Pith. sign in

REVIEW 4 major objections 4 minor 46 references

VecTree-RAG argues that scientific-QA retrieval should split into two mechanisms: vector search to find papers, tree navigation to find evidence, and reports that this split outscores flat-passage retrieval on all three benchmarks tested.

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 03:51 UTC pith:3SIU3GG3

load-bearing objection A sensible and unusually transparent locate-then-read RAG; the evidence-localization results hold up, but the answer-score claims need out-of-family judging before they can be trusted. the 4 major comments →

arxiv 2607.23006 v1 pith:3SIU3GG3 submitted 2026-07-25 cs.IR cond-mat.mtrl-scics.AI

VecTree-RAG: An Agentic Retrieval-Augmented Generation Framework Combining Vector and Tree Retrieval for Efficiency and Accuracy

Xinyan Zhong , Yuwei Shi , Yuqi Wei , Chen Shen , Tianhang Zhou , Zhenghao Wu This is my paper
classification cs.IR cond-mat.mtrl-scics.AI
keywords retrieval-augmented generationagentic retrievaldocument structuretree retrievalevidence localizationscientific question answeringprogressive disclosurelocate-then-read
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.

Scientific question answering needs two things from retrieval: which papers are relevant and where the evidence sits inside them. The paper argues that a single mechanism cannot do both well, because flat passage search flattens the structure that ties a claim to its methods and qualifications. VecTree-RAG therefore assigns the first job to vector retrieval over compact document and section representations and the second to a reasoning-guided walk over a source-verified section tree, reading page text only after the section is chosen. On single-paper, corpus-level, and multi-paper synthesis benchmarks it reports higher answer correctness than four baseline systems, and substantially higher evidence-page precision: QASPER 0.274 versus 0.046 to 0.071, with an index that scales with structural entries rather than chunk count. The stated caveat is that tree navigation depends on generated section summaries whose faithfulness has not yet been verified.

Core claim

On its own terms, the discovery is that a paper's authored section hierarchy is a retrieval interface rather than noise to be flattened. The system builds a verified tree from each PDF, where nodes are sections with titles, summaries, and inclusive page ranges checked against the source. At query time it first ranks compact document and section representations across the corpus, lets the agent walk the tree, and only then exposes full page text. Against four baselines, this locate-then-read design reports the top score on each benchmark, with much sharper evidence localization: QASPER evidence-page precision 0.274 versus 0.046 to 0.071, and MOSAIC evidence precision 0.221 versus 0.019 to 0.0

What carries the argument

The load-bearing object is the source-verified section tree: each node carries an identifier, a title, an LLM-generated summary, and an inclusive page range, with titles checked by deterministic rules against the parsed source pages so every node resolves to evidence. Around that tree sit three machinery pieces: hybrid dense-plus-lexical retrieval over compact document and section entries for the cross-paper layer, a progressive-disclosure tool interface that returns search metadata first, then tree structure, then page text and images, and a tool-using LLM agent that interleaves reasoning and retrieval actions to move from papers to sections to pages. Together these implement the locate-the

Load-bearing premise

The navigation layer depends on LLM-generated section summaries being faithful enough that choosing a section by its summary really lands on the page containing the evidence; the paper states this faithfulness is unverified.

What would settle it

Take the QASPER set and replace each section summary with a random summary drawn from a different section of the same paper, leaving everything else fixed. If evidence precision stays near 0.274 and correctness near 0.800, section summaries are not the mechanism. If precision collapses toward the flat-baseline 0.05 range, the tree-navigation claim is confirmed to ride on summary fidelity.

