Pith. sign in

REVIEW 4 major objections 4 minor 47 references

The paper argues that LLM-based data integration becomes trustworthy, scalable, and cost-efficient only when it moves from static retrieval-augmented generation to agentic RAG, where agents decide on demand whether and what to retrieve.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:07 UTC pith:5SK7DPVA

load-bearing objection A clear, honest vision paper that maps RAG variants onto data integration and sketches an agentic architecture; just don't mistake the roadmap for evidence. the 4 major comments →

arxiv 2607.22319 v1 pith:5SK7DPVA submitted 2026-07-24 cs.DB cs.AI

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

Chuangtao Ma , Arijit Khan This is my paper
classification cs.DB cs.AI
keywords agentic RAGdata integrationknowledge-grounded LLMstrustworthinessschema matchingentity matchingcost-efficient retrievalgraph-based memory
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.

Large language models cannot deliver reliable data integration on their own, this paper argues: their closed-world parametric knowledge produces both missed matches and hallucinated ones, and pairwise comparison at enterprise scale is prohibitively expensive. The authors' central proposition is that integration should be rebuilt around knowledge-grounded generation, moving from naive retrieval-augmented generation (RAG) through graph-structured variants to agentic RAG, in which autonomous agents decide when to retrieve, what granularity of evidence to fetch, and whether to iterate. This shift, they contend, makes decisions evidence-grounded and verifiable—what the paper defines as trustworthiness—while adaptive retrieval, batch processing, and persistent graph-based memory keep cost under control. The paper contributes a vision, a concrete multi-agent architecture, and a list of open problems that still stand between the proposal and enterprise deployment.

Core claim

On the paper's own terms, the central claim is diagnostic and prescriptive: the persistent knowledge gap that makes LLM-based integration unreliable—models matching on static parametric memory, missing domain-specific matches and hallucinating spurious ones—should be closed not by more fine-tuning or more demonstrations but by grounding every integration decision in externally retrieved evidence. The prescribed cure is agentic RAG, which replaces compulsory per-query retrieval and single-pass chain-of-thought reasoning with adaptive, on-demand retrieval and iterative reasoning: an analyzer routes easy queries to direct inference and ambiguous ones to retrieval at variable granularity (text,

What carries the argument

The central mechanism is agentic RAG: a retrieval-augmented generation workflow in which a set of cooperating agents dynamically decide whether retrieval is needed, select the granularity of knowledge to fetch, and iterate between retrieval and reasoning until a decision is accepted. The cost argument rides on two supporting mechanisms: adaptive retrieval, which skips knowledge-base lookups for queries the analyzer judges easy with high confidence, and batch processing with persistent graph-based memory, which reuses shared column and schema context across many pairwise comparisons instead of retrieving per query. Together they replace static per-query RAG with a system whose cost scales wit

Load-bearing premise

The load-bearing premise is that confidence-based routing can tell easy queries from ambiguous ones and skip retrieval for easy queries without ever missing necessary evidence; if that premise fails—the paper itself flags in Section 5 that adaptive retrieval can fail under high query diversity or imbalanced ground-truth distributions—the cost savings disappear and accuracy drops.

What would settle it

Run the proposed adaptive-retrieval system on a schema-matching benchmark with deliberately imbalanced ground truth (for example, 95% non-matching pairs) and high query diversity, against an always-retrieve KG-RAG control. If the adaptive system shows a higher false-negative rate on queries it routed to direct inference, or if its cost savings approach zero because almost every pair is judged ambiguous, the central cost-accuracy trade-off is refuted. A calibration plot of LLM-as-judge confidence against actual match accuracy on easy-labeled queries would settle the premise directly.

