Pith. sign in

REVIEW 3 major objections 6 minor 44 references

KET-RAG: A Cost-Efficient Multi-Granular Indexing Framework for Graph-RAG

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read KET-RAG claims full-Graph-RAG retrieval quality at about one-tenth the indexing cost, by extracting a knowledge-graph skeleton from only the most central chunks and pairing it with a lightweight text-keyword graph.

desk verdict KET-RAG is a genuinely useful hybrid of a PageRank-pruned KG and a keyword-text graph, but the headline cost claim only holds under a favorable comparison to a reimplemented Graph-RAG baseline. read the letter →

arxiv 2502.09304 v2 pith:JCSTYNNR submitted 2025-02-13 cs.IR

classification cs.IR
keywords Retrieval-augmentedgenerationGraph-RAGKnowledgegraphskeletonText-keywordbipartitePageRankMulti-hopquestionansweringIndexingcost
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 KET-RAG, an indexing scheme that builds two complementary structures instead of a full knowledge graph: a knowledge-graph skeleton from a small set of PageRank-selected core chunks, and a text-keyword bipartite graph over all chunks. The claim is that this two-channel index gives retrieval quality comparable to or better than Microsoft's Graph-RAG while reducing indexing cost by over an order of magnitude, and raises generation quality by up to 32.4% while cutting cost by about 20%. If true, multi-hop question answering over large proprietary document collections becomes affordable without sacrificing answer quality.

What carries the argument

The load-bearing object is the two-part KET index $G = G_s \cup G_k$. $G_s$ is a knowledge-graph skeleton produced by running LLM triplet extraction only on the $\lceil \beta |V| \rceil$ text chunks with highest PageRank in an intermediate KNN graph, and $G_k$ is a text-keyword bipartite graph in which each keyword node's description is the concatenation of all sentences containing that keyword, with embedding equal to the average of those sentence embeddings. Retrieval runs the Graph-RAG local-search procedure on $G_s$ with $\theta \lambda$ tokens and a keyword-seeded neighbor retrieval on $G_k$ with the remaining $(1-\theta)\lambda$ tokens, then concatenates the two contexts.

What would settle it

Construct a corpus where answer-bearing sentences are deliberately placed in low-PageRank chunks, for example by cutting lexical and semantic links between those chunks and the rest of the corpus, and compare KET-RAG-P coverage with KET-RAG-U and full Graph-RAG; if coverage falls to the random-selection level, core-chunk selection is not doing the work.

Watch

Extended reading notes

Core claim

The paper argues that a full triplet-level knowledge graph is not necessary for good Graph-RAG behavior. Instead, it claims, LLM triplet extraction can be confined to a PageRank-selected fraction of core chunks, while a keyword-to-text bipartite graph supplies the lightweight memory for everything else. On three multi-hop QA datasets, this combination matches or improves retrieval coverage and generation metrics compared with full graph indexing at a fraction of the cost, with PageRank-based core selection consistently beating random selection.

Load-bearing premise

The framework's quality rests on the PageRank-selected core chunks containing the query-critical facts; if those facts live in low-centrality chunks, the skeleton contributes little and retrieval leans on the keyword channel alone.

Editorial extensions

If this is right

  • LLM extraction cost scales with the $\beta$ fraction of chunks rather than the whole corpus, so the absolute savings grow as document collections grow.
  • The keyword channel alone can replace plain text retrieval: on the tested datasets it outperforms Text-RAG on coverage, EM, F1, and BERTScore with no added retrieval latency.
  • A small $G_s$ share of the context budget already lifts quality over Keyword-RAG, so practitioners can slide $\theta$ to spend more or less of the context on the skeleton depending on their cost-accuracy target.
  • At $\beta=1$ and $\theta=1$ the framework reduces to the full Graph-RAG local-search configuration, making the reported results a controlled study of how much quality survives when most triplets are skipped.