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

If this is right

  • If the reported results hold, passage-level vector stores in scientific retrieval can be replaced by structural indexes holding N(1+S) embeddings, with page-resolved source text retained only for shortlisted documents.
  • Evidence localization becomes inspectable: each answer can be traced to a physical page, which supports sentence-level citations and auditability.
  • Ablations show that corpus-level vector routing is the main token saver: dropping it raises per-query cost from about 128k to 490k tokens and lowers LitQA2 accuracy from 0.925 to 0.811.
  • Tree navigation is complementary rather than redundant: removing it raises per-query tokens to about 195k and slightly lowers accuracy, so both mechanisms contribute to efficiency and correctness.
  • The efficiency claim is scoped: multi-turn agent inference remains more expensive than single-call retrieval on QASPER and LitQA2; the gain is in index compactness and selective reading, not total query-time tokens.

Where Pith is reading between the lines

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

  • The same two-mechanism split may transfer to other hierarchical corpora, such as legal rulings, clinical guidelines, or patents, where section structure carries meaning beyond scientific papers.
  • Because evidence recall is lower (0.737 versus 0.817 to 0.899 on QASPER), a natural extension is a final passage-level re-query restricted to the read pages, a hybrid the paper does not test.
  • Progressive disclosure itself is not ablated, so readers cannot yet tell whether the tree layer or the staged payload discipline drives the evidence-precision gain.
  • MOSAIC was generated and judged with the same model family, so the headline composite gains should be re-checked with independent, out-of-family adjudication before being treated as settled.

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

4 major / 4 minor

Summary. VecTree-RAG proposes splitting scientific-QA retrieval by granularity: vector retrieval (SPECTER2 + BM25 over document/section entries) for corpus-level paper discovery, and reasoning-guided traversal of source-verified section trees for within-paper evidence localization. The paper reports evaluations on 300 QASPER questions, 54 LitQA2 questions, and a new 49-question MOSAIC benchmark, claiming the highest observed answer scores on all three benchmarks (0.800, 0.925, 0.547) and substantially higher evidence-page precision on QASPER (0.274 vs. 0.046–0.071) than Dense RAG, reranked Dense RAG, RAPTOR, and Search-o1. It also reports lower query-time token use when vector routing is present, and a compact N(1+S) index. The paper is careful to scope its claims, disclosing unverified summary faithfulness, the self-constructed nature of MOSAIC, and the same-family judge setup.

Significance. If the empirical claims survive scrutiny, the core architectural idea is valuable and timely: decoupling corpus-level discovery from within-paper localization is a natural fit for scientific literature QA, and the page-resolved, source-verified index offers practical traceability. The paper also makes useful methodological contributions: it releases code and evaluation harness, reports a section-title verification rate, provides deterministic agreement diagnostics, and is unusually transparent about limitations. However, the headline answer-accuracy claims rest on a same-family LLM judge and a self-constructed, auto-screened benchmark, so the comparative results must be treated as promising but not yet established.

