edges are harmless' heuristic; a natural stress test is to measure recall and buffer growth under high-update workloads where edge sets change dramatically between versions.
  • The same prefix-sharing idea could apply to other dynamic graph workloads that need versioned adjacency views, such as temporal or streaming graph analytics, wherever stale edges are acceptable.
  • Because the interface separates neighbor traversal from distance computation, it may also support heterogeneous accelerators (e.g., GPU distance kernels with CPU graph traversal) without changing algorithm code.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents ANNLib, a C++ library/framework for graph-based ANNS that decouples algorithmic logic from graph-container data structures behind a small interface. It contributes an edge-agent abstraction, modular collect/prune primitives, built-in Vamana/HNSW/HCNNG support, and a new snapshottable container called the chrono prefix array. Four applications are assembled: plain search, batch deletion/updates, filtered search (Stitched-vamana and Filtered-vamana), and historical snapshots. Experiments compare QPS-recall, insertion time, deletion time, and snapshot memory/build time against ParlayANN, DiskANN, Milvus, and Weaviate, reporting broadly competitive or superior performance and up to 70.4% snapshot memory savings. The central claim is that decoupling and modularity do not cost performance and that advanced applications can be built with minimal code.

Significance. If the results hold, ANNLib is a useful systems contribution: it demonstrates that a modular library can match or beat specialized, heavily optimized ANNS implementations across several applications, and the edge-agent and prefix-shared-array ideas are novel and interesting. The paper ships code, uses strong external baselines, and honestly discloses the low-recall throughput gap on DEEP and OpenAI. However, the snapshot claim currently rests on an unvalidated assumption about stale edges in the chrono prefix array, and the absence of replicate-run statistics weakens the quantitative claims. With those points addressed, the paper would be a solid systems-venue candidate.

major comments (3)
  1. [3.3.5, Figs. 8-9] The snapshot claim is load-bearing for the paper's central claim that ANNLib achieves 'comparable or even better performance' for historical snapshots. Section 3.3.5 explicitly retains stale edges in the prefix-shared array and defends this with an assertion: 'We argue that it is usually safe to contain these extra edges...'. Section 5.5, however, reports only memory and build time; no experiment measures query recall or search latency on snapshot versions, and no experiment studies how effective degree grows under update- or delete-heavy workloads. Since beam-search routing and robustPrune are designed to control degree, stale-edge accumulation could inflate search time and distort recall. This is fixable in scope by adding snapshot query-quality measurements (recall and latency as a function of number of versions and update/delete rate).
  2. [Figs. 4-9 (Section 5)] None of the performance figures reports variance, error bars, or the number of replicate runs. QPS-recall curves are typically noisy, and the paper includes single-point claims such as '1.58x faster on average, up to 8.11x on the last batch of OpenAI' (Section 5.1) and 'up to 70.4% memory savings' (Section 5.5). Without repeated-run statistics, it is hard to tell whether the reported advantages are significant. Please add standard deviations or error bars, or state that curves are representative and justify with at least a sensitivity analysis.
  3. [Section 5.4, Fig. 4] The filtered-search comparison to Milvus and Weaviate may underestimate those baselines. The label set is stored as a comma-separated string and matched with the `like` operator, which is a substring scan rather than an exact-metadata filter path. The paper says it follows Filtered-DiskANN's settings, but this is not equivalent to how Milvus and Weaviate would be configured in production. At minimum, describe the exact baseline configuration (index type, parameter sweep, filter operator) and consider an additional comparison using a native metadata filter if the systems support one.
minor comments (6)
  1. [5.5, Figs. 8-9] Please define the baselines 'Naive', 'PAM', 'Chrono_array', and 'Chrono_prefix', and explain what the 'chronological-list scheme' is; otherwise the memory-savings percentages cannot be interpreted.
  2. [5.1] The statement that the insertion speedup 'primarily stems from copy elision' is asserted without an ablation. Since the copy-elision property of the edge-agent interface is one of the paper's distinctive contributions, please provide a microbenchmark with copy elision disabled or otherwise substantiate the causal attribution.
  3. [Abstract and Section 4] The 'minimal programming effort' claim is not quantified. Consider reporting lines of code or changed lines for the four applications built on ANNLib.
  4. [Section 5] Please state how QPS-recall curves were generated: parameter sweep over which hyperparameters, number of queries, warm-up runs, and the meaning of the specificity percentages in Fig. 4.
  5. [Table 1] The '#Pts.' column for label rows is ambiguous; it appears to be average labels per point rather than number of points. Please rename/clarify.
  6. [3.2.1] The sentence containing 'Pkd-tree' appears to be a typo; please fix.

