Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

A graph-based agentic system lifts commercial-registry QA correctness from 0.26 to 0.83, the paper reports.

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

2026-08-02 16:14 UTC pith:DJH4DJMD

load-bearing objection A genuinely useful agentic GraphRAG system whose evaluation rests on a judge the paper itself shows to be unreliable—likely a real effect, but the headline numbers are not established. the 4 major comments →

arxiv 2605.18770 v2 pith:DJH4DJMD submitted 2026-04-15 cs.IR cs.AI

Agentic Graph Retrieval-Augmented Generation for Auditable Commercial Registry Analysis

Arthur Capozzi , Dirk Helbing This is my paper
classification cs.IR cs.AI
keywords GraphRAGagentic AIknowledge graphsentity resolutioncommercial registryquestion answeringretrieval-augmented generationauditability
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.

The paper sets out to show that an agentic system built on a knowledge graph can make formally open but practically unusable commercial registry data answerable in plain language. It constructs a graph of more than five million nodes and 4.7 million relationships from Swiss gazette publications, combining deterministic registry fields with LLM-extracted actors and resolving identities through strict alphabetical tokenization. On a manually curated benchmark, graph-mediated retrieval raises factual correctness from 0.26 for the strongest flat-retrieval baseline to 0.83 for the full system, with comparable gains in relevance and completeness. The authors argue that the architecture's modularity and the evidence dashboard make it a viable template for auditable analysis of public and private registries.

Core claim

Central claim: for expert-oriented exploration of commercial registry data, an agentic architecture that routes natural-language queries through a deterministically resolved knowledge graph is considerably more accurate and robust than standard vector retrieval. The paper's pipeline turns seven million gazette notices into a knowledge graph (strong nodes from verified fields, weak nodes from LLM extraction, NameHub identity resolution), then lets a bounded agent (intent router, at most four reflection steps, state-machine-guided synthesis) traverse it. Evaluation reports correctness 0.828 vs 0.143 and information recall 0.837 vs 0.118 on the golden benchmark, and 0.760 vs 0.344 correctness i

What carries the argument

The load-bearing mechanism is the NameHub identity-resolution layer: every company and person name is normalized by lowercasing, stripping punctuation, sorting tokens alphabetically, and concatenating, so 'Doe, John' and 'John Doe' resolve to the same hub deterministically, with no fuzzy tolerance. Around this, the analytical agent couples a zero-shot intent router (restricting tools by query type), a bounded reflection loop (max four tool calls with backend-injected failure feedback), and a strict state machine that forces response formatting and next-step suggestions. The graph stores companies, persons, events, and NameHubs linked by typed edges, enabling multi-hop traversal that flat vec

Load-bearing premise

The evaluation assumes that floating-point scores assigned by an LLM judge are valid measures of answer quality, although the paper itself documents the judge giving near-zero scores to factually correct, well-formatted answers.

What would settle it

Have human experts score the same 60 golden questions without seeing which system produced the answer; if the human-rated correctness gap between the graph agent and the vector baseline is far smaller than 0.83 vs 0.14, or if re-running the LLM judge with different formatting instructions reverses the ordering, the central claim collapses.

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

If this is right

  • Multi-hop and temporal questions over registries—such as all companies connected to a given person or top companies by capital in a city—become answerable in natural language, not just exact-match lookups.
  • A deterministic identity layer plus a tool-restricting router makes each answer traceable to concrete graph evidence, so outputs can be audited in compliance-sensitive settings.
  • The same ingestion and agent design transfers to other gazettes and public registries; the Swiss case is harder than most due to three official languages and decentralized cantonal formats.
  • Graph traversal requires on average 1.7 tool calls per query at about 10–12 seconds, versus about 3.5 seconds for the vector baseline—accuracy gains come at a latency cost that may matter in interactive use.
  • Entity-resolution precision of 97.15% on sampled hubs means alias merging is conservative; errors concentrate on abbreviated legal suffixes and middle initials, keeping false merges low.

Where Pith is reading between the lines

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

  • The judge-strictness examples the paper quotes suggest the absolute scores understate true answer quality, but they also mean the headline 0.83-vs-0.14 gap may partly reflect formatting and verbosity preferences rather than pure factual capability; a human-scored audit would disentangle these.
  • The paper attributes gains to specific components (bounded reflection, intent routing, LLM enrichment) via ablations, but the body reports no ablation numbers; those attribution claims are therefore not yet independently verifiable from this text.
  • A testable extension: apply the same pipeline to a second registry, such as a single-language gazette, and check whether the correctness gap persists; the architecture's own logic predicts it will be at least as large, because the hard parts—alias resolution and unstructured text—are simpler there.
  • If the judge-sensitivity issue is systematic, agentic systems in this domain need evaluation protocols that separate substantive correctness from formatting, perhaps by resetting the judge prompt to ignore verbosity.