major comments (4)
  1. [Section 4.2; Appendix E; Tables 2 and 4] All three headline answer scores are produced by DeepSeek-V4-Pro judging DeepSeek-V4-Flash outputs, and MOSAIC reference answers/reasoning chains were drafted by the same family. The paper correctly notes in Appendix E that symmetric scoring does not eliminate family/style bias and that out-of-family regrading is the appropriate test, but it does not perform that test. The deterministic agreement check covers only yes/no QASPER and option-selection LitQA2 outputs, not free-text responses. Since the central claim is 'highest observed answer quality,' this single confound leaves the answer-accuracy comparison open. Please regrade all systems with at least one out-of-family judge and report per-system agreement/disagreement.
  2. [Section 4.1; Eq. (5)] MOSAIC is constructed and screened automatically by DeepSeek-V4-Pro without human adjudication, and the same model family scores it. Additionally, the composite in Eq. (5) uses SPECTER2 cosine similarity, the same embedding model family used by VecTree-RAG's retrieval (Section 3.2). This creates a circularity risk for the MOSAIC composite and coverage numbers. Expert validation of at least a sample of MOSAIC items, plus a sensitivity analysis of the 0.75/0.25 weights, is needed before the MOSAIC comparison can support the paper's conclusions.
  3. [Section 4.6; Table 5] The LitQA2 ablations are presented as evidence that vector routing and tree navigation matter, but the reported 95% CIs overlap: full system 0.925 [0.85,0.98] vs. −vector 0.811 [0.70,0.91] (overlap 0.85–0.91) and vs. −tree 0.904 [0.83,0.98]. The token-cost differences are large, but no statistical test or error bar is given for them. Please report paired significance tests (e.g., bootstrap over questions with paired seeds) or explicitly restrict the ablation claim to descriptive token-cost observations.
  4. [Table 3 and Section 4.4] The evidence-precision comparison is not apples-to-apples. Baseline chunks inherit every paragraph identifier they span, while VecTree-RAG's page selections are unit choices; this inflates the denominator for baselines and can overstate the 0.274 vs. 0.046–0.071 gap. The paper acknowledges the inheritance rule but still frames the gap as a headline result. Please report a matched comparison (e.g., evidence at chunk level for VecTree's read pages, or precision-recall curves) to support the localization claim.
minor comments (4)
  1. [Appendix E] The title 'Judge Robustness' overstates what is measured; consider 'Limited Judge Diagnostic' or 'Deterministic Agreement on Unambiguous Outputs'.
  2. [Code Availability] Typo: 'Code A vailability' should be 'Code Availability'.
  3. [Section 2] PageIndex is cited via blog posts; if a peer-reviewed or preprint version exists, cite it, otherwise note these are non-archival sources.
  4. [Table 8] The MOSAIC Dense RAG entry uses expanded-context k=232, which is not cost-matched to the other systems; the text discloses this, but a table footnote would prevent misinterpretation.

Circularity Check

0 steps flagged

No significant circularity: the method's predictions are tested against external benchmarks with deterministic evidence metrics; disclosed judge-family and MOSAIC construction issues are validity caveats, not circular derivation.

full rationale

I walked the paper's claimed derivation chain. The central architectural claim—vector retrieval for corpus-level paper discovery and tree navigation for within-paper evidence localization—is evaluated on QASPER (300 questions from the official test split), LitQA2 (an open-access subset), and MOSAIC. QASPER and LitQA2 are external benchmarks; QASPER's deterministic evidence-precision metric (0.274 vs. 0.046–0.071) is computed against gold paragraph identifiers independent of any LLM judge. The LitQA2 ablations report descriptive accuracy and token-cost differences between architecture variants; these are comparisons among the system's own components, not predictions derived from the same fitted values. No parameter is fitted to a target and then renamed as a prediction: the relevance gates (dense ≥0.35, BM25 ≥1.0, etc.) are stated heuristics, and the paper explicitly declines to claim inferential significance for the ablations. The MOSAIC benchmark and the DeepSeek-family judge are acknowledged limitations: Section 4.1 states MOSAIC was 'constructed and screened automatically without independent human validation or adjudication,' and Appendix E admits 'symmetric scoring does not eliminate family or style bias' and that 'Regrading all outputs with multiple out-of-family judges remains the appropriate robustness test.' These are important external-validity concerns, but they are not circularity: the MOSAIC composite is a defined metric (Eq. 5), not a quantity that the method derives from itself, and the paper does not invoke a self-citation chain, an imported uniqueness theorem, or an ansatz-by-citation to force its conclusion. The citation of PageIndex is related work, not load-bearing self-citation (author lists do not overlap). Therefore, under the strict standard of exhibiting a specific reduction to inputs, no circular step is found.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical or mathematical entities. The free parameters are hand-set retrieval gates, fusion constants, and metric weights; the axioms are the key domain assumptions about structure, summary faithfulness, judge reliability, and embedding validity. The summary-faithfulness axiom is the most fragile and is explicitly unverified.

