Pith. sign in

REVIEW 5 major objections 6 minor 52 references

HASH-RAG: Bridging Deep Hashing with Retriever for Efficient, Fine Retrieval and Augmented Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Hash-RAG replaces embedding vectors with binary hash codes, cutting retrieval time by roughly 90% while preserving recall and improving exact-match generation scores by 1.4–4.3% on three QA benchmarks.

desk verdict Useful idea, but the central efficiency claim is unverified due to an index-size/corpus-scale mismatch that needs an honest fix. read the letter →

arxiv 2505.16133 v4 pith:GMOTA4WA submitted 2025-05-22 cs.IR

classification cs.IR MSC 68P2068T50
keywords deephashingretrieval-augmentedgenerationHammingdistanceasymmetricpropositionchunkingprompt-guidedretrievalopen-domainquestionansweringbinaryhashcodes
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

This paper proposes Hash-RAG, a retrieval-augmented generation pipeline that replaces continuous embedding vectors with binary hash codes computed through deep hashing. It claims this substitution cuts retrieval time to about a tenth of conventional dense retrievers while preserving recall, and that a companion module called Prompt-Guided Chunk-to-Context (PGCC) turns retrieved hash-indexed propositions and their source documents into a prompt that improves generation accuracy. The experiments on NQ, TriviaQA, and HotpotQA report a 1.4–4.3% exact-match improvement over retrieval and non-retrieval baselines. The practical stake is that RAG systems could serve much larger knowledge bases at lower cost if hashing-based retrieval delivers most of the recall of dense retrieval.

What carries the argument

The load-bearing mechanism is asymmetric deep supervised hashing. Queries are encoded by a BERT encoder followed by a scaled tanh layer approximating the sign function to produce binary codes; propositions are not encoded by a neural network at all. Instead, their codes are learned directly by alternating optimization of a pairwise loss that matches the inner products of binary codes against a similarity matrix, which avoids training an encoder over millions of propositions. A second mechanism, the Prompt-Guided Chunk-to-Context module, chunks documents into self-contained propositions (atomic factual units) and indexes each proposition back to its source document. Retrieval proceeds by expanding the Hamming radius around the query code, and the prompt feeds both propositions and documents to the generator.

What would settle it

Reproduce Table 1 on the advertised corpus scale: Prop-WIKI is reported to contain 261,125,423 propositions, so a 768-bit hash index for all of them should occupy roughly 25 gigabytes, more than five times the 4.6 GB reported. If the reported latency and recall were measured on a smaller index, the 90% retrieval-time reduction may not hold at the claimed scale, which would settle the speed claim either way.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that deep hashing can be integrated into RAG without sacrificing retrieval quality. The Hash-Based Retriever learns 768-bit binary codes for queries and propositions, retrieves by Hamming distance instead of inner product over float vectors, and reduces query latency to roughly 10% of conventional dense retrievers while keeping or slightly improving recall@20 and recall@100 on three open-domain QA benchmarks. The PGCC module then supplies the generator with retrieved propositions together with their original documents and a prompt that instructs the model to integrate both sources; the authors report that this outperforms RAG baselines by 1.4–4.3% in exact match. The paper frames the contribution as an efficiency–accuracy coordination: speed comes from binary codes, accuracy from proposition-level chunking and prompt-guided context.

Load-bearing premise

The pipeline assumes the knowledge base is static: proposition hash codes are produced by an optimization over the full corpus, so any new or edited document would require retraining the hash codes before it can be retrieved.

Editorial extensions

If this is right

  • At 768-bit code length, the hash index for the same corpus is roughly an order of magnitude smaller than a float-vector dense index, so larger knowledge bases fit in the same memory budget.
  • Retrieval latency drops to about 10% of conventional dense retrievers (from roughly 457 ms for DPR to about 42 ms), making the approach usable where query-time budget is tight.
  • Proposition-level chunking alone improves recall@20 over sentence- and paragraph-level chunking, and PGCC with prompts further lifts exact match, so the efficiency gain does not force a generation-quality trade-off.
  • The alternating optimization learns proposition codes without training a proposition encoder, which reduces training time relative to full-database deep hashing baselines such as DSH and DHN.

