Pith. sign in

REVIEW 4 major objections 6 minor 47 references

XGraphRAG: Interactive Visual Analysis for Graph-based Retrieval-Augmented Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read XGraphRAG gives GraphRAG developers a visual path from a wrong answer to the specific recalls that caused it.

desk verdict Useful first stab at visual GraphRAG debugging, but the LLM-based recall reconstruction is unvalidated and the user study tasks give away the answers. read the letter →

arxiv 2506.13782 v1 pith:BJSKINTK submitted 2025-06-10 cs.IR cs.AI

Ke Wang , Bo Pan , Yingchaojie Feng , Yuwei Wu , Jieyi Chen , Minfeng Zhu , Wei Chen This is my paper
classification cs.IRcs.AI
keywords graph-basedretrieval-augmentedgenerationvisualanalyticsLLMinvocationtracingretrievaldiagnosisinteractivevisualizationknowledgegraphuserstudy
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

Graph-based Retrieval-Augmented Generation (GraphRAG) answers questions by building a knowledge graph from documents and making many large-language-model calls during construction and querying, which makes failures hard to explain. The paper proposes XGraphRAG, a visual analysis system that helps developers find which retrieved pieces of the graph (recalls) led a GraphRAG system astray, and then trace those recalls back through the pipeline to the exact LLM invocation and source text where information was lost or distorted. Its two-stage workflow first compares the generated answer against ground truth to flag 'missing' and 'unexpected' recalls, then lets users inspect global topic-level and local entity-level relevance and drill into the LLM's extraction, merge, and summarization behavior. The authors evaluate the system with a user study involving 14 RAG engineers and report that, compared with the Kotaemon baseline, XGraphRAG significantly improved task accuracy while reducing completion time.

What carries the argument

The central object is the comparative inference trace: two parallel Question-Recalls-Inferring-Answer chains, one reconstructed from the generated answer and one from the ground truth. Each chain is built by an LLM that expands ground-truth facts, reverse-infers the reasoning steps, and attaches the relevant graph recalls to each step; the actual-answer side inherits the recalls the GraphRAG system really retrieved. This symmetrical representation turns 'wrongness' into a two-way discrepancy: recalls essential to the correct chain but absent from the actual chain (missing), and recalls present in the actual chain but unrelated to the correct one (unexpected). The supporting machinery is the four linked views — QA, Inference Trace, Topic Explore, Entity Explore — plus the LLM Invocation View, which together make each recall traceable forward to the inference steps that used it and backward to the merge, extraction, and summarization calls and source chunks that produced it.

What would settle it

Take a set of GraphRAG questions, run the full system, and have independent annotators mark which recalls were truly missing or unexpected from the system's own retrieval logs; if the LLM-reconstructed traces disagree with those log-based traces on a substantial share of cases, Stage 1's foundation is broken.

Watch

Extended reading notes

Core claim

XGraphRAG's central claim is that the diagnosis of a wrong GraphRAG answer can be reduced to locating suspicious recalls and tracing their provenance. The system reconstructs the inference process of both the actual generated answer and the ground-truth answer in a Question-Recalls-Inferring-Answer template, so that each inference step names the entity, relationship, or topic-report recalls it relied on. Comparing the two chains exposes what the authors call Missing Recalls (pieces essential to the correct reasoning but absent from the actual retrieval) and Unexpected Recalls (pieces that entered the actual reasoning but are not part of the correct answer). Users then move to graph-based analysis: the Topic Explore View examines global semantic communities, the Entity Explore View examines local connectivity around a suspicious entity, and the LLM Invocation View shows the actual prompts, inputs, and outputs of the extraction, merge, and summarization calls that produced each recall. The paper argues that this chain of views turns an opaque LLM pipeline into a traceable object, and its user study reports that participants using XGraphRAG completed diagnostic tasks more accurately and faster than with the Kotaemon baseline.

Load-bearing premise

Stage 1 assumes the LLM can reliably judge answers, expand ground-truth facts, reconstruct both inference chains, and classify recalls as missing or unexpected, but the paper reports no check of that reconstruction against human annotations or the system's actual traced steps.