free parameters (6)
  • Pre-fusion relevance gate thresholds = dense ≥0.35 or BM25 ≥1.0
    Hand-set thresholds deciding which documents enter RRF; no tuning or sensitivity analysis is reported (Section 3.4).
  • Post-fusion gate thresholds = 40% of leading RRF (corpus >50 docs); strong-match dense ≥0.55, BM25 ≥4.0
    Hand-set routing heuristics that shape the candidate set presented to the agent (Section 3.4).
  • RRF constant k = 60
    Standard, but still a free choice affecting fusion (Section 3.4, Eq. 4).
  • MOSAIC composite weights = 0.75 factuality-F1 + 0.25 scaled SPECTER cosine
    Weights chosen by the authors for the headline MOSAIC score; changing them changes the ranking (Section 4.3, Eq. 5).
  • Agent limits = 30 iterations, 10 pages/call, 400-token system prompt
    Implementation choices that affect cost/accuracy tradeoffs and the efficiency comparison (Section 3.3).
  • Section-splitting thresholds = 20,000 tokens and 10 pages
    Indexing heuristic that bounds node size; no universal guarantee, affects tree structure (Section 3.2).
axioms (5)
  • domain assumption Scientific papers have informative section structures that can be recovered from PDFs
    The whole method depends on the presence and recoverability of section hierarchy; Section 5 admits it 'may transfer poorly to unstructured documents.'
  • domain assumption Generated section summaries are faithful enough to guide navigation
    Explicitly unverified in Section 5: 'Navigation depends on generated summaries whose faithfulness is not currently verified.' This is load-bearing for the tree-localization claim.
  • domain assumption DeepSeek-V4-Pro judge reliably evaluates DeepSeek-V4-Flash answers
    Headline scores rely on an LLM judge from the same model family as the generator; Appendix E acknowledges possible family bias and that mechanical checks cover only yes/no and MCQ subsets.
  • domain assumption SPECTER2 embeddings are valid for both retrieval and answer-quality scoring
    SPECTER2 is used in the index (Section 3.4) and in the MOSAIC composite metric (Eq. 5); the paper does not validate this dual use.
  • domain assumption QASPER synthetic page-per-paragraph mapping is a fair coordinate system for evidence localization
    QASPER evidence is measured on synthetic page identifiers that equal paragraphs; this favors page-reading agents over chunk-inheritance (Section 4.1).

pith-pipeline@v1.3.0-alltime-deepseek · 13601 in / 10425 out tokens · 94896 ms · 2026-08-01T03:51:11.332780+00:00 · methodology

0 comments
read the original abstract

Scientific question answering requires a retrieval system to solve two distinct problems: identifying which papers are relevant and locating the supporting evidence within those papers. Conventional retrieval-augmented generation typically addresses both through similarity search over fixed-length passages, flattening document structure and separating scientific claims from their methodological and argumentative context. We present VecTree-RAG, an agentic framework that assigns these tasks to complementary retrieval mechanisms. Vector search ranks compact document and section representations across the corpus, whereas reasoning-guided traversal of source-verified section trees localizes evidence within shortlisted papers. Full text is retained in a page store and exposed progressively only after structural localization. We evaluate VecTree-RAG on 300 QASPER questions, an open-access subset of 54 LitQA2 questions, and 49 multi-document MOSAIC questions. Compared with Dense RAG, reranked Dense RAG, RAPTOR, and Search-o1, VecTree-RAG obtained the highest observed answer score on all three benchmarks, reaching 0.800 LLM-judge correctness on QASPER, 0.925 accuracy on LitQA2, and a 0.547 composite score on MOSAIC. On QASPER, its evidence-page precision was 0.274, compared with 0.046--0.071 for the baselines. LitQA2 ablations further showed that the complete vector--tree architecture required fewer inference tokens than variants without tree navigation or corpus-level vector routing. These results indicate that vector retrieval narrows the corpus-level search space and tree navigation concentrates reading on structurally relevant evidence. Although multi-turn inference remains more expensive than single-call retrieval, VecTree-RAG provides a structure-aware and traceable architecture for scientific literature question answering.