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. The paper proposes an agentic GraphRAG architecture for analyzing the Swiss Official Gazette of Commerce (SHAB). It builds a Neo4j knowledge graph from structured fields and LLM-extracted weak nodes, resolves identity through a deterministic NameHub tokenization layer, and provides an analytical agent with intent routing, bounded reflection, and a strict state-machine response synthesizer, plus an audit dashboard. The evaluation uses four tiers: entity-resolution precision, agent tool trajectories, RAGAS-style answer-quality metrics judged by GPT-5, and multi-turn conversational metrics. The central claim is that the graph-based agent substantially outperforms a standard vector-RAG baseline, with reported golden-set Correctness of 0.828 vs 0.143 and automated-set Answer Relevance of 0.689 vs 0.080.

Significance. If the headline claim is validated, the paper would make a useful contribution to expert-oriented public-registry analytics: a modular, auditable GraphRAG pipeline with deterministic identity resolution, an explicit state machine for tool use, and a transparent dashboard. The engineering is substantial, and the authors deserve credit for releasing code, documenting the exact LLM call parameters, and publishing the judge's failure cases in the appendix. The multi-tier evaluation protocol is a useful template. However, the core comparative evidence currently rests on a single weak dense baseline and on GPT-5-as-judge scores that the paper itself shows to be unreliable. The abstract also claims baselines, ablations, and a performance number that do not appear in the body. These issues are load-bearing for the central claim, though they are fixable.

major comments (4)
  1. [§5.4–5.5, Tables 4, 11–12] The headline comparison is produced entirely by GPT-5-as-judge. The paper itself documents severe judge failures: a correct answer identifying the new city scored 0.10 Answer Relevance (Table 11), and an answer correctly capturing a bankruptcy event but reporting the effective date instead of the publication date scored 0.00 Information Recall (Table 12). §6 also reports that excluding the 10 dissolution questions raises golden Correctness from 0.828 to 0.940, a ~0.11 swing driven by the judge's date-format strictness. Because the GraphRAG agent's state-machine output is verbose Markdown with disclaimers while the baseline is terse, a style-correlated judge bias will not cancel merely because the same judge is used for both systems. The Correctness prompt in Table 9 explicitly tells the judge to be tolerant to formatting, yet the documented examples show the judge is not. A human-scored
  2. [Abstract and §5.1] The abstract claims the complete architecture is compared with dense, lexical, and hybrid flat-retrieval baselines and with controlled architectural ablations, and that ablation results show bounded reflection improves answer quality. The body contains no such experiments: §5.1 defines a single all-MiniLM-L6-v2 top-5 dense baseline as the only comparator, and §6 reports no lexical/hybrid baselines and no ablations anywhere. The abstract's 'increases factual correctness from 0.26 ... to 0.83' is also not traceable to Table 4, which reports 0.143→0.828 for the golden benchmark. Either the missing experiments should be added or the abstract must be corrected to state what was actually evaluated. This discrepancy matters because the attribution of gains to specific architectural components is currently unsupported.
  3. [§5, automated dataset; §6] The automated 300-question dataset is generated by 'Reverse RAG': complex Cypher queries extract subgraphs from the Neo4j graph, and GPT-5 synthesizes questions and expected answers from those subgraphs. Evaluating the system on questions derived from the same graph it retrieves from is self-referential; it tests internal consistency rather than generalizable QA ability. The only independent benchmark is the 60-question golden set, and §6 shows that excluding 10 dissolution questions changes headline Correctness from 0.828 to 0.940. After that exclusion, the central quantitative claim rests on 50 manually curated questions scored by the same unreliable judge. The paper should report human scores for the golden set, per question and with uncertainty intervals, and should present the automated set only as a trajectory/coverage diagnostic.
  4. [§5.1, Table 4] The baseline is a flat dense retriever using all-MiniLM-L6-v2 with top-5 chunks over the full SHAB corpus. This is a weak comparator for multi-hop, entity-centric QA: there is no lexical or hybrid retrieval, no reranking, no larger embedding model, and a top-k of 5 cannot plausibly cover multi-hop joins over millions of records. The reported 0.69-point gap may therefore reflect the underpowered baseline rather than an intrinsic advantage of graph-mediated retrieval. A stronger flat baseline (e.g., BM25 combined with dense retrieval, higher k, reranking, or a more capable embedding model) is needed, together with baseline retrieval recall on the golden set, to support the paper's comparative conclusion.
minor comments (4)
  1. [§2.5] The phrase 'in a way that has never been possible before' is an overclaim for a literature that already contains agentic GraphRAG systems (e.g., [41, 42]). Rephrase to describe the specific combination of registry data, deterministic identity resolution, and state-machine control.
  2. [Tables 3–5] Point estimates are reported without confidence intervals or significance tests. Given the small golden set (N=60), the exclusion of 10 questions, and the stochasticity of LLM-based components, bootstrap intervals or per-question scores would help the reader assess robustness.
  3. [Figure 1] Typographical issues: 'Gazzette' should be 'Gazette', and 'creditators' should be 'creditors' in the figure text.
  4. [§5.4] Information Recall is defined as 'the fraction of the expected answer’s core information that the agent’s response successfully captures,' but the scoring prompt in Table 9 does not specify how core information is decomposed. Since this metric drives several conclusions, a deterministic or human-annotated decomposition for at least the golden set would strengthen the methodology.