Reading between the lines

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

  • The asymmetric design could be extended to a two-tier index: keep the optimized hash codes for the stable corpus and route fresh documents through a small learned encoder, patching the static-corpus limitation without full retraining (our inference, not the paper's proposal).
  • Because retrieval is a Hamming-radius scan over binary codes, combining it with product quantization or inverted-file partitioning could lower index size further at a modest recall cost; the paper does not explore this direction.
  • The reported training-speed comparison covers only the hash learning stage; the Propositionizer preprocessing cost is not included in that comparison, so end-to-end index build time would be higher than the figure suggests.
  • The attention heatmap evidence implies PGCC works partly by shifting the model's attention from self-referential diagonal tokens onto proposition tokens, a mechanism that could be probed directly in other RAG baselines.
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

5 major / 6 minor

Summary. The paper proposes Hash-RAG, a retrieval-augmented generation framework that replaces dense-vector search with deep-hashing based retrieval. A query encoder (BERT) is trained against directly learned proposition hash codes via an asymmetric pairwise loss, and a Prompt-Guided Chunk-to-Context (PGCC) module retrieves proposition-level chunks while also supplying the original document context to the generator. Experiments on NQ, TriviaQA, and HotpotQA report recall and EM numbers, claiming a 90% reduction in retrieval time and EM improvements of 1.4-4.3% over baselines. The paper also includes ablations on chunk granularity, prompt design, encoder choice, and training time.

Significance. If the efficiency and accuracy claims are substantiated, Hash-RAG would be a useful demonstration that deep hashing can serve as a practical ANN component in RAG pipelines, with lower storage and latency than dense retrievers. The paper has concrete strengths: it provides a public code link, evaluates on three standard QA benchmarks, compares against several retrieval baselines, and includes ablations of the PGCC module. The main limitation is that the central efficiency claim is not currently supported by the reported index-size numbers, and several headline claims are selectively stated. With clarification and corrected claims, the contribution could be of interest to the IR/RAG community.

major comments (5)
  1. [Section 4.2, Table 1; Section 3.1; Appendix B.2] The index-size numbers in Table 1 are inconsistent with the claimed corpus scale. Section 3.1 fixes the hash code length at l=768 bits, and Appendix B.2 reports Prop-WIKI as containing 261,125,423 propositions. Storing 768-bit codes for all propositions requires 261,125,423 * 96 bytes ≈ 25.1 GB if stored as raw bits, yet Table 1 reports an HbR index size of only 4.6 GB. The same table reports a 64.6 GB DPR index, which matches the standard ~21M-passage NQ corpus (21M * 768 * 4 bytes) rather than a 261M-proposition corpus. This suggests that the retrieval experiments were run on a much smaller corpus than the advertised Prop-WIKI, and the 90% query-time reduction cannot be extrapolated to the claimed scale without clarification. The authors should report the actual number of indexed propositions, the exact storage format, and the corpus used for each row of Table 1.
  2. [Abstract; Section 4.2, Table 2] The abstract claims EM improvements of 1.4-4.3% over retrieval/non-retrieval baselines, but Table 2 does not support this range. On TriviaQA with LLaMA2-7B, Hash-RAG achieves 57.1 EM, identical to the REPLUG baseline (57.1), i.e., a 0.0 improvement; on NQ with LLaMA2-13B the improvement over REPLUG is 5.5 percentage points, which is outside the upper bound of 4.3. The stated range is therefore both incomplete and misleading. The authors should restate the claim using the actual per-dataset, per-model differences, or qualify it as 'up to 5.5%' with the zero-improvement case explicitly reported.
  3. [Section 4.2, Table 1] The headline '90% reduction in retrieval time' is selective. HbR's query time of 42.3 ms is about 91% lower than DPR's 456.9 ms, but compared with the ANN baselines that are the natural efficiency comparators, the reduction is much smaller: PQ is 46.2 ms (about 8% reduction), DSH is 38.1 ms (HbR is slower), and LSH is 28.8 ms (HbR is about 47% slower). Thus the claim 'requires only 10% of the time needed for conventional retrieval methods' holds only against lexical or dense baselines, not against the hashing and quantization baselines listed in the same table. The paper should report speedups over all baselines and define 'conventional methods' precisely.
  4. [Section 3.1, Equations (1)-(8); Limitations] The proposition encoder E_p directly learns binary codes for the fixed knowledge base via alternating optimization, but no mechanism is described for hashing a new or unseen proposition at inference time. The Limitations section acknowledges that the knowledge base is assumed to be static and that incremental updates require retraining, which is an honest caveat. However, Section 3.1 and the abstract present the method as a general retriever without making this static-corpus restriction explicit. The authors should state in Section 3.1 that the current formulation applies to a static corpus whose propositions are hashed during the indexing phase, and clarify whether any inference-time proposition encoder exists.
  5. [Section 5.1, Table 5] The information-bottleneck analysis is presented as validation of proposition-level chunking, but the quantities in Table 5 are not derived from the definitions in Equation (9). The paper reports I(X~; X|Y;Q) without explaining how the relevant distributions are estimated or how the conditional mutual information is computed from the QA datasets. The text itself says 'a potential correlation,' yet the conclusion later states that the approach is 'theoretically optimized and experimentally validated.' This is post-hoc motivation rather than a fitted model prediction, so the language should be softened and the estimation procedure should be described. This does not affect the central recall/EM experiments, but it should be corrected.
minor comments (6)
  1. [Acknowledgments] There are several typos in the Acknowledgments: 'rescarch' should be 'research', 'Scicnce' should be 'Science', 'Burcau' should be 'Bureau', and 'Coopcration' should be 'Cooperation'.
  2. [Section 3.1, Equation (4)] In the regularization term of Equation (4), the subscript of h appears to be pj in the PDF, but based on the surrounding text it should likely be h_pi; please correct this typo.
  3. [Section 3.2, Equation (12)] Equation (12) uses hpj inside an argmax over i; the proposition code should be indexed by the loop variable i (e.g., h_pi) to be consistent with the surrounding text.
  4. [Section 4.3, Table 3 caption] The caption for Table 3 says 'proposition-level chunking achieves significantly superior retrieval performance compared to sentence-level and paragraph-level strategies,' but Table 3 reports results for different encoder versions, not chunking strategies; the caption should be corrected.
  5. [Section 4.1] The sentence 'With more retrieval units, we retrieve additional propositions, map them to source documents, deduplicate, and return the top k unique documents' is awkwardly phrased and should be rewritten for clarity.
  6. [Appendix A] The prompt template is labeled 'Open-domain QA for LLaMA-2-7B,' but Table 2 also reports results with LLaMA2-13B; the prompt appendix should mention whether the same template is used for both model sizes.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: retrieval and generation results are measured against external benchmarks; the only self-citation is minor and non-load-bearing.