Editorial extensions

If this is right

  • A wrong GraphRAG answer can be turned into a concrete, localizable defect: a missing relationship between two entities, a topic report that omitted an entity, or an extraction failure in one source chunk.
  • Developers can systematically collect failure cases from a QA test set because the system automatically flags where the generated inference chain diverges from the ground-truth chain.
  • The same pipeline abstraction (split, extract, merge, summarize, recall, infer) can be used to compare information loss across different GraphRAG implementations, not just the one the system was built on.
  • Because each recall is linked to the LLM invocation that created it, fixing a bad answer becomes a matter of patching a specific construction stage instead of blind prompt or parameter retuning.
  • The reported accuracy and time gains in the user study suggest that such traceability tools can make GraphRAG practical to deploy for developers who currently lack confidence in its outputs.

Reading between the lines

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

  • The paper does not validate Stage 1's LLM reconstruction of inference steps against human annotation or against the GraphRAG system's actual logged steps; if the reconstruction is wrong, the missing and unexpected flags could misdirect the entire downstream analysis.
  • A natural next validation would compare the LLM-reconstructed inference trace with the real recalls each inference step used in the GraphRAG run, measuring divergence directly.
  • The design implies a traceability principle for retrieval systems generally: if every recall exposes its generative LLM invocation and source chunk, debugging becomes a provenance lookup rather than a black-box search.
  • The evaluation uses one news benchmark dataset and GPT-4 as the underlying model; whether the time and accuracy advantages hold across domains, model families, and graph sizes remains open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents XGraphRAG, an interactive visual analysis system for diagnosing failures in Graph-based Retrieval-Augmented Generation (GraphRAG) pipelines. The authors first abstract a common GraphRAG pipeline and derive four design requirements from a formative study with five RAG experts. The proposed workflow has two stages: Stage 1 uses an LLM-assisted analysis framework to judge whether a generated answer is correct, reconstruct inference chains for the actual answer and the ground truth, and flag missing or unexpected recalls; Stage 2 supports local and global graph exploration, LLM invocation inspection, and tracing back to raw chunks. The system is evaluated through a usage scenario on the MultiHop-RAG dataset and a user study comparing XGraphRAG with Kotaemon, reporting improved task accuracy and reduced completion time. The manuscript claims that XGraphRAG helps developers locate suspicious recalls and trace them through the GraphRAG pipeline, and the paper is open-sourced.

Significance. If the central claim holds, XGraphRAG would fill a genuine gap: no existing tool provides end-to-end diagnostic support for GraphRAG systems, where the pipeline spans graph construction, multiple retrieval stages, and numerous LLM invocations. The paper's strengths include a clearly articulated common-pipeline abstraction, design requirements grounded in a formative study, a coherent set of coordinated views (QA/Inference Trace, Topic Explore, Entity Explore, LLM Invocation), a detailed usage scenario, and an open-source prototype. The user study against a popular baseline is a reasonable first step. However, the paper's central promise depends on Stage 1's LLM-based reconstruction of inference steps and ground-truth recalls, and that reconstruction is never validated. The evaluation also inherits that unvalidated ground truth and reports significance without statistical details. The contribution is therefore plausible and potentially valuable, but the evidence as presented does not yet establish that XGraphRAG identifies genuine retrieval failures rather than plausible LLM-invented ones.