Reading between the lines

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

  • The paper selects core chunks once, statically, by PageRank; a natural untested extension is query-aware selection, so chunks that would not be central in the whole corpus can still be extracted when they matter for a query.
  • Because a keyword node's embedding averages every sentence containing the keyword, a polysemous word receives one blended vector; an LLM-generated sense-level description per keyword might sharpen the cosine search, though the paper does not explore this.
  • The reported $\beta$ sensitivity suggests that on corpora with degree distributions even more skewed than the three tested datasets, the budget could be lowered below the default 0.8; that is an extrapolation from the paper's experiments, not a tested claim.
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

3 major / 6 minor

Summary. The paper proposes KET-RAG, a multi-granular indexing framework for Graph-RAG. Instead of extracting triplets from every text chunk, KET-RAG selects a PageRank-central subset of core chunks, builds a knowledge graph skeleton from them, and augments this with a text-keyword bipartite graph constructed from all chunks. Retrieval combines local search on the skeleton with keyword-anchored retrieval on the bipartite graph. The authors evaluate 13 solutions on MuSiQue, HotpotQA, and RAG-QA Arena, reporting that KET-RAG achieves better or comparable retrieval quality than Microsoft's Graph-RAG at a fraction of the indexing cost, and up to 32.4% improvement in generation quality while lowering indexing costs by about 20% relative to Hybrid-RAG.

Significance. If the main empirical finding holds, the paper's central claim is practically significant: a partial knowledge graph skeleton plus a keyword-text bipartite graph can largely substitute for full LLM-based knowledge graph indexing on multi-hop QA benchmarks, at a large cost saving. The paper has notable strengths: a transparent cost model in Section 5.3, a clean decomposition into Skeleton-RAG and Keyword-RAG ablations, comparisons against several external baselines, and a public code release. However, the headline claims are currently weakened by an internal contradiction in the abstract and by comparing the proposed system at a cheap operating point against a baseline at a very expensive operating point, with the baseline being an in-house simplified reimplementation rather than the official Microsoft Graph-RAG. These issues are fixable but are load-bearing for the paper's central message.

major comments (3)
  1. [Abstract and Section 6.2 (Table 2)] The abstract states that KET-RAG "outperforms all competitors in indexing cost, retrieval effectiveness, and generation quality." This is contradicted by Table 2, where Text-RAG, KNNG-RAG, and HyDE have indexing costs of USD 0.01-0.05 on all datasets, while KET-RAG-P costs USD 1.87-4.08. The claim is only defensible among graph-indexing competitors, not absolute. Please either restrict the wording to the Pareto frontier among graph-based methods or report a comparison that explicitly accounts for quality-matched operating points.
  2. [Section 6.2, Tables 2-3] The "over an order of magnitude" cost reduction claim is based on comparing the low-cost KET-RAG-P row (e.g., USD 1.89 on MuSiQue in Table 2) with the high-accuracy MS-Graph-RAG row (USD 24.94 on MuSiQue in Table 3). This crosses configurations: the same low-cost table shows MS-Graph-RAG at USD 2.30, which represents only a 1.2x cost difference. To substantiate an intrinsic cost-quality advantage, please report each baseline at multiple chunk sizes (e.g., ℓ=1200, 600, 300, 150) and compare at matched cost or matched quality, rather than selecting the least favorable baseline operating point.
  3. [Section 6.1] The baseline labeled "MS-Graph-RAG" is not the official Microsoft Graph-RAG implementation. As stated in Section 6.1, it is obtained by setting β=1 and θ=1 within the KET-RAG framework, i.e., a simplified local-search variant that omits community detection and the official prompt pipeline. Since the abstract and Section 7 claim superiority to "Microsoft's Graph-RAG," this is an external-validity concern. Please either run the official implementation or rename the baseline and temper the claims accordingly.
minor comments (6)
  1. [Section 1] Typo: "enbales" should be "enables."
  2. [Section 5.2] Typo: "spitting" should be "splitting."
  3. [Section 6.4] The sentence "particularly when β∈[0.6,0.8] in MuSiQue and β∈[0.2,0.4] in MuSiQue" repeats MuSiQue; one occurrence likely should be HotpotQA.
  4. [Section 6.5] The dataset name is written as "Musique" (e.g., "we take the Musique dataset"); it should be "MuSiQue" for consistency with the rest of the paper.
  5. [Section 5.3] The equation for ITCKET-Index includes only LLM input and embedding costs; the prose separately mentions output token costs but does not integrate them into the equation. Please clarify the scope of the formulas to avoid confusion.
  6. [Table 3 footnote] The exclusion of LightRAG on MuSiQue due to "unexpected behavior" is reported only as a table footnote; this limitation should be acknowledged and discussed in the main text, since it removes a competitor from the high-accuracy comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KET-RAG's headline claims are empirical measurements against external and reimplemented baselines, and its cost equations are accounting identities rather than predictions derived from fitted parameters.

