Pith. sign in

REVIEW 4 major objections 5 minor 59 references

NaviX: A Native Vector Index Design for Graph DBMSs With Robust Predicate-Agnostic Search Performance

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a graph DBMS hosting an HNSW vector index natively can match or beat specialized vector databases for filtered kNN search, by choosing one of three exploration heuristics per candidate node.

desk verdict Genuinely new adaptive-local heuristic with an extensive, mostly convincing evaluation; the robustness claim needs sensitivity analysis and a direct validation of the switching rule. read the letter →

arxiv 2506.23397 v1 pith:QPLRUL2P submitted 2025-06-29 cs.IR cs.DB

classification cs.IRcs.DB
keywords vectorindexgraphDBMSfilteredkNNsearchpredicate-agnosticHNSWprefilteringadaptive-localheuristicretrievalaugmentedgeneration
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

NaviX is a vector index built natively inside a graph DBMS: the lower layer of the HNSW index is stored in the database's own disk-based relationship tables, so vector search reuses the buffer manager, storage, and query processor rather than running as a separate engine. The paper's central claim is that filtered kNN search — finding the k nearest neighbors of a query vector only among the vectors picked by an arbitrary selection subquery — stays efficient and accurate across different subset sizes and correlations if the search chooses its exploration heuristic one candidate at a time instead of once per query. The proposed adaptive-local rule reads each candidate's local selectivity, the fraction of its HNSW neighbors that survived the filter, and switches among three exploration modes accordingly. If the claim holds, a single database system can serve retrieval-augmented generation and other mixed embedding-plus-filter workloads without a separate specialized vector store.

What carries the argument

The load-bearing mechanism is the per-candidate decision rule inside adaptive-local. For each candidate $c_{min}$ the algorithm computes the local selectivity $\sigma_l = |S(nbrs(c_{min}))|/|nbrs(c_{min})|$ with a bit check in a node semimask (a bit mask of the selected vectors), and a high-\sigma_l branch uses the onehop-s heuristic, which explores only selected first-degree neighbors. Otherwise it estimates how many selected vectors lie within two hops of the candidate using $esv = \sigma \times (M+1) \times M$, where $\sigma$ is the global selectivity of the filtered subset and $M$ the maximum HNSW degree; comparing $esv$ against the leniency-scaled threshold $M \times lf$ ($lf=3$ by default) decides between the directed second-degree heuristic, which pays the cost of computing distances to all first-degree neighbors in order to sort them, and the blind second-degree heuristic, which explores neighborhoods in adjacency order. The esv estimate is the part of the mechanism that predicts when directed exploration's extra distance computations stop paying for themselves.

What would settle it

Construct a filtered workload whose selected subset is globally sparse but tightly clustered around the query vectors, on an HNSW index with a skewed degree distribution containing hub nodes with far more than $M+1$ neighbors. Measure whether adaptive-local reaches the paper's 95% recall target at its maximum tested efSearch setting and whether its end-to-end latency beats the blind second-degree heuristic alone; a failure of either check would indicate the esv estimate is miscalibrated. A more direct test is to compare actual selected-vector counts in first- and second-degree neighborhoods against the esv prediction and look for systematic over- or under-prediction beyond the leniency factor.

Watch

Extended reading notes

Core claim

The paper's contribution is a search algorithm, adaptive-local, that runs on the unmodified HNSW graph and is claimed to be robust for predicate-agnostic filtered search. At each step it computes the local selectivity of the current candidate, $\sigma_l = |S(nbrs(c_{min}))|/|nbrs(c_{min})|$, from a bit mask; above a high-selectivity threshold it explores only selected first-degree neighbors (onehop-s), and below it estimates the number of selected vectors in the first- and second-degree neighborhoods as $esv = \sigma \times (M+1) \times M$, comparing that estimate against a leniency-scaled threshold $M \times lf$ to choose between directed second-degree exploration (first-degree neighbors sorted by distance to the query) and blind second-degree exploration (arbitrary adjacency order). The paper reports that this per-candidate adaptation tracks the best fixed heuristic at every selectivity on uncorrelated workloads, beats the global-selectivity adaptive variant by up to 1.7x on correlated workloads, and is competitive with or faster than specialized vector databases, the prior blind second-degree heuristic, disk-based indices, and postfiltering systems at 95% recall.