Circularity Check

1 steps flagged

Tier 1 entity-resolution precision is self-confirming: the 'true positive' test uses the same sorted-token match that defines NameHub merges; the headline GraphRAG-vs-RAG comparison itself is not circular.

specific steps
  1. self definitional [Section 5.2 (Tier 1 evaluation procedure, classification rule after Eq. 1); cf. Section 4.1.4 (Phase 3 identity resolution)]
    "A merge is classified as a true positive if either the similarity score satisfies S(A, B)>0.6 or the alphabetically sorted token sets of the two strings match exactly. This second criterion is important because it correctly captures legal formatting variations such as comma-first name order (e.g., 'Doe, John' versus 'John Doe') and related token-order permutations..."

    The merge rule that creates a NameHub is byte-identical sorted-token keys: 'Nodes are linked to the same central NameHub if the resulting normalized key is identical, byte-for-byte' where generate_hub_key 'lowercases the text, strips non-alphanumeric characters, splits the string into discrete tokens, sorts them alphabetically, and concatenates them.' Therefore any two names merged into the same hub have, by construction, exactly matching sorted token sets. Tier 1's second 'true positive' criterion applies exactly that same sorted-token-set match to the same names, so the reported 97.15% precision is guaranteed by the construction of the hub, not independently verified.

full rationale

I walked the paper's claimed validation chain. The only step that reduces to its own input is Tier 1: the identity-resolution precision is computed on NameHub nodes whose merges were produced by generate_hub_key, and the evaluation's true-positive definition includes exactly the same sorted-token-set match that defines the hub key equality. Thus the 97.15% merge precision is self-confirming rather than an independent audit of entity resolution. This is a real, quotable circularity, but it is contained to a supporting component claim. The headline result—agentic GraphRAG outperforming vector RAG on answer quality—rests mainly on the manually curated golden benchmark (Tier 3) and conversational benchmark (Tier 4), which use human-written reference answers and a fixed vector baseline; those numbers are not fitted parameters or derived from the system's own outputs by construction. The automated dataset is generated from the same Neo4j graph and judged by GPT-5, which is a legitimate evaluation-validity concern (judge calibration, formatting sensitivity), but it is not an equivalence-by-construction of the claimed capability difference. The self-citations in the reference list are background and not load-bearing for the central claim. Overall score 4: one clearly circular evaluation step, while the central empirical comparison has independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 1 invented entities

The system's load-bearing input assumptions are: (1) SHAB structured metadata is true ground truth; (2) sorted-token byte equality is a safe identity-equivalence rule; (3) weak nodes sharing a hub and an event with a strong node are duplicates; (4) gpt-4o-mini extraction at 2,000-char truncation over six hand-picked sub-rubrics captures latent actors; (5) GPT-5 judge scores are valid quality measures — an axiom the paper's own §6 partially contradicts. Free parameters are engineering choices (sub-rubric filter, truncation length, reflection bound, top-k, similarity threshold) rather than fitted model constants, so the circularity burden is moderate, concentrated in the self-derived benchmark rather than in numerical fitting.