full rationale

KET-RAG does not derive its headline results from assumptions that already contain them. The retrieval-quality and generation-quality claims (Abstract; Tables 2-3; Section 6.2) are measurements from three benchmark datasets, not predictions of a model fitted to those datasets. The indexing-cost equations in Section 5.3 are cost-accounting definitions: the statement that Skeleton-RAG with beta=0.8 reduces KG-Index cost by 20% is true by construction and is presented as such, not as an empirical discovery. The 'order of magnitude' cost reduction compares KET-RAG-P in the low-cost configuration with MS-Graph-RAG in the high-accuracy configuration (Tables 2-3); this is a comparison-design choice rather than a circular derivation. The MS-Graph-RAG baseline is implemented as a special case of the KET-RAG framework (Section 6.1: 'KET-RAG simplifies to MS-Graph-RAG by setting beta = 1 and theta = 1'), which raises a legitimate external-validity question about whether the reimplementation faithfully matches Microsoft's official Graph-RAG, but it does not make the comparison circular because the reported qualities are measured rather than implied by the definitions. There are no self-citations, no imported uniqueness theorems, no fitted parameters renamed as predictions, and no ansatz smuggled in via citation. Parameter choices are accompanied by sensitivity analyses (Section 6.5, Figures 3-4). The only concerns are baseline fidelity and operating-point selection, which belong to correctness and validity risk, not circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical/theoretical entities: its nodes are text chunks, keywords, entities and relations extracted from text. The free parameters are hyperparameters of the system, chosen for default settings and examined by sensitivity analysis, not fitted to match a target result. The axioms are standard retrieval assumptions plus the heuristic that PageRank-selected chunks carry the knowledge needed for fine-grained retrieval.

free parameters (7)
  • K (KNN graph degree) = 2
    Number of lexical and semantic neighbors per node in the intermediate KNN graph (Algorithm 3 Lines 3-6). Chosen as default; sensitivity analysis (Table 5) shows minor impact.
  • β (core chunk budget ratio) = 0.8
    Fraction of chunks sent to LLM for triplet extraction (Algorithm 3 Line 8). Controls the cost-quality trade-off; Figure 3 shows coverage is fairly stable for β in [0.6, 1.0].
  • θ (retrieval context split) = 0.4
    Proportion of the context budget given to the skeleton channel vs the keyword channel (Algorithm 4). Chosen by default; Figure 4 shows performance is robust for θ in [0.2, 0.6].
  • τ (sub-chunk splits) = 0 (low-cost) or 3 (high-accuracy)
    Controls the granularity of text sub-chunks in the keyword graph (Algorithm 3 Line 10). Smaller sub-chunks improve quality (Table 4) at some cost.
  • ℓ (input chunk size) = 1200 (low-cost) or 150 (high-accuracy)
    Token length of text chunks; from prior work (Edge et al. 2024). Smaller chunks improve quality but raise embedding cost.
  • λ (context length limit) = 12000
    Maximum tokens in the retrieved context for generation; follows the default of MS-Graph-RAG local search. Same across all methods.
  • PageRank teleport α = not stated (standard default ≈0.85)
    Teleport probability in Eq. (1); paper never fixes its value, so treated as an implicit standard choice.
assumptions (4)
  • domain assumption PageRank on the KNN graph ranks chunks by their structural importance for retrieval (Eq. 1, Section 5.1)
    The core-chunk selection depends on this. The paper motivates it with Figure 2 (degree distribution) but does not prove that PageRank high centrality implies answer-bearing entities.
  • domain assumption Cosine similarity of text embeddings (and co-occurring keywords) captures semantic relatedness between chunks and queries
    Used throughout for seed selection in both channels (Algorithms 2, 3, 4). This is a standard assumption in the RAG literature.
  • domain assumption An LLM can extract correct (entity, relation, entity) triplets from a text chunk with the prompting in Algorithm 1
    The skeleton graph inherits the accuracy of LLM extraction; the paper relies on GPT-4o-mini for this and does not audit extraction errors.
  • domain assumption Sentences containing a keyword, averaged into one embedding, serve as a useful representation of the keyword
    Keyword node construction (Algorithm 3 Lines 12-13). The averaging step is a heuristic not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KET-RAG: A Cost-Efficient Multi-Granular Indexing Framework for Graph-RAG." pith.science (2026). https://pith.science/paper/JCSTYNNR