Load-bearing premise

The adaptive switch rests on the estimate $esv = \sigma \times (M+1) \times M$: it assumes every HNSW node has about $M+1$ neighbors and that the global selectivity of the filtered subset predicts the local count of selected vectors near any candidate; if real HNSW degrees are uneven or the selected vectors cluster into hotspots, the heuristic switch is miscalibrated and the robustness advantage over fixed heuristics lapses.

Editorial extensions

If this is right

  • An unmodified HNSW graph can serve predicate-agnostic filtered kNN search, so no predicate-specific index construction or graph densification is needed.
  • A graph DBMS can host the index natively in its existing disk-based relationship storage, letting vector search inherit the buffer manager, columnar storage, and query planner of the database.
  • Per-candidate heuristic switching keeps search efficient when the filtered subset correlates with the query's neighborhood, the case where a single global heuristic choice misfires.
  • At high selectivities the upfront cost of evaluating the selection subquery dominates, so the performance win over postfiltering systems is concentrated at low-to-medium selectivities, where filtered search is hardest.
  • Zero-copy distance computation inside the buffer manager frames speeds up vector search by up to 1.6x, an optimization that transfers to any database vector operator.

Reading between the lines

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

  • The esv estimate could be sharpened by replacing the assumed uniform $(M+1)$ degree with each node's actual stored degree and per-node selected-neighbor counts; a version of adaptive-local that caches such local statistics is a natural testable extension the paper does not explore.
  • The same per-candidate decision space could be ported to other graph-based proximity indices whose filtered search currently depends on label-aware densified graphs, potentially removing the need for that densification.
  • The paper's distinction between distance computations on selected vectors and total distance computations suggests a portable diagnostic for any filtered-search system: plot the two against each other across selectivities to locate where directed exploration's overhead exceeds its search-effectiveness gain.
  • Because the paper's join-based selection subqueries show prefiltering cost rising with selectivity, an optimizer that estimates the selection subquery's cost could feed that estimate directly into the heuristic choice, an integration the paper leaves implicit.
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

4 major / 5 minor

Summary. This paper presents NaviX, a native HNSW-based vector index implemented inside the Kuzu graph DBMS, targeting predicate-agnostic filtered kNN search via prefiltering. The paper defines a space of fixed search heuristics (onehop-s, blind and directed second-degree exploration) and two adaptive variants, adaptive-global and adaptive-local, the latter using per-candidate local selectivity to choose among onehop-s, blind, and directed at each iteration. The implementation leverages Kuzu's CSR storage for adjacency lists and introduces an in-buffer-manager distance computation optimization. The evaluation covers four datasets (GIST, Tiny, Arxiv, and a new 15.4M-vector Wiki/DBpedia dataset), selectivities from 1% to 100%, uncorrelated/positively/negatively correlated workloads, 1-hop and 2-hop join selection subqueries, and comparisons against ACORN, Weaviate, Milvus, DiskANN, FilteredDiskANN, iRangeGraph, PGVectorScale, and VBase. The central claim is that adaptive-local is efficient and robust, outperforming fixed heuristics and adaptive-global especially under correlation.

Significance. If the claims are upheld, NaviX is a meaningful systems contribution: it shows that an unmodified HNSW graph can be embedded natively in a GDBMS and deliver predicate-agnostic filtered search that is competitive with specialized vector databases and postfiltering systems, while supporting arbitrary join-based selection subqueries. The adaptive-local idea of switching search heuristics per candidate based on local selectivity is a plausible and potentially generalizable design insight for prefiltering. The paper's strengths include an extensive benchmark suite, a real large-scale graph workload with correlated queries, an open-source artifact, and an ablation of the zero-copy distance optimization. The main weakness is that the robustness claim rests on hand-tuned thresholds and an unvalidated estimator, as detailed below.