Circularity Check

0 steps flagged

No significant circularity: ANNLib's claims are benchmark measurements over independently published modules; the stale-edge caveat is an unvalidated assumption, not a circular derivation.

full rationale

ANNLib is a systems/benchmark paper, not a derivation chain. The central claim—that decoupling algorithm and data structure yields comparable or better performance—is supported by direct experiments against external baselines (ParlayANN, DiskANN, Milvus, Weaviate, Filtered-DiskANN) and by implementing published algorithms (Vamana, HNSW, HCNNG, FreshDiskANN deletion, Filtered-DiskANN filtered search). There is no fitted parameter later reported as a prediction: insertion time, QPS-recall curves, deletion time/recall, filtered search, and snapshot memory/build time are all measured outputs. The only self-citations are to published prior systems used as infrastructure or baselines: CPAM/PaC-trees [22], ParlayANN [40], and PAM [56]. These are not invoked as uniqueness theorems or as unverified justification for the framework's central contribution; they are peer-reviewed, code-released components that ANNLib reuses and compares against. Section 3.3.5's justification for retaining stale edges ('We argue that it is usually safe to contain these extra edges...') is an unvalidated assumption and a genuine correctness risk, and the snapshot experiments indeed do not measure recall or query latency. But it is not circular: the paper does not define the chrono prefix array in terms of its claimed performance, nor does it fit parameters to snapshot quality and then 'predict' that quality. The claimed memory savings are direct measurements of the structure's layout. Therefore no step reduces to its own inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

No fitted theory is claimed; the ledger instead captures the design knobs and inherited heuristics the results rest on. The prefix-shared buffer limit is an unstated knob shaping the snapshot trade-off; the 'stale edges are harmless' premise is asserted; and the deletion/filtering applications inherit the correctness of prior algorithms wholesale. The one new entity (chrono prefix array) has direct empirical evidence in the paper.

free parameters (2)
  • prefix-shared array buffer size limit = unstated; example uses limit=4 (Fig. 3)
    The chrono prefix array allocates a new buffer once a batch of new edges exceeds the limit (Section 3.3.5). This knob directly controls the memory-vs-insertion-time trade-off measured in Figs. 8-9, yet the paper never states how it was chosen or whether results are sensitive to it.
  • beam width L and degree bound r = swept per experiment (r=32/48 in snapshot tests)
    Standard ANNS hyperparameters varied to trace QPS-recall curves; they are experiment settings, not fitted constants, but every recall value in the figures depends on them.
axioms (5)
  • domain assumption Graph-based ANNS heuristics (beam search + robustPrune + degree-bound pruning) yield high recall as assumed by Vamana/HNSW/HCNNG.
    Inherited from DiskANN [55] and HNSW [39] without re-derivation; Sections 2.2 and 3.2 build on these heuristics as given.
  • domain assumption The prefix-shared array's retained extra edges are few and harmless to recall.
    Stated in Section 3.3.5: 'such deviation is small in practice and extra edges in beam search do not impair the query recall.' Unproved; drives the snapshot memory results.
  • domain assumption FreshDiskANN's mark-and-consolidate deletion restores graph navigability.
    Section 4.2 adopts the deletion algorithm of [53] wholesale; the post-consolidation recall numbers in Fig. 7 depend on it.
  • standard math CPAM (PaC-trees) and PAM functional trees are correct and race-free as published.
    The tree-of-seq containers build on [22, 56]; ANNLib inherits their parallel functional-collection properties (Sections 3.3.4, 4.4).
  • ad hoc to paper Copy elision through the edge-agent interface is the true cause of the insertion speedup.
    Section 5.1 attributes a 1.58x average and up to 8.11x insertion speedup over ParlayANN to 'copy elision' via the interface design; this explanation is asserted post hoc and is not isolated by an ablation (same algorithm, with and without the agent).
invented entities (1)
  • chrono prefix array (prefix-shared array with reference-counted buffer and cursors) independent evidence
    purpose: store per-vertex edge histories across snapshots by appending only the increment and sharing unchanged prefixes, avoiding full graph copies
    Implemented and directly benchmarked in Figs. 8-9 with claimed up-to-70.4% memory reduction; the falsifiable handle is the released code and reproducible memory/time measurements.

