Pith. sign in

REVIEW 5 major objections 6 minor 30 references

KeyKnowledgeRAG (K^2RAG): An Enhanced RAG method for improved LLM question-answering capabilities

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read KeyKnowledgeRAG combines summarization, knowledge graphs, and hybrid search to beat naive RAG pipelines on both accuracy and efficiency.

desk verdict A clear-headed engineering report on a RAG pipeline that combines known pieces; the evaluation is too weak to support the headline numbers, but the paper is honest and deserves a real referee. read the letter →

arxiv 2507.07695 v2 pith:VS5OSWCC submitted 2025-07-10 cs.CL cs.AI

Hruday Markondapatnaikuni , Basem Suleiman , Abdelkarim Erradi , Shijing Chen This is my paper
classification cs.CLcs.AI
keywords retrieval-augmentedgenerationknowledgegraphshybridretrievaltextsummarizationquestionansweringquantizedlargelanguagemodelsMultiHopRAGresourceefficiency
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

KeyKnowledgeRAG (K2RAG) is a retrieval-augmented generation pipeline that tries to show that answer accuracy and resource efficiency can be improved together rather than traded off. Its core claim is that summarising the document corpus before building the sparse and dense vector stores and the knowledge graph cuts information-store creation time by 93% on average, while a hybrid retriever with $\lambda=0.8$ (80% dense, 20% sparse), knowledge-graph topic expansion, and sub-question decomposition improve the relevance of what the generator sees. On the MultiHopRAG benchmark, K2RAG reports the best mean answer cosine similarity (0.57) and the best third-quartile similarity (0.82) among the five pipelines tested, while running in about 5GB of VRAM versus 14.3–18.1GB for the naive pipelines and answering faster than naive knowledge-graph RAG. The paper itself notes in Section 5.5 that the mean gain is modest and the advantage concentrates in certain question types, so the central claim is specifically that the combination helps on a substantial share of questions, not on all of them.

What carries the argument

The load-bearing mechanism is a two-stage retrieve-and-condense loop. First the full corpus is replaced by summaries produced by a Longformer-based summarizer; those summaries are chunked with Eq. (1) and indexed into a GraphRAG-style knowledge graph (300-token chunks, 100-token overlap) and into dense and sparse vector stores (256-token chunks, 20-token overlap). At query time, the knowledge graph returns topic-relevant passages, the Longformer summarizer condenses them, and the condensed result is split into 128-token chunks; a quantized 7B LLM turns each chunk into a sub-question. Each sub-question retrieves the top-$k$ chunks from the vector stores using the hybrid score $\mathrm{scr}(q,c,\lambda) = \lambda\,d(q,c) + (1-\lambda)\,s(q,c)$ with $\lambda = 0.8$, the sub-answers are summarized again, and the final answer is generated from the concatenated summaries. The intended effect is that every context given to the generator stays short and topic-focused, which is how the paper claims to avoid the needle-in-a-haystack failure mode.

What would settle it

Measure retrieval recall on the summarized corpus: for each MultiHopRAG question, check whether the gold evidence passages appear in the top-10 hybrid-retrieval results computed from summaries; if recall falls sharply relative to full-corpus retrieval while answer similarity stays at 0.57, the claimed retrieval-quality improvement is not what drives the answer scores. A cleaner ablation would remove the knowledge-graph and sub-question steps and see whether Q3 drops from 0.82 toward the ~0.67 of the naive pipelines.

Watch

Extended reading notes

Core claim

The paper's central claim is that four ingredients combined in one pipeline — corpus-level summarisation with a Longformer-based book-summary model, a knowledge graph built on those summaries, hybrid dense-plus-sparse retrieval weighted at $\lambda = 0.8$, and a quantised 7B LLM that generates sub-questions and final answers — give a RAG system that is simultaneously more accurate and cheaper to build and run than four naive single-store pipelines. The evidence is the MultiHopRAG evaluation: K2RAG reaches a mean answer cosine similarity of 0.57 and a third-quartile similarity of 0.82, compared with means of 0.54–0.56 and Q3 values of 0.63–0.67 for the naive baselines. Corpus summarisation (about 89% average size reduction, taking roughly 25 minutes) cuts dense-store creation from 4010 seconds to 441 seconds and knowledge-graph creation from 64,785 seconds to 3,915 seconds, and the pipeline's total VRAM footprint is about 5GB versus 14.3–18.1GB. The paper presents this as evidence that a deliberately lightweight, summary-first pipeline can beat naive RAG on answer alignment while removing the main scalability bottlenecks.