full rationale

The paper's central claims—90% retrieval-time reduction and EM improvements—are supported by experiments against external baselines (BM25, DPR, Contriever, MEVI, LSH, DSH) on standard QA benchmarks (NQ, TriviaQA, HotpotQA). The hash-based retriever (HbR) learns query and proposition codes through an asymmetric pairwise loss (Eqs. 3–8), and recall is computed on held-out test sets, so the retrieval numbers are not derived from the paper's own assumptions by construction. The information-bottleneck discussion (Section 3.2, 5.1, Eq. 9) is post-hoc motivation: the paper adopts proposition units from external prior work (Min et al., 2023) and does not solve the IB Lagrangian to derive chunking; Table 5 reports correlations, not a fitted prediction. The prompt-guided chunk-to-context module is evaluated through EM comparisons with baseline RAG systems, again externally measured. The only self-citation is to ReAct (Yao et al., 2022/2023), a co-author's prior work cited for the general claim that RAG alleviates hallucination; this citation is not load-bearing for any of the paper's new results. The acknowledged limitation that the knowledge base is static (Limitations section) is a constraint on generalization, not evidence of circularity. No step in the derivation reduces, by definition or fitted parameter, to the paper's own inputs. Thus the paper is largely self-contained in its empirical evaluation, with only a minor non-load-bearing self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central retrieval result rests on the assumptions that binary codes preserve semantic relevance, that propositions are the right retrieval unit, and that the directly-learned proposition codes can be obtained for the whole static corpus. Several hyperparameters (α, γ, w_k, code length) are chosen without full disclosure, and the corpus size is not reconciled with the advertised index.

free parameters (5)
  • Hash code length l = 768 bits
    Set equal to BERT embedding dimension d=768; no search or ablation over code length is reported, so the choice is unvalidated.
  • tanh scaling schedule σ = 0.1
    Hyperparameter controlling the β schedule in Eq 2; set by hand, no sensitivity analysis beyond γ.
  • Regularization weight γ = not reported (stable in 1 < γ < 500)
    Appears in Eq 4; the paper reports robustness over a range but does not state the value used in the main experiments.
  • Hybrid scoring weight α = not reported (cross-validated)
    In Eq 10, α balances document-level and proposition-level scores; optimized via cross-validation, value not disclosed.
  • Proposition weights w_k = not reported (cross-validated)
    Weights for propositions in hybrid scoring Eq 10; values not disclosed.
assumptions (4)
  • domain assumption Sign function can be approximated by scaled tanh without loss of retrieval fidelity.
    Used throughout Section 3.1, Eq 2; the approximation converges to sign only as β→∞, but finite β is used.
  • domain assumption Hamming distance over learned binary codes approximates semantic relevance for open-domain QA.
    The entire HbR retrieval relies on this, Eq 11; no analysis of hash code quality beyond recall numbers.
  • domain assumption Propositions are self-contained atomic units sufficient for answer generation.
    Borrowed from FactScore; the paper's ablation supports it for HotpotQA, but it is assumed for all datasets.
  • ad hoc to paper The information bottleneck objective justifies proposition chunking.
    Eq 9 is presented as theoretical grounding, but the objective is never optimized or measured in training; it is post-hoc.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HASH-RAG: Bridging Deep Hashing with Retriever for Efficient, Fine Retrieval and Augmented Generation." pith.science (2026). https://pith.science/paper/GMOTA4WA