pith-pipeline@v1.3.0-alltime-deepseek · 19466 in / 21941 out tokens · 182132 ms · 2026-08-01T17:34:44.421423+00:00 · methodology

0 comments
read the original abstract

Approximate Nearest Neighbor Search (ANNS) plays a pivotal role in modern deep learning pipelines. Recently, many ANNS systems have been proposed to either provide broad functionality or reach high performance. However, it is yet difficult to achieve both with minimal programming efforts. We propose ANNLib to address the gap. ANNLib is a library that provides a programming framework for achieving high performance and flexible functionality in ANNS systems, based on popular graph-based ANNS algorithms. We carefully decouple and independently optimize both the algorithm and the data structure components of an ANNS system. In addition, we integrate state-of-the-art algorithms and data structures into ANNLib as modules, along with our new designs. Users can choose combinations of components to implement sophisticated settings with high performance, such as filter search, fully dynamic updates, and historical queries on snapshots. Our experiments show that our new solution provides a simple interface for various applications and achieves comparable or even better performance than previous work, specifically for each application.

Figures

Figures reproduced from arXiv: 2607.17582 by Jingbo Su, Yan Gu, Yihan Sun, Zheqi Shen, Zijin Wan.

Figure 1
Figure 1. Figure 1: An overview of ANNLib. Users can combine dif￾ferent components, including the base algorithm, algorithm modules and the graph container data structure, as needed by their applications. to-use interface over performance, and thus underper￾form specialized algorithms. Conversely, performance￾optimized systems [1, 13, 43, 53, 55] consist of code re￾fined over years with sophisticated optimizations, mak￾ing th… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of the beam search and the robust prune. gorithms build on a common primitive, beam search, discussed below. Beam Search. A beam search traverses a graph 𝐺 = (𝑉 , 𝐸) from a starting point 𝑠 ∈ 𝑉 towards the query point 𝑞. As shown in Alg. 1, it maintains a candidate queue C with the capacity of beam width 𝐿, which ini￾tially only contains 𝑠 and keeps updating iteratively. At each iteration, the alg… view at source ↗
Figure 3
Figure 3. Figure 3: shows an example: a new version appends only the increment (e.g., 𝐴 at 𝑡=2) and may retain a harmless extra edge (e.g., 𝐷). Once the buffer hits its size limit, a fresh array is allocated (at 𝑡=3). C D E C E A C B ⋅ len=3 C D E ⋅ len=3 C D E A ⋅ len=4 ⋅ len=3 C D E A ⋅ len=4 ⋅ len=2 C B timeline neighbors prefix array (limit=4) t=1 t=2 t=3 -D +A -A +B [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: QPS-recall of Filtered Search. The throughput-recall curves performed on varying dataset and labels; x-axis for recall and y-axis for QPS. Each dataset uses three labels for query, indicating the specificity on the top. 25pc means 25% base points contains the queried label. 1 3 5 7 9 Batches (each with 10% points) 50 100 150 Batch Insertion Time (s) BIGANN-100M-128d 1 3 5 7 9 100 125 150 DEEP-100M-96d 1 3 … view at source ↗
Figure 5
Figure 5. Figure 5: Incremental Insertion Time. x-axis indicates ten even-divided batches to insert; y-axis shows the per-batch insertion time. 0.25 0.50 0.75 1.00 Recall 0.0 0.5 1.0 1.5 Throughput (QPS) ×10 6 BIGANN-100M-128d 0.25 0.50 0.75 1.00 0 1 2 ×10 6 Deep-100M-96d 0.7 0.8 0.9 2 4 ×10 4 Cohere-10M-768d 0.25 0.50 0.75 1.00 0 1 2 ×10 5 OpenAI-5M-1536d ANNlib ParlayANN Milvus [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: QPS-recall of regular query. The optimal throughput-recall tradeoff over parameter sweeping on varying dataset. layANN is known for fast construction and high query throughput due to its lock-free parallel scheme. We evaluate insertion (Section 5.1) and query performance (Section 5.2) relative to ParlayANN on Vamana and HNSW. DiskANN [52] supports batch deletion proposed in FreshDiskANN [53], and incorpora… view at source ↗
Figure 7
Figure 7. Figure 7: Deletion Time and Search Quality. Our main algorithm is “consolidation-ANNlib”. “FreshDiskANN” is the original code from [53]. Other solutions are baselines for comparison. “Rebuild” means to always fully rebuild the graph upon deletions. “mark” means to only mark the deleted points without removing them. 0 10 20 30 40 50 Batches (each with 2% points) 5 10 15 Memory usage (GiB) Deg=32, Input=1M 0 10 20 30 … view at source ↗
Figure 8
Figure 8. Figure 8: Memory usage for snapshots. Incrementally insert 50 batches and monitor the memory usage in varying degree and base size. 0 10 20 30 40 50 Batches (each with 2% points) 0.1 0.2 0.3 0.4 Batch Insertion Time (s) Deg=32, Input=1M 0 10 20 30 40 50 2 4 Deg=32, Input=10M 0 10 20 30 40 50 0.2 0.4 0.6 Deg=48, Input=1M 0 10 20 30 40 50 2 4 6 8 Deg=48, Input=10M Naive, single Naive, snapshots PAM, single PAM, snapsh… view at source ↗
Figure 9
Figure 9. Figure 9: Build time for snapshots. Per-batch build time of incrementally inserting 50 batches and saving all historical versions. avoiding full edge copies matters more. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