major comments (4)
  1. [3.2 (esv formula)] The esv formula in Section 3.2 is stated with a generic sigma; when used for adaptive-local, sigma is the local selectivity sigma_l of the candidate's first-degree neighborhood. The formula esv = sigma x (M+1) x M then assumes both that each node has about M+1 neighbors and that the density of selected vectors in the second-degree neighborhood equals the density in the first-degree neighborhood. Neither assumption is tested. This is load-bearing because the claimed advantage of adaptive-local is largest in the correlated workloads (Section 5.3, Figure 10): under positive correlation, a good candidate's second-degree neighborhood is likely denser than its first-degree neighborhood, so esv underestimates the number of selected second-degree vectors and biases the decision toward blind; under negative correlation the bias is reversed. A miscalibrated switch would reduce adaptive-local to one of the fixed heuristics it is supposed to dominate. I ask for a validation experiment comparing esv to the actual number of selected first- and second-degree neighbors on the uncorrelated and both correlated Wiki workloads, or a sensitivity analysis that varies the esv formula.
  2. [3.2 and 5.1.5-5.1.6] The parameters governing the adaptive decisions are hand-tuned on the evaluation suite: Section 3.2 states 'We found 50% to be a safe choice' for the onehop-s upper bound and sets the leniency factor to lf=3 by default, while Section 5.1.6 fixes the upper-layer sampling ratio at 5%. No sensitivity analysis is provided in the main text or the appendix. Because the paper's central claim is robustness across datasets, selectivities, and correlations, the lack of a parameter sweep leaves open the possibility that the reported superiority over fixed heuristics is an artifact of tuning ub, lf, and the sampling ratio to these workloads. I request a sweep (e.g., ub in {30%,40%,50%,60%}, lf in {1,2,3,4,5}, sampling ratio in {2%,5%,10%}) on at least the Wiki dataset, with the main comparison curves replotted for each setting.
  3. [5.1.7] All latency results are averages over 50 queries with 5 repeated runs (Section 5.1.7), but no standard deviations, error bars, or significance tests are reported anywhere. Several headline comparisons are small (e.g., NaviX vs FAISS-Navix differences in Section 5.5, and NaviX vs adaptive-g on uncorrelated workloads in Figure 10), and buffer-manager and OS cache effects can easily shift sub-millisecond averages. Adding per-query latency distributions or standard deviations for the main Wiki and Arxiv comparisons would let the reader judge whether the claimed 1.3-1.7x advantages are systematic.
  4. [5.1.7 vs 5.8] The DiskANN and FilteredDiskANN experiments in Section 5.8 were executed on a different machine (132GB RAM, 32 vCPUs, 2.5TB SSDs) than the machine used for NaviX and all other baselines (180GB RAM, 32 vCPUs), as stated in Section 5.1.7. Since Section 5.8's conclusions are about relative disk-I/O behavior (e.g., NaviX-cold being 25-27x slower than DiskANN), the hardware difference, especially SSD model and filesystem state, could materially affect the quantitative gaps. I ask for a like-for-like run of NaviX on the DiskANN machine, or at minimum a random/sequential read bandwidth benchmark of both machines to bound the hardware effect.
minor comments (5)
  1. [3.2] There is a symbol inconsistency: the text defines 'esv' but then says 'If esn is less than M'; it should read 'esv'.
  2. [4.1 and 4.2.1] Section 4.1 says 'We did not modify any of these core components,' but Section 4.2.1 describes extending the storage manager interfaces to run distance computations directly on buffer-manager frames; please reconcile these statements.
  3. [Table 1] The onehop-s row says 'only unfiltered nodes'; from Section 3.1 this should be 'only selected nodes', since 'unfiltered' is ambiguous and could mean the opposite.
  4. [Figure 11] The y-axis label 'Heuristic calls' should specify that the values are fractions of candidate explorations, and the caption should state that the distribution is for the Wiki negatively correlated workload.
  5. [5.5] Please clarify whether the ACORN baselines in Section 5.5 use the original ACORN implementation or the improved blind variant described in Section 3.1; the current wording is ambiguous and affects how the comparison should be interpreted.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the adaptive-local heuristic is validated empirically against external systems rather than being defined into its own conclusion.

full rationale

The paper's central claim—that adaptive-local is efficient and robust under varying selectivities and correlations—is established experimentally in Section 5, with comparisons against external systems including ACORN, Weaviate, Milvus, DiskANN, FilteredDiskANN, iRangeGraph, PGVectorScale, and VBase. The adaptive decision rule in Section 3.2 uses esv = σ × (M+1) × M, compared against a leniency-scaled threshold M×lf; this is a stated heuristic decision rule, not a "prediction" whose target is definitionally equal to the rule's inputs. No equation in the paper reduces to another equation by construction, and no fitted parameter is renamed as a predicted result. The 50% onehop-s threshold and lf=3 are hand-picked and disclosed; whether they generalize to other workloads is a legitimate correctness/robustness concern, but it is not circularity because the evaluation measures actual latency and recall against independent baselines. The self-citations to Kuzu [10,15] are background descriptions of the system NaviX is built on, not load-bearing justifications of the search-heuristic claims, and no uniqueness theorem is imported from prior work by the authors. Thus, no circular step meets the evidentiary bar.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central robustness claim rests on two hand-tuned thresholds (50% and lf=3), a uniform-density estimate for selected-neighbor counts, and workload assumptions. No new entities or forces are introduced; the contributions are algorithmic and engineering.