Load-bearing premise

The whole efficiency gain rests on the assumption that shrinking each article to roughly 11% of its original size still leaves enough of the answer-bearing facts intact for the indexes to retrieve what a question needs; if an 89% reduction drops those facts, the training-time advantage comes at the cost of answer quality.

Editorial extensions

If this is right

  • Information-store updates become practical in production: knowledge-graph refresh drops from roughly 18 hours to about 1 hour, and dense-vector creation from 67 minutes to about 7 minutes, after a one-time 25-minute summarization pass.
  • Deployment moves to smaller hardware: the full pipeline fits in roughly 5GB of VRAM, about a threefold reduction from the 14.3–18.1GB needed by the compared naive pipelines.
  • A quantized 7B generator plus a small summarizer can replace a full-precision inference model without sacrificing answer alignment on MultiHopRAG, lowering the cost of hosting RAG systems.
  • The Q3 similarity of 0.82 means that for the upper quarter of questions the system's answers land close to the ground truth, which is the outcome that matters in practice even when the mean gain is small.

Reading between the lines

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

  • Ablating the knowledge-graph and sub-question stages separately would show whether the Q3 jump comes from graph-driven topic decomposition or from summarization alone; the paper does not isolate these factors.
  • Because retrieval recall on the summarized corpus is never reported, the 93% speed-up could in principle come at the cost of losing answer-bearing facts; a recall@10 comparison on the 609-article corpus would settle this.
  • The modest mean gain suggests K2RAG's real advantage may be on multi-hop questions whose answers straddle several documents; a per-category breakdown (bridge, comparison, temporal) would test this directly.
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 K2RAG, a retrieval-augmented generation framework that combines a knowledge graph, hybrid dense/sparse retrieval, Longformer-based summarization at indexing and query time, and a quantized Mistral-7B generator. The method is evaluated on the MultiHopRAG dataset against four naive RAG pipelines (semantic, keyword, hybrid, and knowledge-graph-only). The reported results are a mean answer cosine similarity of 0.57 versus 0.54–0.56 for the baselines, a third-quartile similarity of 0.82 versus 0.63–0.67, an average reduction in information-store training time of 93%, and a memory footprint of 5 GB versus 14.3–18.1 GB for the baselines.

Significance. If the efficiency results are substantiated, the paper makes a practically useful contribution: summarizing the corpus before indexing and using a quantized generator plausibly reduce indexing time and memory footprint, and the modular pipeline is clearly described with concrete research goals. However, the central accuracy claim is not established by the evidence presented. The reported accuracy advantage is a small difference in a single soft metric, with no statistical testing, no exact-match or token-level F1 scores, and no retrieval-recall measurement on the summarized corpus. The paper's own limitation section, §5.5, describes the accuracy gain as 'slight' and limited to certain question types, which contradicts the abstract's language of 'notable improvements.' The two internal numerical inconsistencies identified below further reduce confidence in the quantitative record. The efficiency claims are therefore not yet anchored to a demonstrated preservation of question-answering capability.