@misc{pith2026250516133,
  author       = {Pith},
  title        = {Pith review of: HASH-RAG: Bridging Deep Hashing with Retriever for Efficient, Fine Retrieval and Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GMOTA4WA}},
  note         = {Machine review of arXiv:2505.16133}
}
read the original abstract

Retrieval-Augmented Generation (RAG) encounters efficiency challenges when scaling to massive knowledge bases while preserving contextual relevance. We propose Hash-RAG, a framework that integrates deep hashing techniques with systematic optimizations to address these limitations. Our queries directly learn binary hash codes from knowledgebase code, eliminating intermediate feature extraction steps, and significantly reducing storage and computational overhead. Building upon this hash-based efficient retrieval framework, we establish the foundation for fine-grained chunking. Consequently, we design a Prompt-Guided Chunk-to-Context (PGCC) module that leverages retrieved hash-indexed propositions and their original document segments through prompt engineering to enhance the LLM's contextual awareness. Experimental evaluations on NQ, TriviaQA, and HotpotQA datasets demonstrate that our approach achieves a 90% reduction in retrieval time compared to conventional methods while maintaining considerate recall performance. Additionally, The proposed system outperforms retrieval/non-retrieval baselines by 1.4-4.3% in EM scores.

Figures

Figures reproduced from arXiv: 2505.16133 by the authors.

Figure 1
Figure 1. Framework of Deep Supervised Hashing with Pairwise Similarity. The framework computes similarity-preserving loss by aligning pairwise rela￾tionships between hash codes and their corresponding ground truth, which provide a novel perspective for op￾timizing RAG efficiency. expands the capabilities of LLMs in few- or zero￾shot settings (Brown et al., 2020; Chowdhery et al., 2023), which is now widely considered a stand… view at source ↗
Figure 2
Figure 2. Framework Overview. (a) Training. The hash-based encoder generates compact query hash codes, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Prompt Guidance Attention Heat Map ficiency, thereby degrading generation accuracy. Prompt Guidance on Attention To investigate how prompts influence attention mechanisms dur￾ing LLM text generation, we employ Recall@1 to identify a document providing optimal factual sup￾port, thereby validating the effectiveness of prompt optimization. We generate comparative attention heatmaps ( [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Hyperparameter γ on NQ dataset. Conclusion We bridge deep hashing with retrieval-augmented generation for efficient, fine-grained knowledge retrieval and context-augmented generation, bal￾ancing the trade-off between the query processing time and recall. Not only as an…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 21 canonical work pages

  1. [1]

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023

  2. [2]

    Erik Bernhardsson. 2015. Annoy (approximate nearest neighbors oh yeah). URL https://github.com/spotify/annoy

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

  4. [4]

    Riccardo Cantini, Fabrizio Marozzo, Giovanni Bruno, and Paolo Trunfio. 2021. Learning sentence-to-hashtags semantic mapping for hashtag recommendation on microblogs. ACM Transactions on Knowledge Discovery from Data (TKDD), 16(2):1--26

  5. [5]

    Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Philip S Yu. 2017. Hashnet: Deep learning to hash by continuation. In Proceedings of the IEEE international conference on computer vision, pages 5608--5617

  6. [6]

    Moses S Charikar. 2002. Similarity estimation techniques from rounding algorithms. In Proceedings of the thiry-fourth annual ACM symposium on Theory of computing, pages 380--388

  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 neighborhood search. Advances in Neural Information Processing Systems, 34:5199--5212

  8. [8]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1--113

Show all 52 references
  1. [9]

    Thanh-Toan Do, Anh-Dzung Doan, and Ngai-Man Cheung. 2016. Learning to hash with binary deep neural network. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 219--234. Springer

  2. [10]

    T Gao, X Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP 2021-2021 Conference on Empirical Methods in Natural Language Processing, Proceedings

  3. [11]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

  4. [12]

    Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In Annual Conference of the North American Chapter of the Association for Computational Linguistics

  5. [13]

    Wenyu Huang, Mirella Lapata, Pavlos Vougiouklis, Nikos Papasarantopoulos, and Jeff Pan. 2023. Retrieval augmented generation with rich answer encoding. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-P...