Figures

Figures reproduced from arXiv: 2607.23006 by Chen Shen, Tianhang Zhou, Xinyan Zhong, Yuqi Wei, Yuwei Shi, Zhenghao Wu.

Figure 1
Figure 1. Figure 1: Architecture and progressive disclosure in VecTree-RAG. a, [PITH_FULL_IMAGE:figures/full_fig_p004_1.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

46 extracted references · 6 linked inside Pith

  1. [1]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  2. [2]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2020

  3. [3]

    The Probabilistic Relevance Framework:

    Robertson, Stephen and Zaragoza, Hugo , journal =. The Probabilistic Relevance Framework:

  4. [4]

    and Clarke, Charles L

    Cormack, Gordon V. and Clarke, Charles L. A. and Buettcher, Stefan , booktitle =. Reciprocal Rank Fusion Outperforms

  5. [5]

    , booktitle =

    Cohan, Arman and Feldman, Sergey and Beltagy, Iz and Downey, Doug and Weld, Daniel S. , booktitle =

  6. [6]

    Singh, Amanpreet and D'Arcy, Mike and Cohan, Arman and Downey, Doug and Feldman, Sergey , booktitle =

  7. [7]

    , booktitle =

    Sarthi, Parth and Abdullah, Salman and Tuli, Aditi and Khanna, Shubh and Goldie, Anna and Manning, Christopher D. , booktitle =

  8. [9]

    From Local to Global: A Graph

    Edge, Darren and Trinh, Ha and Cheng, Newman and Bradley, Joshua and Chao, Alex and Mody, Apurva and Truitt, Steven and Larson, Jonathan , journal =. From Local to Global: A Graph

  9. [10]

    Guo, Zirui and Xia, Lianghao and Yu, Yanhua and Ao, Tu and Huang, Chao , journal =

  10. [11]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =

  11. [12]

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. Self-

  12. [13]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , pages =

    A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers , author =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , pages =

    2021

  13. [15]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Precise Zero-Shot Dense Retrieval without Relevance Labels , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  14. [16]

    Transactions of the Association for Computational Linguistics , volume =

    Lost in the Middle: How Language Models Use Long Contexts , author =. Transactions of the Association for Computational Linguistics , volume =

  15. [17]

    Marker: Convert

    Paruchuri, Vikas , year =. Marker: Convert

  16. [19]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  17. [20]

    Jiang, Ziyan and Ma, Xueguang and Chen, Wenhu , journal =

  18. [21]

    Khattab, Omar and Zaharia, Matei , booktitle =

  19. [22]

    Findings of the Association for Computational Linguistics: EMNLP , pages =

    Measuring and Narrowing the Compositionality Gap in Language Models , author =. Findings of the Association for Computational Linguistics: EMNLP , pages =

  20. [25]

    2025 , note =

    Zhang, Mingtian and Tang, Yu and. 2025 , note =

    2025

  21. [26]

    2026 , month = may, note =

    2026

  22. [27]

    2024 , note =

    Xiao, Shitao and Liu, Zheng and Zhang, Peitian and Muennighoff, Niklas and Lian, Defu and Nie, Jian-Yun , journal =. 2024 , note =

    2024

  23. [28]

    Self- RAG : Learning to retrieve, generate, and critique through self-reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self- RAG : Learning to retrieve, generate, and critique through self-reflection. In International Conference on Learning Representations (ICLR), 2024

    2024

  24. [29]

    Walking down the memory maze: Beyond context limit through interactive reading

    Howard Chen, Ramakanth Pasunuru, Jason Weston, and Asli Celikyilmaz. Walking down the memory maze: Beyond context limit through interactive reading. arXiv preprint arXiv:2310.05029, 2023

    Pith/arXiv arXiv 2023

  25. [30]