@misc{pith2026250209304,
  author       = {Pith},
  title        = {Pith review of: KET-RAG: A Cost-Efficient Multi-Granular Indexing Framework for Graph-RAG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCSTYNNR}},
  note         = {Machine review of arXiv:2502.09304}
}
read the original abstract

Graph-RAG constructs a knowledge graph from text chunks to improve retrieval in Large Language Model (LLM)-based question answering. It is particularly useful in domains such as biomedicine, law, and political science, where retrieval often requires multi-hop reasoning over proprietary documents. Some existing Graph-RAG systems construct KNN graphs based on text chunk relevance, but this coarse-grained approach fails to capture entity relationships within texts, leading to sub-par retrieval and generation quality. To address this, recent solutions leverage LLMs to extract entities and relationships from text chunks, constructing triplet-based knowledge graphs. However, this approach incurs significant indexing costs, especially for large document collections. To ensure a good result accuracy while reducing the indexing cost, we propose KET-RAG, a multi-granular indexing framework. KET-RAG first identifies a small set of key text chunks and leverages an LLM to construct a knowledge graph skeleton. It then builds a text-keyword bipartite graph from all text chunks, serving as a lightweight alternative to a full knowledge graph. During retrieval, KET-RAG searches both structures: it follows the local search strategy of existing Graph-RAG systems on the skeleton while mimicking this search on the bipartite graph to improve retrieval quality. We evaluate 13 solutions on three real-world datasets, demonstrating that KET-RAG outperforms all competitors in indexing cost, retrieval effectiveness, and generation quality. Notably, it achieves comparable or superior retrieval quality to Microsoft's Graph-RAG while reducing indexing costs by over an order of magnitude. Additionally, it improves the generation quality by up to 32.4% while lowering indexing costs by around 20%.

Figures

Figures reproduced from arXiv: 2502.09304 by the authors.

Figure 1
Figure 1. The illustration of KET-RAG: the indexing stage in ○1 -○3 and the retrieval stage in ○4 -○5 . (1) KET-RAG first organizes the input text chunks in T into a KNN graph, where chunks are linked if they exhibit sufficient lexical or semantic similarity. This serves as an intermediate structure for building the final graph G. (2) Next, KET-RAG selects a 𝛽 fraction of core chunks according to their structural importance i… view at source ↗
Figure 2
Figure 2. Log-log Plot of the degree distribution of the KNN [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Answer quality by varying 𝛽. decrease of 3.5% and 7.4% in the low-cost setting on MuSiQue and HotpotQA, respectively. However, in high-accuracy settings, there is no performance drop, and in the case of HotpotQA, even a slight improvement is observed. These results suggest that Skeleton￾RAG effectively balances efficiency and quality, making it a viable alternative to full-scale knowledge graph indexing. Text-keywor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 19 canonical work pages

  1. [1]

    Mohannad Alhanahnah, Yazan Boshmaf, and Benoit Baudry. 2024. DepesRAG: Towards Managing Software Dependencies using Large Language Models. arXiv preprint arXiv:2405.20455 (2024)