major comments (4)
  1. [Section 4.2.1–4.2.2] The Stage 1 inference-chain reconstruction is the epistemic foundation of the entire workflow, yet it is never validated. Section 4.2 says the system uses an LLM-assisted framework to 'reconstruct the inference process for generated answers and ground truth,' and Section 4.2.1 describes IA as inference steps generated when the GraphRAG system processes recalls, while IG is produced by GPT-4 expanding facts, reverse-inferring steps, and selecting entities and relationships from fact subgraphs. The manuscript does not state whether IA is a logged trace of the actual GraphRAG behavior or another LLM reconstruction, and no evidence is provided that the reconstructed recalls correspond to what the retriever actually used. There is no human-annotation agreement study, no precision/recall comparison between the flagged recalls and the real retriever output, and no sanity check on a sample of cases. If GPT-4's reverse inference produces a plausible but incorrect chain, every 'missing recall' or 'unexpected recall' shown to the user is an artifact, and all downstream views in Sections 5 and 6 inherit that error. This is a load-bearing gap, not an ablation detail.
  2. [Section 6.2.4 / Table 1] The claim that 'using XGraphRAG significantly improved task accuracy while reducing time consumption' rests on p-values (p1 = 0.00012, p2 = 0.00012, p3 = 0.0033, p4 = 0.0022, p5 = 0.012, p6 = 0.0051) reported without the statistical test used, the unit of analysis, whether accuracy and time were analyzed separately, or any effect sizes or confidence intervals. With 14 participants and six tasks, the reader cannot assess whether the effects are large, consistent across participants, or robust to multiple-comparison correction. Please report the test procedure, effect sizes, and per-task descriptive statistics for both accuracy and time, and clarify whether the p-values are per-task, pooled, or adjusted.
  3. [Section 6.2.4] The user-study tasks are engineered known-answer tasks that presuppose the very outputs that Stage 1 was meant to produce. For example, the T3 prompt tells participants that 'the question was answered incorrectly because the Google entity was not recalled' and asks them to find information related to that missing entity; T6 similarly states that the relationship between Taylor Swift and Travis Kelce is missing in a specified chunk. This design measures whether users can navigate to a pre-labeled target, not whether XGraphRAG correctly identifies genuine suspicious recalls from an unconstrained analysis. The task-accuracy measure is therefore better characterized as completion of a guided navigation task, and the ground truth for those tasks is imported from the same unvalidated Stage 1 reconstruction discussed above. The evaluation should separately validate Stage 1 outputs against human annotations or logged retrieval behavior before using them as task ground truth.
  4. [Section 6.2.2 and Section 4.2] The evaluation pipeline uses GPT-4 both as the LLM inside the GraphRAG system being analyzed and as the LLM that judges answers and reconstructs inference chains. The manuscript does not discuss the resulting self-consistency risk: the judge may be internally consistent with the generator's errors rather than faithful to the actual retrieval process. A concrete remedy is to validate Stage 1 reconstructed recalls and answer judgments against human annotators and against logged recalls from the GraphRAG system on a sample of cases, reporting agreement metrics such as Cohen's kappa or precision/recall of the flags. Without such a check, the distinction between 'missing recall' and 'hallucinated gap' remains unexamined.
minor comments (6)
  1. [Section 6.2.2 and Section 6.2.4] The text refers to 'the XGraphRAG[6] system' in two places, but reference [6] is Microsoft's GraphRAG, not XGraphRAG; these bracket citations should be corrected or removed.
  2. [Table 1] The table is difficult to read: it lists paired ratios such as '14/14' alongside time values without clear column headers indicating which value is accuracy and which is time, and the use of '×' markers is unexplained. Please restructure the table with labeled columns, units, and explicit accuracy counts.
  3. [Section 5.2] There is a typographical error in '(Fig 5C)' where the closing parenthesis is unbalanced; it should read '(Fig 5C)'.
  4. [Reference list] Reference [33] is incomplete: 'Kotaemon team, 2024' lacks a title, URL, and version information, which is insufficient for a tool used as the baseline.
  5. [Section 6.2.3] The procedure says tasks were assigned in randomized order but does not specify how many test cases were used per task, whether each participant saw a different test case, or how the cases were balanced across systems; please clarify this to make Table 1 interpretable.
  6. [Fig. 1 caption] The caption states that XGraphRAG contains four views, but it then labels five panels (A, B, C, D, E); please reconcile the count or describe the QA and Inference Trace views as two coordinated panels within one view.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity: Stage 1's LLM reconstruction is an unvalidated heuristic, but the paper does not reduce any prediction to its inputs, and tasks T3-T6 presuppose rather than test those reconstructions.

full rationale

XGraphRAG is a visual analytics system, not a model with fitted parameters; there is no equation in the paper whose output equals an input by construction. The common GraphRAG pipeline (Sec 3.1) is derived from an external 30-work survey seeded by third-party systems (GraphRAG, KAG, LightRAG), and the evaluation uses the external MultiHop-RAG dataset and a Kotaemon baseline. Stage 1 (Sec 4.2) does use GPT-4 to reconstruct the actual-answer inference chain (IA) and a reverse-inferred ground-truth chain (IG), then flags Missing/Unexpected Recalls by comparing them. This creates a genuine validity risk: if the LLM hallucinates a plausible chain, downstream 'suspicious recalls' inherit the error, and no human-annotation or precision/recall validation is reported against the real retriever behavior. But that is a validation gap, not circularity: the flags are not fitted parameters renamed as predictions, and the paper never defines ground truth as the output of its own system. The user-study tasks T3-T6 explicitly give participants the missing-recall premise ('It is already known ... the Google entity was not recalled'), so they test navigation/traceability, not Stage-1 detection; T1-T2 scoring criteria are under-specified but no text shows the labels are generated by the same Stage-1 reconstruction. The self-citations ([19], [20], [22], [37], etc.) are related-work or general method pointers and are not load-bearing; none supplies a uniqueness theorem or forbids alternatives. No pattern 1-6 is present.