free parameters (5)
  • LLM-extraction sub-rubric filter = HR01, KK02, KK03, KK06, LS01, LS02
    Phase 2 (§4.1.3) extracts weak actors only from these six sub-rubrics (~541k events of 2.18M). Events outside this set get no LLM extraction, which bounds what the graph can answer.
  • Event text truncation length = 2000 characters
    Phase 2 truncates full_text to 2,000 chars (§4.1.3), so actors mentioned later in long legal notices cannot be extracted. Hand-chosen to optimize API cost/context.
  • Tier 1 true-positive similarity threshold = S > 0.6
    Evaluation criterion in Eq. 1 (§5.2); a hand-chosen threshold defining merge precision. Changing it changes the reported 97.15%.
  • Agent reflection loop bound = 4 iterations
    Max tool iterations (§4.2.2, Table 1); bounds ARS=1.7 and latency 12.5s. A different bound would change both cost and correctness.
  • Baseline top-k = 5 chunks
    Vector baseline retrieves top-5 chunks (§5.1); no sweep over k is reported, a material configuration for a fairness comparison.
axioms (6)
  • domain assumption Structured SHAB metadata fields are verified ground truth
    Phase 1 treats registry metadata (UID, legal form, address) as authoritative 'strong nodes' (§4.1.2).
  • domain assumption Byte-exact alphabetical tokenization is a valid identity-resolution equivalence
    generate_hub_key merges any two name strings with identical sorted-token keys (§4.1.4); conservative, but merges distinct people sharing identical name tokens and splits middle-initial variants.
  • domain assumption A weak node sharing a NameHub and an Event with a strong node is a duplicate
    Weak-node absorption (§4.1.4) deletes such weak nodes; false if two distinct persons with the same name appear in the same event.
  • domain assumption gpt-4o-mini at T=0 with enforced JSON performs reliable NER on registry text
    Phase 2 (§4.1.3) depends on this for all 718,736 weak nodes; no manual accuracy audit of the extraction is reported.
  • domain assumption GPT-5 LLM-as-a-judge scores approximate expert quality judgments
    Tiers 3-4 (§5.4-5.5) rest on this; §6 documents the judge as systematically harsh, penalizing correct answers to 0.10/0.00, so the paper's own analysis partially contradicts this axiom.
  • domain assumption Reverse-RAG synthetic questions and answers derived from the graph are valid ground truth
    The automated dataset (§5) generates Q/A from Cypher-extracted subgraphs; its validity inherits the correctness of the graph itself, which is exactly what is under test.
invented entities (1)
  • NameHub identity-resolution node independent evidence
    purpose: Abstract bridging node linking orthographic variants of the same company or person to a single canonical identity for stable graph traversal.
    Introduced in §4.1.3/4.1.4; an engineering abstraction rather than a physical entity. Its key is deterministic and recomputable from public registry strings, giving an external falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 25842 in / 22539 out tokens · 168385 ms · 2026-08-02T16:14:37.907274+00:00 · methodology

0 comments
read the original abstract

Public commercial registries are formally open, yet their practical analysis remains difficult because relevant facts are scattered across millions of records that combine structured metadata, multilingual legal notices, temporal events, and entity aliases. This paper presents a controlled, tool-mediated agentic GraphRAG architecture for auditable natural-language analysis of such registries. The proposed pipeline transforms publications from the Swiss Official Gazette of Commerce into a Neo4j knowledge graph comprising over five million nodes and 4.7 million relationships. It combines deterministic ingestion of structured registry fields, LLM-assisted extraction of latent actors from unstructured notices, and a deterministic identity-resolution layer. An analytical agent operates on this graph through intent routing, restricted graph tools, bounded reflection, and state-machine-guided response synthesis. We evaluate the system using a multi-tier protocol covering answer quality, retrieval behavior, entity resolution, and multi-turn conversational performance. The complete architecture is compared with dense, lexical, and hybrid flat-retrieval baselines and with controlled architectural ablations. On a manually curated benchmark, graph-mediated retrieval increases factual correctness from 0.26 for the strongest flat-retrieval baseline to 0.83 for the complete system, with comparable improvements in relevance and completeness. Ablation results show that bounded reflection improves answer quality while intent routing and LLM-based graph enrichment improve reliability in difficult entity resolution tasks. An exploratory dashboard displays the graph evidence and execution traces underlying each response, allowing users to inspect how answers were produced.

Figures

Figures reproduced from arXiv: 2605.18770 by Arthur Capozzi, Dirk Helbing.

Figure 1
Figure 1. Figure 1: Data ingestion pipeline for constructing the SHAB knowledge graph. Raw SHAB data combines unstructured [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the analytical agent. The user query is first processed by a zero-shot intent router that restricts [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Exploratory Dashboard operating in global search mode. The left navigation panel displays deterministic [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The Exploratory Dashboard operating in Dossier Mode. Since the user selected an entity in the left sidebar, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The central dossier panel of the Exploratory Dashboard renders a 2D force-directed network to visualize [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Towards Trustworthy and Cost-Efficient Data Integration: From Na\"ive RAG to Agentic RAG

    cs.DB 2026-07 conditional novelty 4.0

    The paper argues that agentic RAG with adaptive retrieval, iterative reasoning, and graph memory is the path to trustworthy and cost-efficient data integration, and sketches a six-agent architecture.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Agentic retrieval-augmented generation: A survey on agentic rag.arXiv preprint arXiv:2501.09136, 2025

    Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag.arXiv preprint arXiv:2501.09136, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    From llm reasoning to autonomous ai agents: A comprehensive review, 2026

    Mohamed Amine Ferrag, Norbert Tihanyi, and Merouane Debbah. From llm reasoning to autonomous ai agents: A comprehensive review, 2026

    2026

  3. [3]