78 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Hnswlib - fast approximate nearest neighbor search

    2019. Hnswlib - fast approximate nearest neighbor search. Web- page. Retrieved December 16, 2019 from https://github.com/ nmslib/hnswlib

    2019

  2. [2]

    Apache Lucene

    2023. Apache Lucene. https://lucene.apache.org/

    2023

  3. [3]

    Microsoft Bing Search Engine

    2023. Microsoft Bing Search Engine. https://www.bing.com/new

    2023

  4. [4]

    Pinecone: Vector Database for Vector Search

    2023. Pinecone: Vector Database for Vector Search. https: //www.pinecone.io/

    2023

  5. [5]

    Weaviate: The AI Native Vector Database

    2023. Weaviate: The AI Native Vector Database. https://weaviate. io/

    2023

  6. [6]

    ANNlib: A Development Framework for Efficient Approx- imate Nearest Neighbor Search

    2025. ANNlib: A Development Framework for Efficient Approx- imate Nearest Neighbor Search. https://github.com/ucrparlay/ ANNlib-pub

    2025

  7. [7]

    Laurent Amsaleg and Hervé Jegou. 2010. Datasets for approx- imate nearest neighbor search. http://corpus-texmex.irisa.fr/. [Online; accessed 20-May-2018]

    2010

  8. [8]

    Martin Aumüller, Erik Bernhardsson, and Alexander John Faith- full. 2020. ANN-Benchmarks: A benchmarking tool for approx- imate nearest neighbor algorithms.Information Systems87 (2020)

    2020

  9. [9]

    ANN Benchmarks Authors. 2023. ANN-Benchmarks. https: //ann-benchmarks.com/index.html

    2023

  10. [10]

    Artem Babenko and Victor Lempitsky. 2016. Efficient indexing of billion-scale datasets of deep descriptors. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2055–2063

    2016

  11. [11]

    Dmitry Baranchuk and Artem Babenko. 2021. Benchmarks for Billion-Scale Similarity Search. Webpage. Retrieved March 16, 2023 from https://research.yandex.com/blog/benchmarks-for- billion-scale-similarity-search

    2021

  12. [12]

    Blelloch, Daniel Anderson, and Laxman Dhulipala

    Guy E. Blelloch, Daniel Anderson, and Laxman Dhulipala. 2020. ParlayLib - A Toolkit for Parallel Algorithms on Shared-Memory Multicore Machines. InACM Symposium on Parallelism in Al- gorithms and Architectures (SPAA). ACM, 507–509. https: //doi.org/10.1145/3350755.3400254

    arXiv 2020

  13. [13]