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

No numerical parameters are fitted to data; the system depends on qualitative assumptions. The most important is that the LLM-as-judge reliably reconstructs inference steps. The common-pipeline abstraction and the representative test-set assumption are also load-bearing. There are no new physical or conceptual entities.

assumptions (4)
  • domain assumption GPT-4 can reliably judge answer correctness and reconstruct inference steps from recalls.
    Section 4.2 uses LLM-assisted judgment and reverse inference to build comparison traces; no reliability analysis is provided.
  • domain assumption The common GraphRAG pipeline (split, extract, merge, summarize, recall, infer) generalizes across GraphRAG systems.
    Section 3.1 abstracts this pipeline from 30 works; Section 7 acknowledges topic reports differ by system.
  • domain assumption The modified MultiHop-RAG test questions represent real GraphRAG debugging scenarios.
    Section 6.2.2 filters questions to known entities and has experts verify them, which may not cover open-ended failures.
  • domain assumption Visual encodings such as node size, edge thickness, and topic circles convey relevance accurately.
    Section 5 assumes these cues aid analysis; the user study provides indirect, task-specific evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XGraphRAG: Interactive Visual Analysis for Graph-based Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/BJSKINTK

@misc{pith2026250613782,
  author       = {Pith},
  title        = {Pith review of: XGraphRAG: Interactive Visual Analysis for Graph-based Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BJSKINTK}},
  note         = {Machine review of arXiv:2506.13782}
}
read the original abstract

Graph-based Retrieval-Augmented Generation (RAG) has shown great capability in enhancing Large Language Model (LLM)'s answer with an external knowledge base. Compared to traditional RAG, it introduces a graph as an intermediate representation to capture better structured relational knowledge in the corpus, elevating the precision and comprehensiveness of generation results. However, developers usually face challenges in analyzing the effectiveness of GraphRAG on their dataset due to GraphRAG's complex information processing pipeline and the overwhelming amount of LLM invocations involved during graph construction and query, which limits GraphRAG interpretability and accessibility. This research proposes a visual analysis framework that helps RAG developers identify critical recalls of GraphRAG and trace these recalls through the GraphRAG pipeline. Based on this framework, we develop XGraphRAG, a prototype system incorporating a set of interactive visualizations to facilitate users' analysis process, boosting failure cases collection and improvement opportunities identification. Our evaluation demonstrates the effectiveness and usability of our approach. Our work is open-sourced and available at https://github.com/Gk0Wk/XGraphRAG.

Figures

Figures reproduced from arXiv: 2506.13782 by the authors.

Figure 1
Figure 1. XGraphRAG contains four views: (A&B) The QA & Inference Trace View allows users to locate suspicious recalls. (C) The Topic Explore View supports global relevance analysis on the graph for suspicious retrievals. (D) The Entity Explore View supports local relevance analysis on the graph for suspicious retrievals. (E) The LLM Invocation View presents details for each LLM invocation. Abstract—Graph-based Retrieval-Augm… view at source ↗
Figure 2
Figure 2. Common Pipeline of GraphRAG, which is typically divided into an [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the workflow: The user first prepares a testing pair [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: QA View & Inference Trace View. Once the user inputs the test [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: The Entity Explore View enables investigation of entity recalls [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 5
Figure 5. Figure 5: Topic Explore View, which displays all topics with explicit nesting [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: The LLM Invocation View has three variations, showing LLM [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 11 canonical work pages

  1. [1]

    Abd-Alrazaq, R

    A. Abd-Alrazaq, R. AlSaad, D. Alhuwail, A. Ahmed, P. M. Healy, S. Latifi, S. Aziz, R. Damseh, S. A. Alrazak, J. Sheikh, et al. Large language models in medical education: opportunities, challenges, and future directions. JMIR Medical Education, 9(1):e48291, 2023. doi: 10.2196/48291 1

    doi:10.2196/48291 2023

  2. [2]

    M. T. Agler, B. A. Wrenn, S. H. Zinder, and L. T. Angenent. Waste to bioproduct conversion with undefined mixed cultures: the carboxylate platform.Trends in biotechnology, 29(2):70–78, 2011. doi: 10.1016/j. tibtech.2010.11.006 1

    doi:10.1016/j 2011

  3. [3]