    Structugraphrag: Structured document- informed knowledge graphs for retrieval-augmented generation.Proceedings of the AAAI Symposium Series, 4(1): 242–251, Nov

    Xishi Zhu, Xiaoming Guo, Shengting Cao, Shenglin Li, and Jiaqi Gong. Structugraphrag: Structured document- informed knowledge graphs for retrieval-augmented generation.Proceedings of the AAAI Symposium Series, 4(1): 242–251, Nov. 2024. doi: 10.1609/aaaiss.v4i1.31798

    doi:10.1609/aaaiss.v4i1.31798 2024

  4. [4]

    T-grag: A dynamic graphrag framework for resolving temporal conflicts and redundancy in knowledge retrieval

    Dong Li, Yichen Niu, Ying Ai, Xiang Zou, Biqing Qi, and Jianxing Liu. T-grag: A dynamic graphrag framework for resolving temporal conflicts and redundancy in knowledge retrieval. InProceedings of the 33rd ACM International Conference on Multimedia, MM ’25, page 11880–11889, New York, NY , USA, 2025. Association for Computing Machinery. ISBN 9798400720352....

    arXiv 2025

  5. [5]

    Open data and transparency: Opportunities and challenges in the swiss context.Public Performance & Management Review, 43(3):662–686, 2020

    Tereza Cahlikova and Vincent Mabillard. Open data and transparency: Opportunities and challenges in the swiss context.Public Performance & Management Review, 43(3):662–686, 2020. doi: 10.1080/15309576.2019. 1657914

    arXiv 2020

  6. [6]

    Financial risk management and explainable, trustworthy, responsible ai.Frontiers in Artificial Intelligence, V olume 5 - 2022, 2022

    Sebastian Fritz-Morgenthal, Bernhard Hein, and Jochen Papenbrock. Financial risk management and explainable, trustworthy, responsible ai.Frontiers in Artificial Intelligence, V olume 5 - 2022, 2022. ISSN 2624-8212. doi: 10.3389/frai.2022.779799

    arXiv 2022

  7. [7]

    The role of explainable ai in enhancing trust and decision-making in financial services.Journal of Applied Finance & Banking, 15(5):49–62, 2025

    Ian Staley. The role of explainable ai in enhancing trust and decision-making in financial services.Journal of Applied Finance & Banking, 15(5):49–62, 2025

    2025

  8. [8]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Sys...

    2023

  9. [9]

    A survey on llm-as-a-judge.The Innovation, page 101253, 2026

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Zhouchi Lin, Bowen Zhang, Lionel Ni, Wen Gao, Yuanzhuo Wang, and Jian Guo. A survey on llm-as-a-judge.The Innovation, page 101253, 2026. ISSN 2666-6758. doi: https://doi.org/10.1016/j.xinn.2025.101253

    arXiv 2026

  10. [10]

    Effective automatic feature engineering on financial statements for bankruptcy prediction

    Xinlin Wang, Zsófia Kräussl, Maciej Zurad, and Mats Brorsson. Effective automatic feature engineering on financial statements for bankruptcy prediction. In2023 3rd International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), pages 1–8, 2023. doi: 10.1109/ICECCME57830. 2023.10252608

    arXiv 2023

  11. [11]

    Data mining techniques for the detection of fraudulent financial statements.Expert Systems with Applications, 32(4):995–1003, 2007

    Efstathios Kirkos, Charalambos Spathis, and Yannis Manolopoulos. Data mining techniques for the detection of fraudulent financial statements.Expert Systems with Applications, 32(4):995–1003, 2007. ISSN 0957-4174. doi: https://doi.org/10.1016/j.eswa.2006.02.016

    doi:10.1016/j.eswa.2006.02.016 2007

  12. [12]

    Bankruptcy prediction using case-based reasoning, neural networks, and discriminant analysis.Expert Systems with Applications, 13(2):97–108, 1997

    Hongkyu Jo, Ingoo Han, and Hoonyoung Lee. Bankruptcy prediction using case-based reasoning, neural networks, and discriminant analysis.Expert Systems with Applications, 13(2):97–108, 1997. ISSN 0957-4174. doi: https://doi.org/10.1016/S0957-4174(97)00011-0

    doi:10.1016/s0957-4174(97)00011-0 1997

  13. [13]

    Feature-weighted counterfactual-based explanation for bankruptcy prediction

    Soo Hyun Cho and Kyung shik Shin. Feature-weighted counterfactual-based explanation for bankruptcy prediction. Expert Systems with Applications, 216:119390, 2023. ISSN 0957-4174. doi: https://doi.org/10.1016/j.eswa.2022. 119390

    doi:10.1016/j.eswa.2022 2023

  14. [14]