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

If this is right

  • If the vision is realized, integration decisions become auditable: every accepted match can be traced to retrieved evidence, which is the paper's operational definition of trustworthiness.
  • Easy matching pairs no longer trigger expensive knowledge-base lookups, so cost scales with the number of ambiguous cases rather than with the total number of comparisons.
  • Batch retrieval and prompting exploit shared column profiles and overlapping schema attributes so that one retrieval and one inference call serve many pairs, cutting token consumption without sacrificing match quality.
  • Persistent graph-based agent memory replaces short-lived caches, so repeated or related integration tasks avoid cold-start retrieval and multi-hop evidence remains stable across queries.
  • Fine-tuning and large labeled datasets become less central to integration: external grounding, not supervision, becomes the primary lever for accuracy.

Where Pith is reading between the lines

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

  • The paper's trustworthiness definition is qualitative; a natural operationalization is to measure what fraction of accepted decisions are entailed by the retrieved evidence, and to test whether adaptive routing preserves that entailment rate while cutting cost.
  • The cost-efficiency claim is workload-dependent: in domains where nearly every candidate pair is ambiguous (dirty data, heavy abbreviation use), skipping retrieval saves little, so the benefit is concentrated in workloads with a long tail of easy pairs—an implicit boundary the paper does not quantify.
  • The iterative reject/re-plan loop effectively turns the knowledge base into an active-learning target: a testable extension is to mine rejected answers to update graph memory, so repeated failures become cheaper and faster on later queries.
  • LLM-as-judge confidence is the linchpin of the routing decision; replacing it with a calibrated uncertainty measure (such as token-level entropy) and comparing routing outcomes would make the proposal testable before full deployment.

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 is a vision/survey paper arguing that LLM-based data integration should move from parametric, fine-tuned, and in-context-learning paradigms to knowledge-grounded paradigms, specifically a trajectory from naive RAG through GraphRAG and KG-RAG to Agentic RAG. It proposes a multi-agent architecture (Analyzer, Planner, Retriever, Reasoning, Decision, Evaluator agents) that combines adaptive retrieval, iterative reasoning, batch processing, and graph-based persistent memory to achieve what the authors call trustworthy, scalable, and cost-efficient data integration. The paper also lists open research gaps, including memory conflict, batch-retrieval noise, parameter tuning, autonomous tool orchestration, and the lack of enterprise benchmarks for trustworthiness. The central claim is that Agentic RAG, with on-demand retrieval and adaptive granularity, overcomes the limitations of static RAG and enables cost-efficient, trustworthy integration.

Significance. If the proposed vision were realized and validated, it would provide a useful organizing framework for the growing literature on RAG-based data integration and a concrete architectural template for multi-agent systems in this domain. The paper is clearly structured, the running examples (MRN vs. Pat ID, measurement time vs. perfac date) are helpful, and the taxonomy from flat-text RAG to graph-based and agentic variants is a valuable framing. The paper also deserves credit for explicitly acknowledging several limitations and open problems in Section 5, and for not overselling the maturity of the field. However, the manuscript's title and abstract promise 'trustworthy and cost-efficient' integration as a near-term capability, while the body repeatedly states such outcomes as established facts. The load-bearing mechanisms—especially adaptive retrieval gating and LLM-as-judge confidence—are neither experimentally validated nor supported by a quantitative model. At present the paper is a research roadmap rather than a validated system; the central claim needs a substantial evidential or epistemic qualification before it can be accepted as a sound contribution.