major comments (5)
  1. [§4.4, Table 2, Figure 13] Table 2 lists the Naive Knowledge Graph Search Q3 value as 0.4, while Figure 13 shows the same pipeline with Q3 = 0.63. These two exhibits are the sole evidence for the accuracy comparison, so this discrepancy directly undermines every reported accuracy value and must be resolved before the comparison can be trusted.
  2. [§3.3.2, §4.4, §5.5] Answer quality is evaluated exclusively with cosine similarity (Eq. 6), and no error bars, confidence intervals, or significance tests are reported. The mean differences (0.57 versus 0.54–0.56 on 2,555 questions) are small enough to be within sampling noise, and the Q3 advantage is explicitly attributed by the authors to a limited subset of question types in §5.5. The paper should report exact match, token-level F1, and a paired significance test (for example, a bootstrap or Wilcoxon test), and should align the abstract's 'notable improvements' with the more cautious language in §5.5.
  3. [§3.1.1, §4.3] The 93% training-time reduction depends on the assumption that the 89%-compressed summaries preserve the facts needed to answer MultiHopRAG questions, but the paper never measures retrieval recall on the summarized corpus or compares per-question answerability against the full corpus. The only validation is the same weak cosine-similarity metric. Without a retrieval-recall or passage-level evaluation, the efficiency gain cannot be interpreted as preserving QA capability.
  4. [§4.6, §5.4] The VRAM comparison is confounded: K2RAG uses Mistral-7B-Q4 for generation, while the naive pipelines use an unquantized Mistral-7B-FP16. The reported 14.3 GB versus 5 GB difference is therefore mostly attributable to model quantization, not to the K2RAG architecture. To support the scalability claim, the comparison should hold the generator precision fixed or provide a per-component memory breakdown.
  5. [§4.3, Table 1, Figure 12] There is a second internal inconsistency in the training-time results: Table 1 reports Sparse Vector Summarized Corpus as 12 s, while Figure 12 shows 27 s. Because the training-time reduction is a central efficiency claim, this discrepancy needs to be corrected and explained.
minor comments (6)
  1. [§1, Goal 1] The phrase 'Spare and Dense vector database creation times' should be 'Sparse and Dense'; the same typo appears elsewhere in the paper.
  2. [§2] 'Research Research Goal 1' and 'Research Research Goal 3' appear in the discussion of Blended RAG; these should be corrected.
  3. [§3.1.3, Eq. (4)] The weights λ = 0.8 and 0.2 are described in §3.1 as 'optimized,' but they are taken from Mandikal et al. [18] without tuning on MultiHopRAG. Please state explicitly that these values are inherited from prior work, not optimized on this dataset.
  4. [§3.1.1] The summarizer is referred to as a Longformer-based model and cited to LongT5 [29], but the text names pszemraj/led-base-book-summary, which is a Longformer Encoder-Decoder (LED) model. Please cite the correct source or clarify the model architecture.
  5. [§4.4] The four baseline distributions in Figure 13 are described as 'very similar,' but the panels use different scales; plotting all pipelines on a common scale or as overlaid densities would make the comparison easier to read.
  6. [Overall] The paper does not mention code availability; releasing the evaluation scripts and configuration files would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: K2RAG's accuracy, speed, and memory claims rest on direct measurements and externally sourced parameters, not on self-referential derivations.

full rationale

The paper's central claims are not circular. The hybrid weighting parameter λ=0.8 is explicitly borrowed from Mandikal et al. [18], an external prior work, and is applied uniformly to K2RAG and the naive hybrid pipeline; it is not fitted to MultiHopRAG and then 'predicted' from it. The evaluation metric (Eq. 6) computes cosine similarity between the generated answer and the ground-truth answer, while the retrieval scoring function (Eq. 4) scores question–chunk relevance; these are different quantities, and nothing in Eq. 6 is an input to retrieval, summarization, or generation. Thus the reported 0.57 mean and 0.82 Q3 are measured outcomes, not identities. Likewise, the 93% training-time reduction and 3× VRAM reduction are direct comparisons of measured indexing times and peak memory use, not derivations that assume the accuracy result. The paper contains no self-citations that bear on the framework's justification; all cited components (GraphRAG, Longformer summarizer, MultiHopRAG, Mandikal et al.) are external. The internal inconsistency between Table 2 and Figure 13 for the naive KG Q3 value, and the paper's own admission in Sec. 5.5 of only a 'slight mean accuracy advantage,' are validity and reporting concerns, not circular reasoning. No fitted parameter is renamed as a prediction, no uniqueness claim is imported from the authors' prior work, and no known result is merely re-labeled. Under the stated standard, the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claims rest on a single transferred hyperparameter (lambda), an unvalidated assumption about summarization preserving answer-relevant facts, and an embedding-similarity evaluation metric. No new entities are postulated.