    Blelloch and Magdalen Dobson

    Guy E. Blelloch and Magdalen Dobson. 2022. Parallel Nearest Neighbors in Low Dimensions with Batch Updates. InProceed- ings of the Symposium on Algorithm Engineering and Experi- ments (ALENEX). SIAM, 195–208. https://doi.org/10.1137/1. 9781611977042.16

    doi:10.1137/1 2022

  14. [14]

    CJ Carey, Jonathan Halcrow, Rajesh Jayaram, Vahab Mirrokni, Warren Schudy, and Peilin Zhong. 2022. Stars: Tera-Scale Graph Building for Clustering and Learning. InAdvances in Neural Information Processing Systems, Vol. 35. 21470–21481

    2022

  15. [15]

    Harrison Chase. 2023. Vector DB text generation. https://python.langchain.com/en/latest/modules/chains/ index_examples/vector_db_text_generation.html

    2023

  16. [16]

    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 Data- base System in SingleStore.Proceedings of the VLDB Endowment 17, 12 (2024), 3772–3785

    2024

  17. [17]

    Qi Chen, Xiubo Geng, Corby Rosset, Carolyn Buractaon, Jingwen Lu, Tao Shen, Kun Zhou, Chenyan Xiong, Yeyun Gong, Paul Bennett, et al . 2024. MS MARCO Web Search: a Large-scale Information-rich Web Dataset with Millions of Real Click Labels. InCompanion Proceedings of the ACM on Web Conference 2024. 292–301

    2024

  18. [18]

    2018.SPTAG: A library for fast approximate nearest neigh- bor search

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

    2018

  19. [19]

    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- hood Search. InAnnual Conference on Neural Information Pro- cessing Systems (NeurIPS). 5199–5212

    2021

  20. [20]

    Cohere. 2023. Wikipedia (en) embedded with cohere.ai multilingual-22-12 encoder. https://huggingface.co/datasets/ Cohere/wikipedia-22-12-en-embeddings. [Online; accessed 22-January-2025]

    2023

  21. [21]

    SpaceV Contributors. 2021. SPACEV1B: A billion-Scale vector dataset for text descriptors. Webpage. Retrieved March 16, 2023 from https://github.com/microsoft/SPTAG/tree/main/datasets/ SPACEV1B

    2021

  22. [22]

    Blelloch, Yan Gu, and Yihan Sun

    Laxman Dhulipala, Guy E. Blelloch, Yan Gu, and Yihan Sun. 2022. PaC-trees: supporting parallel and compressed purely-functional collections(PLDI 2022). Association for Computing Machinery, New York, NY, USA, 108–121. https://doi.org/10.1145/3519939. 3523733

    doi:10.1145/3519939 2022

  23. [23]

    Blelloch, and Julian Shun

    Laxman Dhulipala, Guy E. Blelloch, and Julian Shun. 2019. Low- latency graph streaming using compressed purely-functional trees(PLDI 2019). Association for Computing Machinery, 918–

    2019

  24. [24]

    Etienne Dilocker, Bob van Luijt, Byron Voorbach, Mohd Shukri Hasan, Abdel Rodriguez, Dirk Alexander Kulawiak, Marcin Antas, and Parker Duckworth. 2024. Weaviate. Webpage. https://github.com/weaviate/weaviate

    2024

  25. [25]

    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)

    Pith/arXiv arXiv 2024

  26. [26]

    Magdalini Eirinaki, Jerry Gao, Iraklis Varlamis, and Konstanti- nos Tserpes. 2018. Recommender systems for large-scale social networks: A review of challenges and solutions. , 413–418 pages

    2018

  27. [27]

    Joshua Engels, Benjamin Landrum, Shangdi Yu, Laxman Dhuli- pala, and Julian Shun. 2024. Approximate nearest neighbor search with window filters.arXiv preprint arXiv:2402.00943 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Cong Fu, Changxu Wang, and Deng Cai. 2022. High dimensional similarity search with satellite system graph: Efficiency, scala- bility, and unindexed query compatibility.IEEE Trans. Pattern Anal. Mach. Intell.44, 8 (2022), 4139–4150