major comments (4)
  1. [§4.2 / Figure 4 / §5] The cost-efficiency argument rests on the Analyzer Agent's ability to classify queries as 'easy' or 'ambiguous' via LLM-as-judge confidence and to skip retrieval for easy cases. No evidence is provided that such confidence is well-calibrated or that skipped retrieval does not omit necessary evidence. This is not a peripheral concern: the cost savings in Section 4.3 come precisely from avoiding per-query retrieval, while the trustworthiness gains come from grounding only ambiguous cases. Section 5 itself concedes that 'batch and adaptive retrieval may fail in scenarios with high query diversity or highly imbalanced ground-truth distributions, since the absence of a default retrieval policy can prevent necessary evidence from being retrieved.' That sentence directly undercuts the paper's claim that the proposed design achieves both trustworthiness and cost-efficiency. The authors must eith
  2. [§4.3] The statements that 'Agentic RAG reduces retrieval costs relative to static RAG and GraphRAG' and that persistent graph-based memory 'effectively address[es] latency and scalability challenges in agentic RAG systems for large-scale data integration' are made without comparative experiments, cost measurements, or latency data. These are central to the title's promise of cost-efficiency. The nearby discussion relies on recent or unpublished preprints (e.g., [12], [43], [44]) as if their results were established, but the manuscript does not report their settings or effect sizes, and [12] appears to be an arXiv preprint by the same authors. The paper should either present a concrete cost/latency analysis, clearly attribute each quantitative claim to a citable study with numbers, or replace categorical assertions with explicitly labeled hypotheses.
  3. [§4.4] The six-agent architecture is presented as a design that 'ensures that only verifiable, faithful, and trustworthy results are committed.' This is an unsupported functional claim. In particular, the Evaluator Agent uses LLM-as-judge confidence scores to approve or reject outputs, and the Decision Agent uses those scores to manage memory and re-planning, but no evidence is given that LLM judges are reliable enough for this role in data-integration tasks. Given that trustworthiness is defined as evidence-grounded, verifiable reasoning, the architecture needs at least a pilot study or a formal argument that the judge's confidence is calibrated and that the approve/reject loop does not systematically admit hallucinations or reject correct answers. Without such support, the statement should be softened to describe the intended behavior of the architecture, not its verified behavior.
  4. [§3.3 / Example 3] The KG-RAG example (beneficiary → customer → patient) is used to argue that external structured evidence resolves a specific schema-matching conflict between 'measurement time' and 'perfac date.' It is not clear whether this is a reported experimental result from KG-RAG4SM [20] or an illustrative reconstruction. If it is illustrative, the text should say so; if it is a reported result, the evaluation protocol and accuracy numbers should be cited or summarized. As written, the example functions as evidence for the superiority of KG-RAG without making the evidentiary basis explicit.
minor comments (4)
  1. [References] References [1] and [4] are the same Amalur paper and should be consolidated. Several other references to recent arXiv preprints are used as if peer-reviewed; please mark preprint status where appropriate.
  2. [Figures] Figure 4 contains the typo 'Retreival' (twice) and 'retreival' in the agent list. Figure 6 has awkward text overflow and a misspelling: 'exectue a spefic taks' should be 'execute a specific task.' Figure 3's caption and the surrounding text are repetitive ('it presents a vision and roadmap' appears almost verbatim twice).
  3. [§2.2] The claim that 'RAG reduces the need for fine-tuning and few-shot demonstrations, enabling integration tasks to be performed without extensive labeled datasets' is stated before the survey evidence is presented. Some support is given later, but a forward pointer to Section 3 would improve readability.
  4. [§4.1] The roadmap paragraph is repetitive and does not clearly distinguish the current state of the art from the authors' proposed future work. Consider separating 'what exists' from 'what we propose' with explicit markers.

Circularity Check

1 steps flagged

Trustworthiness is defined as retrieved-knowledge grounding, so the RAG-trustworthiness link is true by definition; otherwise the paper is a vision with independent content and only non-circular self-citations.

specific steps
  1. self definitional [Abstract (definition of trustworthiness); §4.1 Roadmap and Vision]
    "Here, trustworthiness refers to evidence-grounded, verifiable reasoning, where integration decisions are transparently supported by retrieved knowledge, robust against hallucination, and consistent across tasks. ... First, architectural advances in different RAG directly address hallucination and trustworthiness gaps by grounding LLMs with precise evidence chains for complex reasoning."