free parameters (1)
  • Hybrid retrieval weighting lambda = 0.8
    Section 3.1.3: hybrid scoring in Equation (4); the weight is taken from Mandikal et al. and not fit to the MultiHopRAG data, so it is a transferred hyperparameter rather than a fitted value.
assumptions (2)
  • domain assumption The Longformer summarizer preserves corpus facts needed to answer MultiHopRAG questions
    Section 4.4: the summarized corpus is used for all indexes, but the paper does not measure retrieval recall on the summaries, so the fact-preservation premise is unvalidated.
  • domain assumption Cosine similarity between the generated answer and ground truth is a valid proxy for answer accuracy
    Section 3.3.2: evaluation metric defined in Equation (6). The paper relies on this metric and does not report exact-match, F1, or human-judged correctness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KeyKnowledgeRAG (K^2RAG): An Enhanced RAG method for improved LLM question-answering capabilities." pith.science (2026). https://pith.science/paper/VS5OSWCC

@misc{pith2026250707695,
  author       = {Pith},
  title        = {Pith review of: KeyKnowledgeRAG (K^2RAG): An Enhanced RAG method for improved LLM question-answering capabilities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VS5OSWCC}},
  note         = {Machine review of arXiv:2507.07695}
}
read the original abstract

Fine-tuning is an immensely resource-intensive process when retraining Large Language Models (LLMs) to incorporate a larger body of knowledge. Although many fine-tuning techniques have been developed to reduce the time and computational cost involved, the challenge persists as LLMs continue to grow in size and complexity. To address this, a new approach to knowledge expansion in LLMs is needed. Retrieval-Augmented Generation (RAG) offers one such alternative by storing external knowledge in a database and retrieving relevant chunks to support question answering. However, naive implementations of RAG face significant limitations in scalability and answer accuracy. This paper introduces KeyKnowledgeRAG (K2RAG), a novel framework designed to overcome these limitations. Inspired by the divide-and-conquer paradigm, K2RAG integrates dense and sparse vector search, knowledge graphs, and text summarization to improve retrieval quality and system efficiency. The framework also includes a preprocessing step that summarizes the training data, significantly reducing the training time. K2RAG was evaluated using the MultiHopRAG dataset, where the proposed pipeline was trained on the document corpus and tested on a separate evaluation set. Results demonstrated notable improvements over common naive RAG implementations. K2RAG achieved the highest mean answer similarity score of 0.57, and reached the highest third quartile (Q3) similarity of 0.82, indicating better alignment with ground-truth answers. In addition to improved accuracy, the framework proved highly efficient. The summarization step reduced the average training time of individual components by 93%, and execution speed was up to 40% faster than traditional knowledge graph-based RAG systems. K2RAG also demonstrated superior scalability, requiring three times less VRAM than several naive RAG implementations tested in this study.

Figures

Figures reproduced from arXiv: 2507.07695 by the authors.