free parameters (4)
  • onehop-s upper-bound threshold = 50% selectivity
    Hand-tuned in Section 3.2 ('We found 50% to be a safe choice here'); controls when adaptive-global and adaptive-local switch to the onehop-s heuristic.
  • leniency factor (lf) = 3
    Hand-tuned multiplier on M in the esv threshold (Section 3.2); sets the decision boundary between blind and directed heuristics.
  • upper-layer sampling ratio = 5%
    Default fraction of nodes in the in-memory upper HNSW level (Section 5.1.6); affects entry-point quality and memory footprint.
  • efs search width = adjusted per query to reach 95% recall
    Search-time knob that trades latency for recall; the paper adjusts efs for every baseline to align recall, so it is not a fixed constant.
assumptions (4)
  • ad hoc to paper The expected number of selected vectors in the 1st and 2nd degree neighborhoods of a candidate is esv = sigma x (M+1) x M, with sigma the global selectivity.
    Section 3.2 introduces this formula to decide when directed loses its advantage; it is a uniform-density estimate not validated against actual degree distributions.
  • domain assumption Exploring 2nd degree neighbors is sufficient for robust filtered search on the evaluated workloads.
    Section 3.1 states 'in our workloads 2nd degree neighbors are enough'; this bounds the search depth and is not proven for other data distributions.
  • domain assumption The constructed workloads (ID-range filters for uncorrelated, Person birth-date joins for correlated, with o1-generated questions) represent the space of predicate-agnostic queries.
    Section 5.1.3: the robustness claim is extrapolated from one relational schema (Wiki/DBpedia) and synthetic query generation.
  • domain assumption Kuzu's node semimasks, CSR storage, and buffer manager behave as described and are representative of GDBMS infrastructure.
    Section 4.2: NaviX's prefiltering and zero-copy distance computation rely on these Kuzu-specific mechanisms; the paper does not modify these core components.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NaviX: A Native Vector Index Design for Graph DBMSs With Robust Predicate-Agnostic Search Performance." pith.science (2026). https://pith.science/paper/QPLRUL2P

@misc{pith2026250623397,
  author       = {Pith},
  title        = {Pith review of: NaviX: A Native Vector Index Design for Graph DBMSs With Robust Predicate-Agnostic Search Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QPLRUL2P}},
  note         = {Machine review of arXiv:2506.23397}
}
read the original abstract

There is an increasing demand for extending existing DBMSs with vector indices so that they become unified systems capable of supporting modern predictive applications, which require joint querying of vector embeddings together with the structured properties and connections of objects. We present NaviX, a native vector index for graph DBMSs (GDBMSs) that has two main design goals. First, we aim to implement a disk-based vector index that leverages the core storage and query-processing capabilities of the underlying GDBMS. To this end, NaviX is built on the Hierarchical Navigable Small-World (HNSW) graph, which itself is a graph-based structure. Second, we aim to support predicate-agnostic filtered vector search queries, in which the k nearest neighbors (kNNs) of a query vector vQ are searched only within an arbitrary subset S of vectors defined by an ad-hoc selection sub-query QS. We adopt a prefiltering approach that evaluates QS first and passes the full description of subset S to the kNN search operator. We study how to design a prefiltering search algorithm that remains robust under varying selectivities and under different correlations between subset S and query vector vQ. We propose an adaptive algorithm that uses the local selectivity of each vector in the HNSW graph to choose an appropriate heuristic at every iteration of the kNN search. Finally, We demonstrate NaviX's robustness and efficiency through extensive experiments against both existing prefiltering- and postfiltering-based baselines.

Figures

Figures reproduced from arXiv: 2506.23397 by the authors.