  6. [14]

    Piotr Indyk and Rajeev Motwani. 1998. Approximate nearest neighbors: towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pages 604--613

  7. [15]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. https://doi.org/10.48550/ARXIV.2112.09118 Unsupervised dense information retrieval with contrastive learning

  8. [16]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24(251):1--43

  9. [17]

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

  10. [18]

    Qing-Yuan Jiang and Wu-Jun Li. 2018. Asymmetric deep supervised hashing. In Proceedings of the AAAI conference on artificial intelligence, volume 32

  11. [19]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

  12. [20]

    Rong Kang, Yue Cao, Mingsheng Long, Jianmin Wang, and Philip S Yu. 2019. Maximum-margin hamming hashing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8252--8261

  13. [21]

    Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

  14. [22]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, volume 1. Minneapolis, Minnesota

  15. [23]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

  16. [24]

    Hanjiang Lai, Yan Pan, Ye Liu, and Shuicheng Yan. 2015. Simultaneous feature learning and hash coding with deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3270--3278

  17. [25]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

  18. [26]

    Haomiao Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. 2016. Deep supervised hashing for fast image retrieval. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2064--2072

  19. [27]

    Xiao Luo, Daqing Wu, Chong Chen, Jinwen Ma, and Minghua Deng. 2021. Deep unsupervised hashing by global and local consistency. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1--6. IEEE

  20. [28]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting in retrieval-augmented large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5303--5315

  21. [29]

    Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence, 42(4):824--836

  22. [30]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empiric...

  23. [31]

    Vatsal Raina and Mark Gales. 2024. Question-based retrieval using atomic units for enterprise rag. arXiv preprint arXiv:2405.12363

  24. [32]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

  25. [33]

    Stephen E Robertson and Steve Walker. 1997. On relevance weights with little relevance information. In Proceedings of the 20th annual international ACM SIGIR conference on Research and development in information retrieval, pages 16--24

  26. [34]

    Ruslan Salakhutdinov and Geoffrey Hinton. 2009. Semantic hashing. International Journal of Approximate Reasoning, 50(7):969--978

  27. [35]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. 2024. Raptor: Recursive abstractive processing for tree-organized retrieval. arXiv preprint arXiv:2401.18059

  28. [36]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539--68551

  29. [37]

    Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Sch \"a rli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210--31227. PMLR

  30. [38]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. Replug: Retrieval-augmented black-box language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computa...

  31. [39]

    Naftali Tishby, Fernando C Pereira, and William Bialek. 2000. The information bottleneck method. arXiv preprint physics/0004057

  32. [40]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

  33. [41]

    Jingdong Wang, Ting Zhang, Nicu Sebe, Heng Tao Shen, et al. 2017. A survey on learning to hash. IEEE transactions on pattern analysis and machine intelligence, 40(4):769--790

  34. [42]

    Zhiruo Wang, Jun Araki, Zhengbao Jiang, Md Rizwan Parvez, and Graham Neubig. 2023. Learning to filter context for retrieval-augmented generation. arXiv preprint arXiv:2311.08377

  35. [43]

    Yair Weiss, Antonio Torralba, and Rob Fergus. 2008. Spectral hashing. Advances in neural information processing systems, 21

  36. [44]

    Rongkai Xia, Yan Pan, Hanjiang Lai, Cong Liu, and Shuicheng Yan. 2014. Supervised hashing for image retrieval via image representation learning. In Proceedings of the AAAI conference on artificial intelligence, volume 28

  37. [45]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language ...

  38. [46]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629

  39. [47]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

  40. [48]

    Hailin Zhang, Yujing Wang, Qi Chen, Ruiheng Chang, Ting Zhang, Ziming Miao, Yingyan Hou, Yang Ding, Xupeng Miao, Haonan Wang, et al. 2024. Model-enhanced vector index. Advances in Neural Information Processing Systems, 36

  41. [49]

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, and Bin Cui. 2024. Retrieval-augmented generation for ai-generated content: A survey. arXiv preprint arXiv:2402.19473

  42. [50]

    Han Zhu, Mingsheng Long, Jianmin Wang, and Yue Cao. 2016. Deep hashing network for efficient similarity retrieval. In Proceedings of the AAAI conference on Artificial Intelligence, volume 30

  43. [51]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  44. [52]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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