Figure 1
Figure 1. Evolution of RAG since 2020 [5]. A class of techniques known as Retrieval Augment Generate (RAG) systems have been devised as an alternative to fine-tuning where they do not employ any step where a base LLM is modified as in the fine-tuning method but rather knowledge is extracted from documents and indexed in a knowledge holding system to then retrieve [6]. This helps eliminate the computationally expensive associa… view at source ↗
Figure 2
Figure 2. Graph showing how LLMs prefer extracting information from top or bottom of context to answer a question [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Results of Priyanka Mandikal et al. on 2 example queries [18]. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: A high level view of what a RAG pipeline with a reranker looks like. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Results comparing knowledge graph framework with semantic search methods [25]. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: How to generate summaries of corpus. 3.1.2 Indexing The chunking strategy to generate a set of chunks Cd from a single document or text d is defined as: 7 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Chunk sizes and their effects on Answer Faithfulness and Answer Relevancy. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: K2RAG framework overview. Component (a): Knowledge Graph. Component (b): Hybrid Retriever containing Dense and Sparse retrievers. Component (c): Embeddings Model. Step (A): Knowledge Graph results. Step (B): Summarise Knowledge Graph results. Step (C): Generating sub-q…
Figure 9
Figure 9. Figure 9: Structure of competitor pipelines. Component (S): Semantic Search - Dense Vector Retrieval. Component [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Evaluation Pipeline. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Boxplot of corpus size reduction. 4.3 Training Times By performing corpus summarization, sparse and dense vector stores, and knowledge graph creation times reduced by 89%, 97% and 94% respectively with significant time saved for knowledge graph creation from 18 hours …
Figure 12
Figure 12. Figure 12: Information store component training times with summarized and unsummarized corpus. Training on the [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Pipeline Answer Accuracy results. K2RAG achieved the best performance with highest mean and 3rd Quartile value. Pipeline Q1 Mean Q3 Naive Semantic Search 0.40 0.55 0.67 Naive Keyword Search 0.41 0.55 0.67 Naive Hybrid Search 0.4 0.56 0.67 Naive Knowledge Graph Search …
Figure 14
Figure 14. Figure 14: Pipeline Execution Times results. K2RAG is faster than KG Search despite more steps. Pipeline Mean Execution Time Naive Semantic Search 3.31s Naive Keyword Search 3.29s Naive Hybrid Search 3.92s Naive Knowledge Graph Search 117.31s K2RAG 70.25s [PITH_FULL_IMAGE:figur…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages

  1. [1]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M

    Jason Wei, Maarten Bosma, Vincent Y . Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V . Le. Finetuned language models are zero-shot learners, 2022

    2022

  2. [2]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021

    2021

  3. [3]

    Qlora: Efficient finetuning of quantized llms, 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms, 2023

    2023

  4. [4]

    An empirical study of llama3 quantization: From llms to mllms, 2024

    Wei Huang, Xingyu Zheng, Xudong Ma, Haotong Qin, Chengtao Lv, Hong Chen, Jie Luo, Xiaojuan Qi, Xianglong Liu, and Michele Magno. An empirical study of llama3 quantization: From llms to mllms, 2024

    work page 2024

  5. [5]

    Retrieval-augmented generation for large language models: A survey, 2024

    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, 2024

    2024

  6. [6]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    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. Retrieval-augmented generation for knowledge-intensive nlp tasks. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,Advances in Neural Informa...

    2020

  7. [7]

    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, 2023

    2023

  8. [8]

    Fabbri, Caiming Xiong, and Chien-Sheng Wu

    Philippe Laban, Alexander R. Fabbri, Caiming Xiong, and Chien-Sheng Wu. Summary of a haystack: A challenge to long-context llms and rag systems, 2024

    work page 2024

Show all 30 references
  1. [9]

    Searching for best practices in retrieval-augmented generation, 2024

    Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, Ruicheng Yin, Changze Lv, Xiaoqing Zheng, and Xuanjing Huang. Searching for best practices in retrieval-augmented generation, 2024

    2024

  2. [10]

    Relevance filtering for embedding-based retrieval

    Nicholas Rossi, Juexin Lin, Feng Liu, Zhen Yang, Tony Lee, Alessandro Magnani, and Ciya Liao. Relevance filtering for embedding-based retrieval. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM ’24, page 4828–4835, New York, ...

    2024

  3. [11]

    Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia D’amato, Gerard De Melo, Claudio Gutierrez, Sab- rina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel-Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequ...

    2021

  4. [12]

    Blended rag: Improving rag (retriever-augmented generation) accuracy with semantic search and hybrid query-based retrievers

    Kunal Sawarkar, Abhilasha Mangal, and Shivam Raj Solanki. Blended rag: Improving rag (retriever-augmented generation) accuracy with semantic search and hybrid query-based retrievers. In2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (M...