Figure 1
Figure 1. HNSW Index (replicated from reference [23]). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Very low selectivity 𝑆 can disconnect HNSW. 2.2 Predicate Agnostic Search & Prefiltering We briefly discuss the challenge of predicate-agnostic vector search, which is the problem of finding kNNs of𝑣𝑄 over an arbitrary subset 𝑆 of the vectors, that are selected by a selection subquery𝑄𝑆 . In this paper we focus on the prefiltering approach to evaluate predicate￾agnostic search. In this approach, the system first eva… view at source ↗
Figure 3
Figure 3. Space of heuristics in a modified search algorithm. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Optimal selectivity regions of fixed heuristics. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Pictorial depiction of different correlations [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: NaviX search plan. 4.2.1 In-Buffer Manager Distance Computations. We end this sec￾tion with an optimization that may be of interest to system devel￾opers. As we discussed above, during both index construction and search, vectors are scanned through the buffer manager t…
Figure 7
Figure 7. Figure 7: Wiki Schema Wiki: The above datasets contain objects but no connections be￾tween objects. Therefore we can use them with predicate-agnostic queries where the selection subqueries contain simple filters but not joins. Moreover, these filters are uncorrelated with the qu…
Figure 8
Figure 8. Figure 8: Vector search time vs Selectivity for different heuristics within 95% to 95.5% recall [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: t-dc vs s-dc of blind and directed. picks each heuristic when exploring a candidate’s neighborhoods for each selectivity level. Observe that since adaptive-g’s choice is based on the global selectivity, at each selectivity level, it commits to using one heuristic. Inst…
Figure 10
Figure 10. Figure 10: Vector search time vs selectivity for NaviX and [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 16
Figure 16. Figure 16: Execution Time vs Selectivity for postfiltering baselines with >95% recall and within 1% of each other [PITH_FULL_IMAGE:figures/full_fig_p012_16.png]
Figure 18
Figure 18. Figure 18: DiskANN and Filtered-DiskANN benchmarks at [PITH_FULL_IMAGE:figures/full_fig_p013_18.png]
Figure 19
Figure 19. Figure 19: Vector search time vs selectivity for NaviX and adaptive-g within 95% to 95.5% recall [PITH_FULL_IMAGE:figures/full_fig_p016_19.png]
Figure 21
Figure 21. Figure 21: Vector Search Time with and without zero-copy [PITH_FULL_IMAGE:figures/full_fig_p016_21.png]
Figure 20
Figure 20. Figure 20: Execution Time vs Selectivity for postfiltering [PITH_FULL_IMAGE:figures/full_fig_p016_20.png]
Figure 22
Figure 22. Figure 22: Vector Search Time vs Selectivity for ACORN and Faiss-Navix baselines at different recalls [PITH_FULL_IMAGE:figures/full_fig_p017_22.png]
Figure 23
Figure 23. Figure 23: Execution Time vs Selectivity for Weaviate and Milvus baselines at different Recalls [PITH_FULL_IMAGE:figures/full_fig_p017_23.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 55 canonical work pages

  1. [1]

    Kuzu Github Repo

    2024. Kuzu Github Repo. https://github.com/kuzudb/kuzu

  2. [2]

    Daniel Abadi, Peter Boncz, Stavros Harizopoulos, Stratos Idreos, and Samuel Madden. 2012. The Design and Implementation of Modern Column-Oriented Database Systems. Foundations and Trends® in Databases 5, 3 (2012)

  3. [3]

    arXiv. [n.d.]. Arxiv Dataset. https://info.arxiv.org/help/bulk_data/index.html

  4. [4]

    Martin Aumüller, Erik Bernhardsson, and Alexander Faithfull. 2020. ANN- Benchmarks: A benchmarking tool for approximate nearest neighbor algorithms. Inf. Syst. 87, C (2020)

  5. [5]

    Yuzheng Cai, Jiayang Shi, Yizhuo Chen, and Weiguo Zheng. 2024. Navigating La- bels and Vectors: A Unified Approach to Filtered Approximate Nearest Neighbor Search. SIGMOD 2, 6 (2024)

  6. [6]

    Compute Canada. [n.d.]. Compute Canada Cedar. https://docs.alliancecan.ca/ wiki/Cedar

  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. In NeurIPS

  8. [8]

    Doug Cutting. [n.d.]. Apache Lucene. https://lucene.apache.org/

Show all 59 references
  1. [9]