    Alhanahnah and Y

    M. Alhanahnah and Y . Boshmaf. Depsrag: Towards agentic reasoning and planning for software dependency management.CoRR, abs/2405.20455,

    arXiv

  4. [4]

    Arawjo, C

    I. Arawjo, C. Swoopes, P. Vaithilingam, M. Wattenberg, and E. L. Glass- man. Chainforge: A visual toolkit for prompt engineering and llm hypothe- sis testing. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, article no. 304, 18 pages. Association for Computing Machinery, New York, NY , USA, 2024. doi: 10.1145/361390...

    doi:10.1145/3613904 2024

  5. [5]

    Brath, A

    R. Brath, A. Bradley, and D. Jonker. Visualizing llm text style transfer: visually dissecting how to talk like a pirate. InProceedings of IEEE VIS 2023, 2023. 2

    work page 2023

  6. [6]

    D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson. From local to global: A graph rag approach to query- focused summarization.arXiv preprint arXiv:2404.16130, 2024. doi: 10. 48550/ARXIV.2404.16130 1, 2, 3, 8, 9

    arXiv doi:10.48550/arxiv.2404.16130 2024

  7. [7]

    W. Fan, Y . Ding, L. Ning, S. Wang, H. Li, D. Yin, T. Chua, and Q. Li. A survey on RAG meeting llms: Towards retrieval-augmented large language models. In R. Baeza-Yates and F. Bonchi, eds.,Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, pp. 6491–6501. ACM,

    work page 2024

  8. [8]

    Y . Feng, Z. Chen, Z. Kang, S. Wang, M. Zhu, W. Zhang, and W. Chen. Jailbreaklens: Visual analysis of jailbreak attacks against large language models.arXiv preprint arXiv:2404.08793, 2024. doi: 10.48550/arXiv. 2404.08793 2

    arXiv doi:10.48550/arxiv.2404.08793 2024

Show all 47 references
  1. [9]

    doi: 10.1145/3637528.3671470 1

  2. [10]

    Y . Feng, X. Wang, K. Wong, S. Wang, Y . Lu, M. Zhu, B. Wang, and W. Chen. Promptmagician: Interactive prompt engineering for text-to- image creation.IEEE Trans. Vis. Comput. Graph., 30(1):295–305, 2024. doi: 10.1109/TVCG.2023.3327168 8

    2024

  3. [11]

    Y . Feng, X. Wang, B. Pan, K. K. Wong, Y . Ren, S. Liu, Z. Yan, Y . Ma, H. Qu, and W. Chen. Xnli: Explaining and diagnosing nli-based visual data analysis.IEEE Trans. Vis. Comput. Graph., pp. 1–14, 2023. doi: 10. 1109/TVCG.2023.3240003 8

    2023

  4. [12]

    Z. Guo, L. Xia, Y . Yu, T. Ao, and C. Huang. Lightrag: Simple and fast retrieval-augmented generation.arXiv preprint arXiv:2410.05779, 2024. doi: 10.48550/arXiv.2410.05779 1, 3, 8

    2024 doi arXiv

  5. [13]

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, Q. Guo, M. Wang, and H. Wang. Retrieval-augmented generation for large lan- guage models: A survey.CoRR, abs/2312.10997, 2023. doi: 10.48550/ ARXIV.2312.10997 1

    2023 doi arXiv

  6. [14]

    X. He, Y . Tian, Y . Sun, N. V . Chawla, T. Laurent, Y . LeCun, X. Bresson, and B. Hooi. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering.CoRR, abs/2402.07630,

    arXiv

  7. [15]

    B. J. Gutiérrez, Y . Shu, Y . Gu, M. Yasunaga, and Y . Su. Hipporag: Neurobiologically inspired long-term memory for large language models. CoRR, abs/2405.14831, 2025. doi: 10.48550/arXiv.2405.14831 2

    2025 doi arXiv

  8. [16]

    Z. Li, L. Deng, H. Liu, Q. Liu, and J. Du. Unioqa: A unified framework for knowledge graph question answering with large language models, 2024. 2

    2024

  9. [17]

    doi: 10.48550/ARXIV.2402.07630 2

    doi arXiv

  10. [18]

    Kahng, I

    M. Kahng, I. Tenney, M. Pushkarna, M. X. Liu, J. Wexler, E. Reif, K. Kallarackal, M. Chang, M. Terry, and L. Dixon. Llm comparator: Interactive analysis of side-by-side evaluation of large language models. IEEE Transactions on Visualization and Computer Graphics, 31(1):503– 51...