    Jiaming Liu, Chengzhang Li, Peng Ouyang, Jiajia Liu, and Chong Wu. Interpreting the prediction results of the tree-based gradient boosting models for financial distress prediction with an explainable machine learning approach.Journal of Forecasting, 42(5):1112–1137, 2023. doi: https://doi.org/10.1002/for.2931

    doi:10.1002/for.2931 2023

  15. [15]

    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

  16. [16]

    Evaluating retrieval-augmented generation models for financial report question and answering.Applied Sciences, 14(20), 2024

    Ivan Iaroshev, Ramalingam Pillai, Leandro Vaglietti, and Thomas Hanne. Evaluating retrieval-augmented generation models for financial report question and answering.Applied Sciences, 14(20), 2024. ISSN 2076-3417. doi: 10.3390/app14209318

    doi:10.3390/app14209318 2024

  17. [17]

    Finsage: A multi-aspect rag system for financial filings question answering

    Xinyu Wang, Jijun Chi, Zhenghan Tai, Tung Sum Thomas Kwok, Hailin He, Zhuhong Li, Yuchen Hua, Muzhi Li, Peng Lu, Suyucheng Wang, Yihong Wu, Huang Jerry, Jingrui Tian, Fengran Mo, Yufei Cui, and Ling Zhou. Finsage: A multi-aspect rag system for financial filings question answering. InProceedings of the 34th ACM International Conference on Information and K...

    arXiv 2025

  18. [18]

    Hierfinrag—hierarchical multimodal rag for financial document understanding.Informatics, 13(2), 2026

    Quang-Vinh Dang, Ngoc-Son-An Nguyen, and Thi-Bich-Diem V o. Hierfinrag—hierarchical multimodal rag for financial document understanding.Informatics, 13(2), 2026. ISSN 2227-9709. doi: 10.3390/informatics13020030

    doi:10.3390/informatics13020030 2026

  19. [20]

    Mengshoel, and Xindong Wu

    Xiaoyi Fu, Xinqi Ren, Ole J. Mengshoel, and Xindong Wu. Stochastic optimization for market return prediction using financial knowledge graph. In2018 IEEE International Conference on Big Knowledge (ICBK), pages 25–32,

  20. [21]

    Knowledge graph-based event embedding framework for financial quantitative investments

    Dawei Cheng, Fangzhou Yang, Xiaoyang Wang, Ying Zhang, and Liqing Zhang. Knowledge graph-based event embedding framework for financial quantitative investments. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, page 2221–2230, New York, NY , USA, 2020. Association for Computing Ma...

    arXiv 2020

  21. [22]

    Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial forensics.arXiv preprint arXiv:1908.02591, 2019

    Mark Weber, Giacomo Domeniconi, Jie Chen, Daniel Karl I Weidele, Claudio Bellei, Tom Robinson, and Charles E Leiserson. Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial forensics.arXiv preprint arXiv:1908.02591, 2019

    Pith/arXiv arXiv 1908

  22. [23]

    Flowseries: flow analysis on financial networks.Applied Network Science, 10(1):28, 2025

    Arthur Capozzi, Salvatore Vilella, Dario Moncalvo, Marco Fornasiero, Valeria Ricci, Silvia Ronchiadin, and Giancarlo Ruffo. Flowseries: flow analysis on financial networks.Applied Network Science, 10(1):28, 2025. doi: 10.1007/s41109-025-00711-0

    doi:10.1007/s41109-025-00711-0 2025

  23. [24]

    Exploiting graph metrics to detect anomalies in cross-country money transfer temporal networks

    Salvatore Vilella, Arthur Thomas Edward Capozzi Lupi, Giancarlo Ruffo, Marco Fornasiero, Dario Moncalvo, Valeria Ricci, and Silvia Ronchiadin. Exploiting graph metrics to detect anomalies in cross-country money transfer temporal networks. InCompanion Proceedings of the ACM Web Conference 2023, WWW ’23 Companion, page 1245–1248, New York, NY , USA, 2023. A...

    arXiv 2023

  24. [25]

    Weirdnodes: centrality based anomaly detection on temporal networks for the anti-financial crime domain.Applied Network Science, 10(1):14, 2025

    Salvatore Vilella, Arthur Capozzi, Marco Fornasiero, Dario Moncalvo, Valeria Ricci, Silvia Ronchiadin, and Giancarlo Ruffo. Weirdnodes: centrality based anomaly detection on temporal networks for the anti-financial crime domain.Applied Network Science, 10(1):14, 2025. doi: 10.1007/s41109-025-00702-1

    doi:10.1007/s41109-025-00702-1 2025

  25. [26]

    Temporal relational ranking for stock prediction.ACM Trans

    Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua. Temporal relational ranking for stock prediction.ACM Trans. Inf. Syst., 37(2), March 2019. ISSN 1046-8188. doi: 10.1145/3309547

    doi:10.1145/3309547 2019

  26. [27]