    DBpedia. 2024. DBpedia Latest Dataset. https://databus.dbpedia.org/dbpedia/ collections/latest-core

  2. [10]

    Xiyang Feng, Guodong Jin, Ziyi Chen, Chang Liu, and Semih Salihoğlu. 2023. Kùzu Graph Database Management System. In CIDR

  3. [11]

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. 2019. Fast approximate nearest neighbor search with the navigating spreading-out graph. PVLDB 12, 5 (2019)

  4. [12]

    Aristides Gionis, Piotr Indyk, and Rajeev Motwani. 1999. Similarity Search in High Dimensions via Hashing. In VLDB

  5. [13]

    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...

  6. [14]

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. 2020. Accelerating large-scale inference with anisotropic vector quantization. In ICML

  7. [15]

    Pranjal Gupta, Amine Mhedhbi, and Semih Salihoglu. 2021. Columnar Storage and List-based Processing for Graph Database Management Systems. PVLDB 14, 11 (2021)

  8. [16]

    HuggingFace. [n.d.]. MTEB Embedding Leaderboard. https://huggingface.co/ spaces/mteb/leaderboard

  9. [17]

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2008. Hamming Embedding and Weak Geometric Consistency for Large Scale Image Search. In ECCV

  10. [18]

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2011. Product Quantization for Nearest Neighbor Search. IEEE PAMI 33, 1 (2011)

  11. [19]

    Andrew Kane. [n.d.]. PGVector. https://github.com/pgvector/pgvector

  12. [20]

    LangChain. 2024. Enhancing RAG-based application accuracy by constructing and leveraging knowledge graphs. https://blog.langchain.dev/enhancing-rag- based-applications-accuracy-by-constructing-and-leveraging-knowledge- graphs/

  13. [21]

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

  14. [22]

    Jinfeng Li, Xiao Yan, Jian Zhang, An Xu, James Cheng, Jie Liu, Kelvin K. W. Ng, and Ti-chung Cheng. 2018. A General and Efficient Querying Method for Learning to Hash. In SIGMOD

  15. [23]

    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 PAMI 42, 4 (2020)

  16. [24]

    Meta. 2018. Faiss. https://github.com/facebookresearch/faiss

  17. [25]

    Meta. 2018. HNSWLib. https://github.com/nmslib/hnswlib

  18. [26]

    Milvus. 2024. Milvus v2.5.0. https://milvus.io/

  19. [27]

    Ilyas, Umar Farooq Minhas, Jeffrey Pound, and Theodoros Rekatsinas

    Jason Mohoney, Anil Pacaci, Shihabur Rahman Chowdhury, Ali Mousavi, Ihab F. Ilyas, Umar Farooq Minhas, Jeffrey Pound, and Theodoros Rekatsinas. 2023. High-Throughput Vector Similarity Search in Knowledge Graphs. SIGMOD 1, 2 (2023)

  20. [28]

    Gonzalo Navarro. 2002. Searching in metric spaces by spatial approximation. VLDB 11, 1 (2002)

  21. [29]

    Aude Oliva and Antonio Torralba. 2001. Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope. Int. J. Comput. Vision 42, 3 (2001)

  22. [30]

    Naoki Ono and Yusuke Matsui. 2023. Relative NN-Descent: A Fast Index Con- struction for Graph-Based Approximate Nearest Neighbor Search. In ACM MM

  23. [31]

    OpenAI. [n.d.]. OpenAI O1 Large Language Model. https://openai.com/o1/

  24. [32]

    Apache Parquet. [n.d.]. Parquet Documentation. https://parquet.apache.org/ docs/

  25. [33]

    Liana Patel, Peter Kraft, Carlos Guestrin, and Matei Zaharia. 2024. ACORN: Per- formant and Predicate-Agnostic Search Over Vector Embeddings and Structured Data. In SIGMOD

  26. [34]

    Mykhailo Poliakov and Nadiya Shvai. 2024. Multi-Meta-RAG: Improving RAG for Multi-Hop Queries using Database Filtering with LLM-Extracted Metadata. arXiv:arXiv:2406.13213 https://arxiv.org/abs/2406.13213

  27. [35]

    Qdrant. [n.d.]. ANN Filtered Dataset. https://github.com/qdrant/ann-filtering- benchmark-datasets

  28. [36]

    Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. HybridRAG: Integrating Knowledge Graphs and Vector Retrieval Augmented Generation for Efficient Information Extraction. In ICAIF. Association for Computing Machinery, New York, N...