    2025

  11. [19]

    Y . Liu, Z. Wen, L. Weng, O. Woodman, Y . Yang, and W. Chen. SPROUT: an interactive authoring tool for generating programming tutorials with the visualization of large language models.IEEE Transactions on Visu- alization and Computer Graphics, pp. 1–15, 2024. doi: 10.1109/TVCG...

    2024

  12. [20]

    Z. Li, Q. Guo, J. Shao, L. Song, J. Bian, J. Zhang, and R. Wang. Graph neural network enhanced retrieval for question answering of llms.CoRR, abs/2406.06572, 2024. doi: 10.48550/arXiv.2406.06572 2

    2024 doi arXiv

  13. [21]

    Liang, M

    L. Liang, M. Sun, Z. Gui, Z. Zhu, Z. Jiang, L. Zhong, Y . Qu, P. Zhao, Z. Bo, J. Yang, H. Xiong, L. Yuan, J. Xu, Z. Wang, Z. Zhang, W. Zhang, H. Chen, W. Chen, and J. Zhou. KAG: boosting llms in professional domains via knowledge augmented generation.CoRR, abs/2409.13731,

    arXiv

  14. [22]

    doi: 10.48550/ARXIV.2409.13731 3

    doi arXiv

  15. [23]

    B. Peng, Y . Zhu, Y . Liu, X. Bo, H. Shi, C. Hong, Y . Zhang, and S. Tang. Graph retrieval-augmented generation: A survey.arXiv preprint arXiv:2408.08921, 2024. doi: 10.48550/arXiv.2408.08921 1

    2024 doi arXiv

  16. [24]

    J. Lu, B. Pan, J. Chen, Y . Feng, J. Hu, Y . Peng, and W. Chen. AgentLens: Visual analysis for agent behaviors in llm-based autonomous systems. CoRR, abs/2402.08995, Feb. 2024. doi: 10.48550/arXiv.2402.08995 2

    2024 doi arXiv

  17. [25]

    S. Ma, C. Xu, X. Jiang, M. Li, H. Qu, C. Yang, J. Mao, and J. Guo. Think-on-graph 2.0: Deep and faithful large language model reasoning with knowledge-guided retrieval augmented generation.arXiv preprint arXiv:2407.10805, 2024. doi: 10.48550/arXiv.2407.10805 2

    2024 doi arXiv

  18. [26]

    B. Pan, J. Lu, K. Wang, L. Zheng, Z. Wen, Y . Feng, M. Zhu, and W. Chen. Agentcoord: Visually exploring coordination strategy for llm-based multi- agent collaboration.CoRR, abs/2404.11943, 2024. doi: 10.48550/arXiv. 2404.11943 2

    2024 doi arXiv

  19. [27]

    Sarmah, D

    B. Sarmah, D. Mehta, B. Hall, R. Rao, S. Patel, and S. Pasquali. Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. InProceedings of the 5th ACM Inter- national Conference on AI in Finance, ICAIF 2024, Brookl...

    2024 doi

  20. [28]

    T. T. Procko and O. Ochoa. Graph retrieval-augmented generation for large language models: A survey.Available at SSRN, pp. 166–169, 2024. doi: 10.1109/AIxSET62544.2024.00030 1

    2024

  21. [29]

    Ranade and A

    P. Ranade and A. Joshi. Fabula: Intelligence report generation using retrieval-augmented narrative construction. InProceedings of the Interna- tional Conference on Advances in Social Networks Analysis and Mining, ASONAM ’23, p. 603–610. ACM, Nov. 2023. doi: 10.1145/3625007. 3627505 2

    doi

  22. [30]

    M. Sallam. Chatgpt utility in healthcare education, research, and prac- tice: systematic review on the promising perspectives and valid con- cerns. InHealthcare, vol. 11, p. 887. MDPI, 2023. doi: 10.3390/ healthcare11060887 1

    2023

  23. [31]

    Tang and Y

    Y . Tang and Y . Yang. Mdalk: Dynamic co-augmentation of llms and kg to answer alzheimer’s disease questions with scientific literature.CoRR, abs/2405.04819, 2024. doi: 10.48550/arXiv.2405.04819 2

    2024 doi arXiv

  24. [32]

    Strobelt, A

    H. Strobelt, A. Webson, V . Sanh, B. Hoover, J. Beyer, H. Pfister, and A. M. Rush. Interactive and visual prompt engineering for ad-hoc task adaptation with large language models.IEEE Transactions on Visualization and Computer Graphics, 29(1):1146–1156, 2023. doi: 10.1109/TVCG...