    2022

  29. [29]

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. 2019. Fast Approximate Nearest Neighbor Search With The Navigating Spreading-out Graph.Proc. VLDB Endow.12, 5 (2019), 461–474

    2019

  30. [30]

    Siddharth Gollapudi, Neel Karia, Varun Sivashankar, Ravis- hankar Krishnaswamy, Nikit Begwani, Swapnil Raz, Yiyong Lin, Yin Zhang, Neelam Mahapatro, Premkumar Srinivasan, Amit Singh, and Harsha Vardhan Simhadri. 2023. Filtered-DiskANN: Graph Algorithms for Approximate Nearest Neighbor Search with Filters. InProceedings of the ACM Web Conference 2023 (WWW ...

    2023

  31. [31]

    Rentong Guo, Xiaofan Luan, Long Xiang, Xiao Yan, Xiaomeng Yi, Jigao Luo, Qianya Cheng, Weizhi Xu, Jiarui Luo, Frank Liu, et al. 2022. Manu: A Cloud Native Vector Database Management System.VLDB Endowment(2022), 3548

    2022

  32. [32]

    Jui-Ting Huang, Ashish Sharma, Shuying Sun, Li Xia, David Zhang, Philip Pronin, Janani Padmanabhan, Giuseppe Ottaviano, and Linjun Yang. 2020. Embedding-based retrieval in facebook search. InProceedings of the 26th ACM SIGKDD International 14 ANNLib: A Development Framework for Efficient ANNS Conference on Knowledge Discovery & Data Mining. 2553–2561

    2020

  33. [33]

    Shikhar Jaiswal, Ravishankar Krishnaswamy, Ankit Garg, Har- sha Vardhan Simhadri, and Sheshansh Agrawal. 2022. OOD- DiskANN: Efficient and Scalable Graph ANNS for Out-of- Distribution Queries.CoRRabs/2211.12850 (2022). https: //doi.org/10.48550/arXiv.2211.12850

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.12850 2022

  34. [34]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion- scale similarity search with GPUs.IEEE Transactions on Big Data 7, 3 (2019), 535–547

    2019

  35. [35]

    Andrew Kane. 2022. pgvector. Webpage. Retrieved July, 2023 from https://github.com/pgvector/pgvector/

    2022

  36. [36]

    Yann LeCun, Yoshua Bengio, et al. 1995. Convolutional networks for images, speech, and time series.The handbook of brain theory and neural networks3361, 10 (1995), 1995

    1995

  37. [37]

    Jie Li, Haifeng Liu, Chuanghua Gui, Jianyu Chen, Zhenyuan Ni, Ning Wang, and Yuan Chen. 2018. The Design and Implemen- tation of a Real Time Visual Search System on JD E-commerce Platform. InProceedings of the 19th International Middleware Con- ference Industry. Association for Computing Machinery, 9–16. https://doi.org/10.1145/3284028.3284030

    arXiv 2018

  38. [38]

    Yanjun Lin, Kai Zhang, Zhenying He, Yinan Jing, and X Sean Wang. 2025. Survey of Filtered Approximate Nearest Neigh- bor Search over the Vector-Scalar Hybrid Data.arXiv preprint arXiv:2505.06501(2025)

    Pith/arXiv arXiv 2025

  39. [39]

    Malkov and Dmitry A

    Yury A. Malkov and Dmitry A. Yashunin. 2020. Efficient and Ro- bust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs.IEEE Trans. Pattern Anal. Mach. Intell.42, 4 (2020), 824–836

    2020

  40. [40]

    Magdalen Dobson Manohar, Zheqi Shen, Guy Blelloch, Laxman Dhulipala, Yan Gu, Harsha Vardhan Simhadri, and Yihan Sun

  41. [41]

    Leland McInnes. 2020. PyNNDescent for Fast Approximate Nearest Neighbors. Webpage. Retrieved December 15, 2022 from https://pynndescent.readthedocs.io/en/latest/

    2020

  42. [42]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality.Advances in neural information pro- cessing systems26 (2013)

    2013

  43. [43]

    Javier Alvaro Vargas Muñoz, Marcos André Gonçalves, Zanoni Dias, and Ricardo da Silva Torres. 2019. Hierarchical Clustering- Based Graphs for Large Scale Approximate Nearest Neighbor Search.Pattern Recognit.96 (2019)