  29. [37]

    Nova Search. 2024. stella_en_400M_v5 Embedding Model. https://huggingface. co/NovaSearch/stella_en_400M_v5

  30. [38]

    Gaurav Sehgal and Semih Salihoğlu. [n.d.]. Kuzu Navix Repo. https://github. com/gaurav8297/kuzu

  31. [39]

    Suhas Jayaram Subramanya, Devvrit, Rohan Kadekodi, Ravishankar Kr- ishaswamy, and Harsha Vardhan Simhadri. 2019. DiskANN: fast accurate billion- point nearest neighbor search on a single node. In NeurIPS

  32. [40]

    Philip Sun, David Simcha, Dave Dopson, Ruiqi Guo, and Sanjiv Kumar. 2023. SOAR: improved indexing for approximate nearest neighbor search. In NeurIPS

  33. [41]

    Timescale. 2024. PGVectorScale v0.5.1. https://github.com/timescale/ pgvectorscale

  34. [42]

    Antonio Torralba, Rob Fergus, and William T. Freeman. 2008. 80 Million Tiny Images: A Large Data Set for Nonparametric Object and Scene Recognition.IEEE PAMI 30, 11 (2008)

  35. [43]

    Toussaint

    Godfried T. Toussaint. 1980. The Relative Neighbourhood Graph of a Finite Planar Set. Pattern Recognition 12, 4 (1980)

  36. [44]

    Sentence Transformers. [n.d.]. all-MiniLM-L6-v2 Embedding Model. https: //huggingface.co/sentence-transformers/all-MiniLM-L6-v2

  37. [45]

    Ash Vardanian. [n.d.]. SimSIMD. https://github.com/ashvardanian/SimSIMD

  38. [46]

    Ash Vardanian. [n.d.]. USearch. https://github.com/unum-cloud/usearch

  39. [47]

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xi- angyu 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....

  40. [48]

    Weaviate. [n.d.]. Filtered Vector Search | Weaviate - vector database. https: //weaviate.io/developers/weaviate/concepts/filtering

  41. [49]

    Weaviate. 2024. How we speed up filtered vector search with ACORN. https: //weaviate.io/blog/speed-up-filtered-vector-search

  42. [50]

    Weaviate. 2024. Weaviate v1.28.2. https://weaviate.io/

  43. [51]

    Chuangxian Wei, Bin Wu, Sheng Wang, Renjie Lou, Chaoqun Zhan, Feifei Li, and Yuanzhe Cai. 2020. AnalyticDB-V: A hybrid analytical engine towards query fusion for structured and unstructured data. PVLDB 13, 12 (2020)

  44. [52]

    WikiMedia. 2024. WikiMedia enwiki Dump. https://dumps.wikimedia.org/ enwiki/latest/enwiki-latest-pages-articles.xml.bz2

  45. [53]

    Wei Wu, Junlin He, Yu Qiao, Guoheng Fu, Li Liu, and Jin Yu. 2022. HQANN: Efficient and Robust Similarity Search for Hybrid Queries with Structured and Unstructured Constraints. In CIKM

  46. [54]

    Yuexuan Xu, Jianyang Gao, Yutong Gou, Cheng Long, and Christian S. Jensen

  47. [55]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024. Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering. In ACM SIGIR

  48. [56]

    Wen Yang, Tao Li, Gai Fang, and Hong Wei. 2020. PASE: PostgreSQL Ultra-High- Dimensional Approximate Nearest Neighbor Search Extension. In SIGMOD

  49. [57]

    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. In OSDI

  50. [58]

    Chaoji Zuo, Miao Qiao, Wenchao Zhou, Feifei Li, and Dong Deng. 2024. SeRF: Seg- ment Graph for Range-Filtering Approximate Nearest Neighbor Search.SIGMOD 2, 1 (2024). 15 Figure 19: Vector search time vs selectivity for NaviX and adaptive-g within 95% to 95.5% recall A APPENDIX...

  51. [2024]

    SIGMOD 2, 6 (2024)

    iRangeGraph: Improvising Range-dedicated Graphs for Range-filtering Nearest Neighbor Search. SIGMOD 2, 6 (2024)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.