  2. [2]

    Shawn Arnold and Clayton Romero. 2022. The Vital Role of Managing e- Discovery. https://legal-tech.blog/the-vital-role-of-managing-e-discovery

  3. [3]

    Lele Cao, Vilhelm von Ehrenheim, Mark Granroth-Wilding, Richard Anselmo Stahl, Andrew McCornack, Armin Catovic, and Dhiana Deva Cavalcanti Rocha. 2024. CompanyKG: A Large-Scale Heterogeneous Graph for Company Similarity Quantification. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4816–4827

  4. [4]

    Xuanzhong Chen, Xiaohao Mao, Qihan Guo, Lun Wang, Shuyang Zhang, and Ting Chen. 2024. RareBench: Can LLMs Serve as Rare Diseases Specialists?. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4850–4861

  5. [5]

    Andrea Colombo. 2024. Leveraging Knowledge Graphs and LLMs to Support and Monitor Legislative Systems. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 5443–5446

  6. [6]

    Mohammad Dehghan, Mohammad Alomrani, Sunyam Bagga, David Alfonso- Hermelo, Khalil Bibi, Abbas Ghaddar, Yingxue Zhang, Xiaoguang Li, Jianye Hao, Qun Liu, Jimmy Lin, Boxing Chen, Prasanna Parthasarathi, Mahdi Biparva, and Mehdi Rezagholizadeh. 2024. EWEK-QA : Enhanced Web and Efficient Knowledge Graph Retrieval for Citation-based Question Answering Systems....

  7. [7]

    Julien Delile, Srayanta Mukherjee, Anton Van Pamel, and Leonid Zhukov. 2024. Graph-Based Retriever Captures the Long Tail of Biomedical Knowledge. arXiv preprint arXiv:2402.12352 (2024)

  8. [8]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

Show all 44 references
  1. [9]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining ...

  2. [10]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero- shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 1762–1777

  3. [11]

    Ant Group and OpenKG. 2023. Semantic-enhanced Programmable Knowledge Graph (SPG) White paper (v1.0). https://spg.openkg.cn/en-US

  4. [12]

    Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3, 1 (...

  5. [13]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2025. LightRAG: Simple and Fast Retrieval-Augmented Generation. arXiv:2410.05779 [cs.IR] https: //arxiv.org/abs/2410.05779

  6. [14]

    Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su

  7. [15]

    Rujun Han, Yuhao Zhang, Peng Qi, Yumo Xu, Jenyuan Wang, Lan Liu, William Yang Wang, Bonan Min, and Vittorio Castelli. 2024. RAG-QA Arena: Evaluating Domain Robustness for Long-form Retrieval Augmented Question Answering. In Proceedings of the 2024 Conference on Empirical Metho...

  8. [16]

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. In The Thirty-eighth Annual Conference on Neural Informatio...

  9. [17]

    Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al . 2024. Graph chain- of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103 (2024)

  10. [18]

    Rishi Kalra, Zekun Wu, Ayesha Gulley, Airlie Hilliard, Xin Guan, Adriano Koshiyama, and Philip Treleaven. 2024. HyPA-RAG: A Hybrid Parameter Adap- tive Retrieval-Augmented Generation System for AI Legal and Policy Appli- cations. In Proceedings of the 1st Workshop on Customiza...

  11. [19]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing S...

  12. [20]

    Dawei Li, Shu Yang, Zhen Tan, Jae Young Baik, Sukwon Yun, Joseph Lee, Aaron Chacko, Bojian Hou, Duy Duong-Tran, Ying Ding, Huan Liu, Li Shen, and Tian- long Chen. 2024. DALK: Dynamic Co-Augmentation of LLMs and KG to answer Alzheimer‘s Disease Questions with Scientific Literat...