    2019

  44. [44]

    Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. 2022. Text and code embed- dings by contrastive pre-training.arXiv preprint arXiv:2201.10005 (2022)

    Pith/arXiv arXiv 2022

  45. [45]

    Arnold Overwijk, Chenyan Xiong, and Jamie Callan. 2022. ClueWeb22: 10 billion web documents with rich information. InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 3360–3362

    2022

  46. [46]

    Liana Patel, Peter Kraft, Carlos Guestrin, and Matei Zaharia

  47. [47]

    Ninh Pham and Tao Liu. 2022. Falconn++: A Locality-sensitive Filtering Approach for Approximate Nearest Neighbor Search. CoRRabs/2206.01382 (2022). https://doi.org/10.48550/arXiv.2206. 01382 arXiv:2206.01382

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2206 2022

  48. [48]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learning. PmLR, 8748–8763

    2021

  49. [49]

    Acorn: Performant and predicate-agnostic search over vector embeddings and structured data.Proceedings of the ACM on Management of Data2, 3 (2024), 1–27

    2024

  50. [50]

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagen- buchner, and Gabriele Monfardini. 2008. The graph neural net- work model.IEEE transactions on neural networks20, 1 (2008), 61–80

    2008

  51. [51]

    Harsha Vardhan Simhadri, Martin Aumüller, Amir Ingber, Matthijs Douze, George Williams, Magdalen Dobson Manohar, Dmitry Baranchuk, Edo Liberty, Frank Liu, Ben Landrum, et al

  52. [52]

    Jie Ren, Minjia Zhang, and Dong Li. 2020. HM-ANN: Efficient Billion-Point Nearest Neighbor Search on Heterogeneous Mem- ory. InAnnual Conference on Neural Information Processing Sys- tems (NeurIPS), Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (Eds.)

    2020

  53. [53]

    Aditi Singh, Suhas Jayaram Subramanya, Ravishankar Krish- naswamy, and Harsha Vardhan Simhadri. 2021. FreshDiskANN: A Fast and Accurate Graph-Based ANN Index for Streaming Similarity Search.CoRRabs/2105.09613 (2021). arXiv:2105.09613 https://arxiv.org/abs/2105.09613

    Pith/arXiv arXiv 2021

  54. [54]

    Colette Stallbaumer. 2023. Introducing Microsoft 365 copilot. https://www.microsoft.com/en-us/microsoft- 365/blog/2023/03/16/introducing-microsoft-365-copilot-a- whole-new-way-to-work/

    2023

  55. [55]

    Results of the Big ANN: NeurIPS’23 competition.arXiv preprint arXiv:2409.17424(2024)

    Pith/arXiv arXiv 2024

  56. [56]

    Simhadri, Harsha Vardhan and Krishnaswamy, Ravishankar and Srinivasa, Gopal and Subramanya, Suhas Jayaram and An- tonijevic, Andrija and Pryce, Dax and Kaczynski, David and Williams, Shane and Gollapudi, Siddarth and Sivashankar, Varun and Karia, Neel and Singh, Aditi and Jaiswal, Shikhar and Maha- patro, Neelam and Adams, Philip and Tower, Bryan and Pate...

    2023

  57. [57]

    Yukihiro Tagami. 2017. Annexml: Approximate nearest neighbor search for extreme multi-label classification. InProceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining. 455–464

    2017

  58. [58]

    tawalke. 2023. PR: SK Vectordb Connector Work - Merging forked branch; PR from Fork to SK Branch by Tawalke·pull request 83·Microsoft/Semantic-Kernel. https://github.com/ microsoft/semantic-kernel/pull/83

    2023

  59. [59]

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnaswamy, and Rohan Kadekodi

  60. [60]

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, et al . 2021. Milvus: A purpose-built vector data 15 management system. InProceedings of the 2021 International Conference on Management of Data. 2614–2627

    2021

  61. [61]

    Blelloch

    Yihan Sun, Daniel Ferizovic, and Guy E. Blelloch. 2018. PAM: Parallel Augmented Maps. InACM Symposium on Principles and Practice of Parallel Programming (PPoPP)