    The paper stipulatively defines trustworthiness as decisions being 'transparently supported by retrieved knowledge' — the defining operation of RAG. It then presents RAG's grounding of LLMs as 'directly address[ing] hallucination and trustworthiness gaps.' The claimed connection between RAG and trustworthiness is therefore entailed by the definition rather than established by evidence. The substantive question — whether retrieved evidence actually reduces hallucination and improves correctness — is assumed in the definition, making the 'trustworthy integration' contribution partly tautological.

full rationale

This is a vision/survey paper with no fitted parameters, no predictive equations, and no claimed derivation from first principles, so most of its argument is architectural and taxonomic rather than circular. The one clear reduction-by-definition is the trustworthiness framing: trustworthiness is defined as evidence-grounded reasoning 'transparently supported by retrieved knowledge,' and RAG is then credited with addressing trustworthiness by grounding LLMs in retrieved evidence. That makes the RAG–trustworthiness connection true by stipulation, not by empirical demonstration. The paper also relies on the authors' own prior work ([12], [20]) for empirical claims that KG-RAG and CE-RAG4EM improve accuracy and lower cost; these are load-bearing for the narrative but they are external preprints with independent experiments, and the roadmap also cites many outside works, so they do not by themselves constitute circularity. The §5 admission that batch and adaptive retrieval 'may fail' and can prevent necessary evidence from being retrieved is a genuine correctness risk for the cost-efficiency claim, but it is not a circularity. Overall, the central vision retains independent content in its multi-agent architecture, adaptive-retrieval roadmap, and research-gap analysis; the circularity is limited to the definitional framing of trustworthiness.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 6 invented entities

The paper contains no fitted parameters or quantitative derivations; it is a qualitative vision. The proposed architecture would require hyperparameters (batch size, confidence threshold, retrieval granularity) but none are specified or fitted. All listed axioms are domain assumptions about the benefits of knowledge grounding and agent orchestration, asserted from prior literature rather than demonstrated here.

axioms (6)
  • domain assumption RAG mitigates LLM hallucination and improves integration accuracy when grounded in retrieved context
    Invoked in §2.2 and §3.1 to justify the entire knowledge-grounded approach; supported only by prior citations, not by experiments in this paper.
  • domain assumption Knowledge grounding reduces the need for labeled data in integration tasks
    §2.2: 'Knowledge grounding alleviates dependence on curated examples.' This is a stated benefit with no direct evidence in this preprint.
  • domain assumption Graph-based persistent memory provides stable knowledge grounding at lower cost than repeated retrieval
    §4.3 asserts this, citing [34,43]; it is load-bearing for the cost-efficiency vision and not tested here.
  • domain assumption LLM-as-judge can reliably assess query complexity and answer confidence to route between direct and grounded inference
    The Analyzer and Evaluator agents in §4.4 depend on this; the paper's own §5 notes parameter tuning and adaptive retrieval policies are open problems.
  • domain assumption Multi-agent orchestration reduces latency and cost compared to linear single-agent workflows
    §4.3 'From Linear Agent Call to Multi-Agent Orchestration' asserts parallelization benefits without measurements.
  • domain assumption Enterprise integration queries can be separated into easy and ambiguous classes so that adaptive retrieval preserves accuracy
    Core to the adaptive-retrieval component; §5 acknowledges this can fail with high query diversity or imbalanced ground truth.
invented entities (6)
  • Analyzer Agent no independent evidence
    purpose: Assess query complexity and partition queries into batches
    Described in §4.4 Agent Layer; no implementation or evaluation exists.
  • Planner Agent no independent evidence
    purpose: Decompose tasks, schedule agents, and decide when retrieval is needed
    Described in §4.4 Agent Layer; no implementation or evaluation exists.
  • Retriever Agent no independent evidence
    purpose: Dynamically retrieve knowledge at varying granularity via MCP tools
    Described in §4.4 Agent Layer; no implementation or evaluation exists.
  • Reasoning Agent no independent evidence
    purpose: Generate answers via direct or knowledge-grounded inference
    Described in §4.4 Agent Layer; no implementation or evaluation exists.
  • Decision Agent no independent evidence
    purpose: Accept/reject answers and manage memory read/write
    Described in §4.4 Agent Layer; no implementation or evaluation exists.
  • Evaluator Agent no independent evidence
    purpose: Judge confidence and retrieval quality via LLM-as-judge
    Described in §4.4 Agent Layer; no implementation or evaluation exists.