  13. [21]

    Zijian Li, Qingyan Guo, Jiawei Shao, Lei Song, Jiang Bian, Jun Zhang, and Rui Wang. 2024. Graph Neural Network Enhanced Retrieval for Question Answering of LLMs. arXiv preprint arXiv:2406.06572 (2024)

  14. [22]

    Costas Mavromatis and George Karypis. 2024. GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning. arXiv preprint arXiv:2405.20139 (2024)

  15. [23]

    Sewon Min, Danqi Chen, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2019. Knowledge guided text retrieval and reading for open domain question answering. arXiv preprint arXiv:1911.03868 (2019)

  16. [24]

    Xinyi Mou, Zejun Li, Hanjia Lyu, Jiebo Luo, and Zhongyu Wei. 2024. Unifying Local and Global Knowledge: Empowering Large Language Models as Political Experts with Knowledge Graphs. In Proceedings of the ACM on Web Conference

  17. [25]

    Sai Munikoti, Anurag Acharya, Sridevi Wagle, and Sameera Horawalavithana

  18. [26]

    NebulaGraph. 2023. NebulaGraph Launches Industry-First Graph RAG: Retrieval- Augmented Generation with LLM Based on Knowledge Graphs. https://www. nebula-graph.io/posts/graph-RAG

  19. [27]

    Neo4j. 2023. NaLLM. https://github.com/neo4j/NaLLM

  20. [28]

    Proceedings of the Workshop on AI to Accelerate Science and Engineering (AI2ASE)

    ATLANTIC: Structure-Aware Retrieval-Augmented Language Model for Interdisciplinary Science. Proceedings of the Workshop on AI to Accelerate Science and Engineering (AI2ASE). Held in conjunction with the 38th AAAI Conference on Artificial Intelligence. (2024)

  21. [29]

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph Retrieval-Augmented Generation: A Survey. arXiv preprint arXiv:2408.08921 (2024)

  22. [30]

    Zhuoyi Peng and Yi Yang. 2024. Connecting the Dots: Inferring Patent Phrase Similarity with Retrieved Phrase Graphs. In Findings of the Association for Com- putational Linguistics: NAACL 2024

  23. [31]

    Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Technical Report. Stanford InfoLab

  24. [32]

    Bhaskarjit Sarmah, Benika Hall, Rohan Rao, Sunil Patel, Stefano Pasquali, and Dhagash Mehta. 2024. HybridRAG: Integrating Knowledge Graphs and Vector Retrieval Augmented Generation for Efficient Information Extraction. arXiv preprint arXiv:2408.04948 (2024)

  25. [33]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

  26. [34]

    Juan Ramos et al. 2003. Using tf-idf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning , Vol. 242. Citeseer, 29–48

  27. [35]

    Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr

  28. [36]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024. Retrieval-augmented generation with knowledge graphs for customer service question answering. In Proceedings of the 47th International ACM SIGIR Conference on Research...

  29. [37]

    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 ...

  30. [38]

    Ruijie Wang, Zheng Li, Danqing Zhang, Qingyu Yin, Tong Zhao, Bing Yin, and Tarek Abdelzaher. 2022. RETE: retrieval-enhanced temporal event forecasting on unified query product evolutionary graph. In Proceedings of the ACM Web Conference 2022. 462–472

  31. [39]

    Yingli Zhou, Yaodong Su, Youran Sun, Shu Wang, Taotao Wang, Runyuan He, Yongwei Zhang, Sicong Liang, Xilin Liu, Yuchi Ma, et al. 2025. In-depth Analysis of Graph-based RAG in a Unified Framework. arXiv preprint arXiv:2503.04338 (2025)

  32. [40]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Knowledge graph prompting for multi-document question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 19206–19214

  33. [43]

    Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong, Qi Liu, and Zhaofeng Liu. 2024. Evaluation of retrieval-augmented generation: A survey. In CCF Conference on Big Data. Springer, 102–120

  34. [2022]

    Transactions of the Association for Computational Linguistics 10 (2022), 539–554

    MuSiQue: Multihop Questions via Single-hop Question Composition. Transactions of the Association for Computational Linguistics 10 (2022), 539–554

  35. [2024]

    In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=hkujvAPVsg

  36. [4374]

    doi:10.18653/v1/2024.emnlp-main.249

Pith tools

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