    2018

  62. [62]

    Chuangxian Wei, Bin Wu, Sheng Wang, Renjie Lou, Chaoqun Zhan, Feifei Li, and Yuanzhe Cai. 2020. AnalyticDB-V: a hy- brid analytical engine towards query fusion for structured and unstructured data.Proceedings of the VLDB Endowment13, 12 (2020), 3152–3165

    2020

  63. [63]

    Jingyi Xi, Chenghao Mo, Benjamin Karsin, Artem Chirkin, Mingqin Li, and Minjia Zhang. 2025. VecFlow: A High- Performance Vector Data Management System for Filtered- Search on GPUs.arXiv preprint arXiv:2506.00812(2025)

    Pith/arXiv arXiv 2025

  64. [64]

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

  65. [65]

    Yuming Xu, Hengyu Liang, Jin Li, Shuotao Xu, Qi Chen, Qianxi Zhang, Cheng Li, Ziyue Yang, Fan Yang, Yuqing Yang, et al

  66. [66]

    Wen Yang, Tao Li, Gai Fang, and Hong Wei. 2020. PASE: Post- greSQL Ultra-High-Dimensional Approximate Nearest Neighbor Search Extension(SIGMOD ’20). 2241–2253

    2020

  67. [67]

    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 at- tribute constraint.Advances in Neural Information Processing Systems36 (2023), 15738–15751

    2023

  68. [68]

    Jiaru Zhang, Ruhui Ma, Tao Song, Yang Hua, Zhengui Xue, Chenyang Guan, and Haibing Guan. 2022. Hierarchical Satellite System Graph for Approximate Nearest Neighbor Search on Big Data.ACM/IMS Trans. Data Sci.2, 4 (2022)

    2022

  69. [69]

    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 Operating Systems Design and Implementation (OSDI 23). USENIX Association, 377– 395

    2023

  70. [70]

    Xiaoliang Xu, Chang Li, Yuxiang Wang, and Yixing Xia. 2020. Multiattribute approximate nearest neighbor search based on navigable small world graph.Concurrency and Computation: Practice and Experience32, 24 (2020), e5970

    2020

  71. [71]

    Chaoji Zuo, Miao Qiao, Wenchao Zhou, Feifei Li, and Dong Deng. 2024. SeRF: segment graph for range-filtering approximate nearest neighbor search.Proceedings of the ACM on Management of Data2, 1 (2024), 1–26. 16

    2024

  72. [74]

    Jianjin Zhang, Zheng Liu, Weihao Han, Shitao Xiao, Ruicheng Zheng, Yingxia Shao, Hao Sun, Hanqing Zhu, Premkumar Srini- vasan, Weiwei Deng, et al. 2022. Uni-retriever: Towards learning the unified embedding based retriever in bing sponsored search. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4493–4501

    2022

  73. [77]

    Yanhao Zhang, Pan Pan, Yun Zheng, Kang Zhao, Yingya Zhang, Xiaofeng Ren, and Rong Jin. 2018. Visual search at alibaba. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 993–1001

    2018

  74. [934]

    https://doi.org/10.1145/3314221.3314598

    arXiv

  75. [2016]

    ACM59, 2 (2016), 64–73

    Yfcc100m: The new data in multimedia research.Commun. ACM59, 2 (2016), 64–73

    2016

  76. [2019]

    InAnnual Conference on Neural Infor- mation Processing Systems (NeurIPS)

    DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node. InAnnual Conference on Neural Infor- mation Processing Systems (NeurIPS). 13748–13758

  77. [2023]

    InProceedings of the 29th Symposium on Operating Systems Principles

    SPFresh: Incremental In-Place Update for Billion-Scale Vector Search. InProceedings of the 29th Symposium on Operating Systems Principles. 545–561

  78. [2024]

    In Proceedings of the 29th ACM SIGPLAN Annual Symposium on Prin- ciples and Practice of Parallel Programming (PPoPP ’24)

    ParlayANN: Scalable and Deterministic Parallel Graph- Based Approximate Nearest Neighbor Search Algorithms. In Proceedings of the 29th ACM SIGPLAN Annual Symposium on Prin- ciples and Practice of Parallel Programming (PPoPP ’24). 270–285