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S. Weld. SPECTER : Document-level representation learning using citation-informed transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 2270--2282, 2020

    2020

  26. [31]

    Cormack, Charles L

    Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. Reciprocal rank fusion outperforms Condorcet and individual rank learning methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 758--759, 2009

    2009

  27. [32]

    Smith, and Matt Gardner

    Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A. Smith, and Matt Gardner. A dataset of information-seeking questions and answers anchored in research papers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4599--4610, 2021

    2021

  28. [33]

    Precise zero-shot dense retrieval without relevance labels

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 1762--1777, 2023 a

    2023

  29. [34]

    Retrieval-augmented generation for large language models: A survey

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

    Pith/arXiv arXiv 2023

  30. [35]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769--6781, 2020

    2020

  31. [36]

    ColBERT : Efficient and effective passage search via contextualized late interaction over BERT

    Omar Khattab and Matei Zaharia. ColBERT : Efficient and effective passage search via contextualized late interaction over BERT . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 39--48, 2020

    2020

  32. [37]

    Rodriques, and Andrew D

    Jakub L \'a la, Odhran O'Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G. Rodriques, and Andrew D. White. PaperQA : Retrieval-augmented generative agent for scientific research. arXiv preprint arXiv:2312.07559, 2023

    Pith/arXiv arXiv 2023

  33. [38]

    Laurent, Joseph D

    Jon M. Laurent, Joseph D. Janizek, Michael Ruzo, Michaela M. Hinks, Michael J. Hammerling, Siddharth Narayanan, Manvitha Ponnapati, Andrew D. White, and Samuel G. Rodriques. Lab-bench: Measuring capabilities of language models for biology research. arXiv preprint arXiv:2407.10362, 2024

    Pith/arXiv arXiv 2024

  34. [39]

    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, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 9459--9474, 2020

    2020

  35. [40]

    Search-o1: Agentic search-enhanced large reasoning models

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models. arXiv preprint arXiv:2501.05366, 2025

    Pith/arXiv arXiv 2025

  36. [41]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12: 0 157--173, 2024

    2024

  37. [42]

    PageIndex File System : Massive-scale document search

    PageIndex Team . PageIndex File System : Massive-scale document search. PageIndex Blog, May 2026. https://pageindex.ai/blog/pageindex-filesystem

    2026

  38. [43]

    Marker: Convert PDF to markdown quickly with high accuracy

    Vikas Paruchuri. Marker: Convert PDF to markdown quickly with high accuracy. https://github.com/VikParuchuri/marker, 2024

    2024

  39. [44]

    Smith, and Mike Lewis

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models. Findings of the Association for Computational Linguistics: EMNLP, pages 5687--5711, 2023

    2023

  40. [45]

    The probabilistic relevance framework: BM25 and beyond

    Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3 0 (4): 0 333--389, 2009

    2009

  41. [46]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D. Manning. RAPTOR : Recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations (ICLR), 2024

    2024

  42. [47]

    SciRepEval : A multi-format benchmark for scientific document representations

    Amanpreet Singh, Mike D'Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. SciRepEval : A multi-format benchmark for scientific document representations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5548--5566, 2023

    2023

  43. [48]

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 10014--10037, 2023

    2023

  44. [49]

    C-Pack : Packed resources for general chinese embeddings

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. C-Pack : Packed resources for general chinese embeddings. arXiv preprint arXiv:2309.07597, 2024. BGE reranker (bge-reranker-v2-m3)

    Pith/arXiv arXiv 2024

  45. [50]

    ReAct : Synergizing reasoning and acting in language models

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

    2023

  46. [51]

    PageIndex : Next-generation vectorless, reasoning-based RAG

    Mingtian Zhang, Yu Tang, and PageIndex Team . PageIndex : Next-generation vectorless, reasoning-based RAG . PageIndex Blog, 2025. https://pageindex.ai/blog/pageindex-intro

    2025