pith-pipeline@v1.3.0-alltime-deepseek · 12653 in / 14452 out tokens · 111551 ms · 2026-08-01T05:07:02.565407+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) and AI agents have demonstrated strong potential for data integration in zero-shot and few-shot settings. However, they continue to face significant accuracy and cost challenges in enterprise environments due to a persistent knowledge gap. This paper envisions trustworthy, scalable, and cost-efficient integration through knowledge-grounded LLMs and agents operating within a retrieval-augmented generation (RAG) workflow. Here, trustworthiness refers to evidence-grounded, verifiable reasoning, where integration decisions are transparently supported by retrieved knowledge, robust against hallucination, and consistent across tasks. We trace the evolution from classic RAG to GraphRAG and KG-RAG (knowledge graph-based RAG), highlighting how these paradigms bridge parametric and contextual knowledge. Building on this trajectory, we explore the shift toward Agentic RAG, where autonomous multi-agent systems adaptively plan, retrieve, refine, and reason for complex integration tasks. We examine optimization strategies for cost-efficient integration, addressing computational bottlenecks in large-scale enterprise settings. Finally, we outline open challenges and future directions toward building reliable, explainable, and scalable knowledge-grounded integration systems.

Figures

Figures reproduced from arXiv: 2607.22319 by Arijit Khan, Chuangtao Ma.

Figure 1
Figure 1. Figure 1: Despite their strong reasoning capabilities, LLM-based integration suffers from three interrelated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: LLM for Data Integration. Hallucination and Class Imbalance. LLMs fre￾quently generate incorrect correspondences when re￾lying solely on internal parametric knowledge. In some cases, they miss true matches (false nega￾tives) due to a lack of domain-specific or contextual knowledge. In other cases, they hallucinate spurious matches (false positives), misled by superficial lex￾ical or syntactic cues [12]. Th… view at source ↗
Figure 2
Figure 2. Figure 2: Data Integration with RAG, GraphRAG, and KG-RAG. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Vision Towards Trustworthy and Cost￾Efficient Data Integration with Agentic RAG. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Toward Agentic RAG-based Data Integration. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Structural Sharing in Data Integration. To overcome these limitations, Agentic RAG [42] and GraphRAG [43] introduce a persistent graph-based mem￾ory mechanism that enables memory search directly from the graph rather than re-embedding or re-querying ex￾ternal knowledge bases. This structured memory sup￾ports complex multi-hop reasoning and provides stable knowledge grounding for LLMs. Recent studies highli… view at source ↗
Figure 6
Figure 6. Figure 6: System Architecture of Multi-Agent RAG for Trustwor [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 8 linked inside Pith

  1. [1]

    Amalur: Data integration meets machine learning,

    R. Hai, C. Koutras, A. Ionescu, Z. Li, W. Sun, J. van Schijndel, Y . Kang, and A. Katsifodimos, “Amalur: Data integration meets machine learning,” inICDE, 2023, pp. 3729–3739

    2023

  2. [2]

    Heterogeneous data integration: Challenges and opportunities,

    I. M. Putrama and P. Martinek, “Heterogeneous data integration: Challenges and opportunities,”Data in Brief, vol. 56, pp. 1–23, 2024

    2024

  3. [3]

    Human-in-the-loop rule learning for data integration,

    J. Fan and G. Li, “Human-in-the-loop rule learning for data integration,”IEEE Data Eng. Bull., vol. 41, no. 2, pp. 104–115, 2018

    2018

  4. [4]

    Amalur: Data integration meets machine learning,

    R. Hai, C. Koutras, A. Ionescu, Z. Li, W. Sun, J. Van Schijndel, Y . Kang, and A. Katsifodimos, “Amalur: Data integration meets machine learning,” inICDE, 2023, pp. 3729–3739