    2024

  5. [13]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019

  6. [14]

    Murdick, Devvret Rishi, Jerry Sheehan, Zhihong Shen, Brandon Stilson, Alex D

    Lucy Lu Wang, Kyle Lo, Yoganand Chandrasekhar, Russell Reas, Jiangjiang Yang, Doug Burdick, Darrin Eide, Kathryn Funk, Yannis Katsis, Rodney Michael Kinney, Yunyao Li, Ziyang Liu, William Merrill, Paul Mooney, Dewey A. Murdick, Devvret Rishi, Jerry Sheehan, Zhihong Shen, Brand...

    2020

  7. [15]

    Squad: 100,000+ questions for machine comprehension of text, 2016

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text, 2016

    2016

  8. [16]

    Cohen, Ruslan Salakhutdinov, and Christo- pher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christo- pher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018

    2018

  9. [17]

    Siva Reddy, Danqi Chen, and Christopher D. Manning. CoQA: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249–266, 2019

    2019

  10. [18]

    Sparse meets dense: A hybrid approach to enhance scientific document retrieval.CoRR, abs/2401.04055, 2024

    Priyanka Mandikal and Raymond Mooney. Sparse meets dense: A hybrid approach to enhance scientific document retrieval.CoRR, abs/2401.04055, 2024

    2024 arXiv

  11. [19]

    Scirepeval: A multi-format benchmark for scientific document representations, 2023

    Amanpreet Singh, Mike D’Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. Scirepeval: A multi-format benchmark for scientific document representations, 2023

    2023

  12. [20]

    Document ranking with a pretrained sequence- to-sequence model

    Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. Document ranking with a pretrained sequence- to-sequence model. In Trevor Cohn, Yulan He, and Yang Liu, editors,Findings of the Association for Compu- tational Linguistics: EMNLP 2020, pages 708–718, Online, Novemb...

    2020

  13. [21]

    Re2G: Retrieve, rerank, generate

    Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Naik, Pengshan Cai, and Alfio Gliozzo. Re2G: Retrieve, rerank, generate. In Marine Carpuat, Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz, editors,Proceedings of the 2022 Conference of the North Am...

    2022

  14. [22]

    Fine-tuning llama for multi-stage text retrieval

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. Fine-tuning llama for multi-stage text retrieval. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’24, page 2421–2425, New York, NY , USA, 2024....

    2024

  15. [23]

    Financial time series forecasting with multi-modality graph neural network.Pattern Recognition, 121:108218, 2022

    Dawei Cheng, Fangzhou Yang, Sheng Xiang, and Jin Liu. Financial time series forecasting with multi-modality graph neural network.Pattern Recognition, 121:108218, 2022

    2022

  16. [24]

    Retrieval-augmented generation with knowledge graphs for customer service question answering

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. Retrieval-augmented generation with knowledge graphs for customer service question answering. InProceedings of the 47th International ACM SIGIR Conference on Research and De...

    2024

  17. [25]

    Kg-rag: Bridging the gap between knowledge and creativity, 2024

    Diego Sanmartin. Kg-rag: Bridging the gap between knowledge and creativity, 2024

    2024

  18. [26]

    Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction

    Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In Proceedings of the 5th ACM International Conference on AI in ...

    2024

  19. [27]

    From local to global: A graph rag approach to query-focused summarization, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization, 2024

    2024

  20. [28]

    Ragas: Automated evaluation of retrieval augmented generation, 2023

    Shahul Es, Jithin James, Luis Espinosa-Anke, and Steven Schockaert. Ragas: Automated evaluation of retrieval augmented generation, 2023

    2023

  21. [29]

    LongT5: Efficient text-to-text transformer for long sequences

    Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, and Yinfei Yang. LongT5: Efficient text-to-text transformer for long sequences. In Marine Carpuat, Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz, editors,Findings of the Association...

    2022

  22. [30]

    Multihop-RAG: Benchmarking retrieval-augmented generation for multi-hop queries

    Yixuan Tang and Yi Yang. Multihop-RAG: Benchmarking retrieval-augmented generation for multi-hop queries. InFirst Conference on Language Modeling, 2024. 19

    2024

Pith tools

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