    Beyond the tax haven: a graph analysis of business attraction in swiss municipalities.EPJ Data Science, 15, 2026

    Arthur Capozzi and Damian Dailisan. Beyond the tax haven: a graph analysis of business attraction in swiss municipalities.EPJ Data Science, 15, 2026. doi: 10.1140/epjds/s13688-026-00619-4

    doi:10.1140/epjds/s13688-026-00619-4 2026

  27. [28]

    Modeling momentum spillover with economic links discovered from financial documents

    Andy Chung and Kumiko Tanaka-Ishii. Modeling momentum spillover with economic links discovered from financial documents. InProceedings of the Fourth ACM International Conference on AI in Finance, ICAIF ’23, page 490–497, New York, NY , USA, 2023. Association for Computing Machinery. ISBN 9798400702402. doi: 10.1145/3604237.3626862

    arXiv 2023

  28. [29]

    In-depth analysis of graph-based rag in a unified framework.Proc

    Yingli Zhou, Yaodong Su, Youran Sun, Shu Wang, Taotao Wang, Runyuan He, Yongwei Zhang, Sicong Liang, Xilin Liu, Yuchi Ma, and Yixiang Fang. In-depth analysis of graph-based rag in a unified framework.Proc. VLDB Endow., 18(13):5623–5637, January 2026. ISSN 2150-8097. doi: 10.14778/3773731.3773738

    arXiv 2026

  29. [30]

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

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. 2025

    2025

  30. [31]

    GRAG: Graph retrieval-augmented generation

    Yuntong Hu, Zhihan Lei, Zheng Zhang, Bo Pan, Chen Ling, and Liang Zhao. GRAG: Graph retrieval-augmented generation. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors,Findings of the Association for Computa- tional Linguistics: NAACL 2025, pages 4145–4157, Albuquerque, New Mexico, April 2025. Association for Computational Linguistics. ISBN 979-8-89176-19...

    doi:10.18653/v1/2025.findings-naacl.232 2025

  31. [32]

    Agentai: A comprehensive survey on autonomous agents in distributed ai for industry 4.0.Expert Systems with Applications, 291:128404, 2025

    Francesco Piccialli, Diletta Chiaro, Sundas Sarwar, Donato Cerciello, Pian Qi, and Valeria Mele. Agentai: A comprehensive survey on autonomous agents in distributed ai for industry 4.0.Expert Systems with Applications, 291:128404, 2025. ISSN 0957-4174. doi: https://doi.org/10.1016/j.eswa.2025.128404

    arXiv 2025

  32. [33]

    A review of agentic artificial intelligence: Power of self-driven ai in the future of financial autonomy and enhanced customer engagement

    Aparna Krishna Bhat and Gokulram Krishnan. A review of agentic artificial intelligence: Power of self-driven ai in the future of financial autonomy and enhanced customer engagement. In2025 3rd International Conference on Sustainable Computing and Data Communication Systems (ICSCDS), pages 1160–1165, 2025. doi: 10.1109/ ICSCDS65426.2025.11167368

    arXiv 2025

  33. [34]

    The rise of agentic ai in finance: Opportunities, risks, and human-centric integration.IT Professional, 27(4):19–24, 2025

    Nir Kshetri. The rise of agentic ai in finance: Opportunities, risks, and human-centric integration.IT Professional, 27(4):19–24, 2025. doi: 10.1109/MITP.2025.3585227

    arXiv 2025

  34. [35]

    Elgendy, Mohamed Y .I

    Ibrahim A. Elgendy, Mohamed Y .I. Helal, Mohammed A. Al-Sharafi, Mousa Ahmed Albashrawi, Mohammad S. Al-Ahmadi, Il Jeon, and Yogesh K. Dwivedi. Agentic systems as catalysts for innovation in fintech: exploring opportunities, challenges and a research agenda.Information Discovery and Delivery, 05 2025. ISSN 2398-6247. doi: 10.1108/IDD-03-2025-0068

    doi:10.1108/idd-03-2025-0068 2025

  35. [36]

    Ai agents in finance and fintech: A scientific review of agent-based systems, applications, and future horizons.Computers, Materials and Continua, 86(1):1–34, 2025

    Maryan Rizinski and Dimitar Trajanov. Ai agents in finance and fintech: A scientific review of agent-based systems, applications, and future horizons.Computers, Materials and Continua, 86(1):1–34, 2025. ISSN 1546-2218. doi: https://doi.org/10.32604/cmc.2025.069678. 27 Agentic GraphRAG

    arXiv 2025

  36. [37]

    Enhancing investment analysis: Optimizing ai-agent collaboration in financial research