    2023 doi

  25. [33]

    S. Suh, B. Min, S. Palani, and H. Xia. Sensecape: Enabling multilevel exploration and sensemaking with large language models. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, article no. 1, 18 pages. Association for Computing ...

    2023

  26. [34]

    J. Sun, C. Xu, L. Tang, S. Wang, C. Lin, Y . Gong, H.-Y . Shum, and J. Guo. Think-on-graph: Deep and responsible reasoning of large language model with knowledge graph.CoRR, abs/2307.07697, 2023. doi: 10.48550/arXiv .2307.07697 2

    2023 doi arXiv

  27. [35]

    New Technologies, Development and Applications

    D. Tiro. The possibility of applying chatgpt (ai) for calculations in me- chanical engineering. InInternational Conference “New Technologies, Development and Applications”, pp. 313–320. Springer, 2023. 1

    2023

  28. [36]

    Tang and Y

    Y . Tang and Y . Yang. Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries.CoRR, abs/2401.15391, 2024. doi: 10. 48550/ARXIV.2401.15391 7

    2024 doi arXiv

  29. [37]

    K. Team. Kotaemon team, 2024. 3

    2024

  30. [38]

    A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. F. Tan, and D. S. W. Ting. Large language models in medicine.Nature medicine, 29(8):1930–1940, 2023. doi: 10.1038/s41591-023-02448-8 1

    1930 doi

  31. [39]

    J. Wu, J. Zhu, and Y . Qi. Medical graph RAG: towards safe medical large language model via graph retrieval-augmented generation.CoRR, abs/2408.04187, 2024. doi: 10.48550/ARXIV.2408.04187 1

    2024 doi arXiv

  32. [40]

    X. Wang, N. Anwer, Y . Dai, and A. Liu. Chatgpt for design, manufacturing, and education.Procedia CIRP, 119:7–14, 2023. doi: 10.1016/j.procir. 2023.04.001 1

    2023 doi

  33. [41]

    X. Wang, Z. Wu, W. Huang, Y . Wei, Z. Huang, M. Xu, and W. Chen. Vis+ai: integrating visualization with artificial intelligence for efficient data analysis.Front. Comput. Sci., 17(6), 12 pages, June 2023. doi: 10. 1007/s11704-023-2691-y 4

    2023

  34. [42]

    Wardat, M

    Y . Wardat, M. A. Tashtoush, R. AlAli, and A. M. Jarrah. Chatgpt: A revolutionary tool for teaching and learning mathematics.Eurasia Journal of Mathematics, Science and Technology Education, 19(7):em2286, 2023. doi: 10.29333/ejmste/13272 1

    2023 doi

  35. [43]

    G. Ye. nano-graphrag: A simple, easy-to-hack GraphRAG implementation. https://github.com/gusye1234/nano-graphrag, 2024. 8

    2024

  36. [44]

    S. Wu, O. Irsoy, S. Lu, V . Dabravolski, M. Dredze, S. Gehrmann, P. Kam- badur, D. S. Rosenberg, and G. Mann. Bloomberggpt: A large language model for finance.CoRR, abs/2303.17564, 2023. doi: 10.48550/ARXIV. 2303.17564 1

    2023 doi arXiv

  37. [45]

    Z. Xu, M. J. Cruz, T. W. Matthew Guevara, M. Deshpande, X. Wang, and Z. Li. Retrieval-augmented generation with knowledge graphs for customer service question answering.CoRR, abs/2404.17723, 2024. doi: 10.1145/3626772.3661370 2

    2024 arXiv

  38. [46]

    H. Yang, X. Liu, and C. D. Wang. Fingpt: Open-source financial large language models.CoRR, abs/2306.06031, 2023. doi: 10.48550/ARXIV. 2306.06031 1

    2023 doi

  39. [2024]

    doi: 10.48550/arXiv.2405.20455 2

    doi arXiv

Pith tools

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