    2023

  5. [5]

    Unicorn: A unified multi-tasking model for supporting matching tasks in data integration,

    J. Tu, J. Fan, N. Tang, P. Wang, G. Li, X. Du, X. Jia, and S. Gao, “Unicorn: A unified multi-tasking model for supporting matching tasks in data integration,”Proc. ACM Manag. Data, vol. 1, no. 1, pp. 1–26, 2023

    2023

  6. [6]

    Large language models for data discovery and integration: Challenges and opportunities,

    J. Freire, G. Fan, B. Feuer, C. Koutras, Y . Liu, E. Pe ˜na, A. S. R. Santos, C. T. Silva, and E. Wu, “Large language models for data discovery and integration: Challenges and opportunities,”IEEE Data Eng. Bull., vol. 49, no. 1, pp. 3–31, 2025

    2025

  7. [7]

    Can foundation models wrangle your data?

    A. Narayan, I. Chami, L. J. Orr, and C. R ´e, “Can foundation models wrangle your data?”Proc. VLDB Endow., vol. 16, no. 4, pp. 738–746, 2022

    2022

  8. [8]

    LLM-Matcher: A name-based schema matching tool using large language models,

    M. Parciak, B. Vandevoort, F. Neven, L. M. Peeters, and S. Vansummeren, “LLM-Matcher: A name-based schema matching tool using large language models,” inSIGMOD Companion, 2025, pp. 203–206

    2025

  9. [9]

    Entity matching using large language models,

    R. Peeters, A. Steiner, and C. Bizer, “Entity matching using large language models,” inEDBT, 2025, pp. 529–541

    2025

  10. [10]

    In-context clustering-based entity resolution with large language models: A design space exploration,

    J. Fu, H. Tang, A. Khan, S. Mehrotra, X. Ke, and Y . Gao, “In-context clustering-based entity resolution with large language models: A design space exploration,”Proc. ACM Manag. Data, vol. 3, no. 4, pp. 1–28, 2025

    2025

  11. [11]

    Archetype: A novel framework for open-source column type annotation using large language models,

    B. Feuer, Y . Liu, C. Hegde, and J. Freire, “Archetype: A novel framework for open-source column type annotation using large language models,”Proc. VLDB Endow., vol. 17, no. 9, pp. 2279–2292, 2024. 11

    2024

  12. [12]

    Cost-efficient RAG for entity matching with LLMs: A blocking-based exploration,

    C. Ma, Z. Zhang, A. Khan, S. Schelter, and P. Groth, “Cost-efficient RAG for entity matching with LLMs: A blocking-based exploration,”CoRR, vol. abs/2602.05708, 2026

    arXiv 2026

  13. [13]

    Unveiling challenges for LLMs in enterprise data engineering,

    J. Bodensohn, U. Brackmann, L. V ogel, A. Sanghi, and C. Binnig, “Unveiling challenges for LLMs in enterprise data engineering,”Proc. VLDB Endow., vol. 19, no. 2, pp. 196–209, 2025

    2025

  14. [14]

    Bootstrapping self-improvement of language model programs for zero- shot schema matching,

    N. Seedat and M. van der Schaar, “Bootstrapping self-improvement of language model programs for zero- shot schema matching,” inICML, vol. 267, 2025, pp. 53 791–53 826

    2025

  15. [15]

    Mind the data gap: Bridging large language models (LLMs) to enterprise data integration,

    M. Kayali, F. Wenz, N. Tatbul, and C ¸ . Demiralp, “Mind the data gap: Bridging large language models (LLMs) to enterprise data integration,” inCIDR, 2025, pp. 1–6

    2025

  16. [16]

    GATHER: convergence-centric hyper-entity retrieval for zero-shot cell-type annotation,