    Xuewen Han, Neng Wang, Shangkun Che, Hongyang Yang, Kunpeng Zhang, and Sean Xin Xu. Enhancing investment analysis: Optimizing ai-agent collaboration in financial research. InProceedings of the 5th ACM International Conference on AI in Finance, ICAIF ’24, page 538–546, New York, NY , USA, 2024. Association for Computing Machinery. ISBN 9798400710810. doi: ...

    arXiv 2024

  37. [38]

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

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-RAG: Learning to retrieve, gen- erate, and critique through self-reflection. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  38. [39]

    RAG-critic: Leveraging automated critic-guided agentic workflow for retrieval augmented generation

    Guanting Dong, Jiajie Jin, Xiaoxi Li, Yutao Zhu, Zhicheng Dou, and Ji-Rong Wen. RAG-critic: Leveraging automated critic-guided agentic workflow for retrieval augmented generation. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors,Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (...

  39. [40]

    Vincent Poor

    Ruichen Zhang, Hongyang Du, Yinqiu Liu, Dusit Niyato, Jiawen Kang, Sumei Sun, Xuemin Shen, and H. Vincent Poor. Interactive ai with retrieval-augmented generation for next generation networking.IEEE Network, 38(6): 414–424, 2024. doi: 10.1109/MNET.2024.3401159

    arXiv 2024

  40. [41]

    Hm-rag: Hierarchical multi-agent multimodal retrieval augmented generation

    Pei Liu, Xin Liu, Ruoyu Yao, Junming Liu, Siyuan Meng, Ding Wang, and Jun Ma. Hm-rag: Hierarchical multi-agent multimodal retrieval augmented generation. MM ’25, page 2781–2790, New York, NY , USA, 2025. Association for Computing Machinery. ISBN 9798400720352. doi: 10.1145/3746027.3754761

    arXiv 2025

  41. [42]

    Ka-rag: Integrating knowledge graphs and agentic retrieval- augmented generation for an intelligent educational question-answering model.Applied Sciences, 15(23), 2025

    Fangqun Gao, Shu Xu, Weiyan Hao, and Tao Lu. Ka-rag: Integrating knowledge graphs and agentic retrieval- augmented generation for an intelligent educational question-answering model.Applied Sciences, 15(23), 2025. ISSN 2076-3417. doi: 10.3390/app152312547

    doi:10.3390/app152312547 2025

  42. [43]

    Keying hash functions for message authentication

    Mihir Bellare, Ran Canetti, and Hugo Krawczyk. Keying hash functions for message authentication. In Neal Koblitz, editor,Advances in Cryptology — CRYPTO ’96, pages 1–15, Berlin, Heidelberg, 1996. Springer Berlin Heidelberg. ISBN 978-3-540-68697-2

    1996

  43. [44]

    Cypher: An evolving query language for property graphs

    Nadime Francis, Alastair Green, Paolo Guagliardo, Leonid Libkin, Tobias Lindaaker, Victor Marsault, Stefan Plantikow, Mats Rydberg, Petra Selmer, and Andrés Taylor. Cypher: An evolving query language for property graphs. InProceedings of the 2018 International Conference on Management of Data, SIGMOD ’18, page 1433–1445, New York, NY , USA, 2018. Associat...

    arXiv 2018

  44. [45]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InNeurIPS 2022 Foundation Models for Decision Making Workshop, 2022

    2022

  45. [46]

    Where llm agents fail and how they can learn from failures.arXiv preprint arXiv:2509.25370, 2025

    Kunlun Zhu, Zijia Liu, Bingxuan Li, Muxin Tian, Yingxuan Yang, Jiaxun Zhang, Pengrui Han, Qipeng Xie, Fuyang Cui, Weijia Zhang, et al. Where llm agents fail and how they can learn from failures.arXiv preprint arXiv:2509.25370, 2025

    arXiv 2025

  46. [47]

    Siva Reddy, Danqi Chen, and Christopher D. Manning. Coqa: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249–266, 05 2019. ISSN 2307-387X. doi: 10.1162/tacl_a_00266

    doi:10.1162/tacl_a_00266 2019

  47. [48]

    QuAC: Question answering in context

    Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. QuAC: Question answering in context. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii, editors,Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2174–2184, Brussels, Belgium, October-No...

    doi:10.18653/v1/d18-1241 2018

  48. [49]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 5776–5788, 2020

    2020

  49. [50]

    RAGAs: Automated evaluation of retrieval augmented generation

    Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. RAGAs: Automated evaluation of retrieval augmented generation. In Nikolaos Aletras and Orphee De Clercq, editors,Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 150–158, St. Julians, Malta, March 202...

    doi:10.18653/v1/2024.eacl-demo.16 2024

  50. [2018]

    26 Agentic GraphRAG

    doi: 10.1109/ICBK.2018.00012. 26 Agentic GraphRAG

    arXiv 2018

  51. [2025]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long.179

    doi:10.18653/v1/2025.acl-long.179 2025