    Z. Zhang, F. Jiang, S. Qin, J. Zhao, and M. Yang, “GATHER: convergence-centric hyper-entity retrieval for zero-shot cell-type annotation,”CoRR, vol. abs/2605.06403, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Structured multi-step reasoning for entity matching using large lan- guage model,

    R. Bopardikar, J. Wang, and J. Zou, “Structured multi-step reasoning for entity matching using large lan- guage model,”CoRR, vol. abs/2511.22832, 2025

    arXiv 2025

  18. [18]

    Assistem: Domain instruction tuning for en- hanced entity matching,

    J. B. Mugeni, S. J. Lynden, T. Amagasa, and A. Matono, “Assistem: Domain instruction tuning for en- hanced entity matching,” inPAKDD, vol. 15874, 2025, pp. 115–127

    2025

  19. [19]

    Fine-tuning large language models for entity matching,

    A. Steiner, R. Peeters, and C. Bizer, “Fine-tuning large language models for entity matching,” inICDEW. IEEE, 2025, pp. 9–17

    2025

  20. [20]

    Knowledge graph-based retrieval-augmented generation for schema matching,

    C. Ma, S. Chakrabarti, A. Khan, and B. Moln ´ar, “Knowledge graph-based retrieval-augmented generation for schema matching,”CoRR, vol. abs/2501.08686, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Knowledge-enriched schema mapping: A preliminary case study of e-MedSolution system,

    C. Ma, “Knowledge-enriched schema mapping: A preliminary case study of e-MedSolution system,” E¨otv¨os Lor´and University, Tech. Rep, 2020

    2020

  22. [22]

    ReMatch: Retrieval enhanced schema matching with LLMs,

    E. Sheetrit, M. Brief, M. Mishaeli, and O. Elisha, “ReMatch: Retrieval enhanced schema matching with LLMs,”CoRR, vol. abs/2403.01567, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    RACOON: An LLM-based framework for retrieval-augmented column type annotation with a knowledge graph,

    L. L. Wei, G. Xiao, and M. Balazinska, “RACOON: An LLM-based framework for retrieval-augmented column type annotation with a knowledge graph,” inTRL@NeurIPS, 2024

    2024

  24. [24]

    RACT: Retrieval augmented column-table learn- ing and prediction for multi-table schema matching,

    L. Traeger, E. Khwaileh, A. Behrend, and G. Karabatis, “RACT: Retrieval augmented column-table learn- ing and prediction for multi-table schema matching,”CoRR, vol. abs/2606.07843, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Construm: A structure-guided LLM framework for context-aware schema matching,

    H. Chen, Z. Zhang, and H. V . Jagadish, “Construm: A structure-guided LLM framework for context-aware schema matching,”CoRR, vol. abs/2601.20482, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    ArchRAG: Attributed community-based hierarchical retrieval-augmented generation,

    S. Wang, Y . Fang, Y . Zhou, X. Liu, and Y . Ma, “ArchRAG: Attributed community-based hierarchical retrieval-augmented generation,” inAAAI, vol. 40, no. 19, 2026, pp. 15 868–15 876

    2026

  27. [27]

    Schema matching on graph: Iterative graph exploration for efficient and explainable data integration,

    M. Jeon, J. Suh, and S. Cho, “Schema matching on graph: Iterative graph exploration for efficient and explainable data integration,”CoRR, vol. abs/2511.20285, 2025

    arXiv 2025

  28. [28]

    Towards the next generation of agent systems: From RAG to agentic AI,

    Y . Zhou and S. Wang, “Towards the next generation of agent systems: From RAG to agentic AI,” in LLM+Graph@VLDB, 2025, pp. 1–7

    2025

  29. [29]

    A collaborative multi-agent approach to retrieval-augmented generation across diverse data,

    A. Salve, S. Attar, M. Deshmukh, S. Shivpuje, and A. M. Utsab, “A collaborative multi-agent approach to retrieval-augmented generation across diverse data,”CoRR, vol. abs/2412.05838, 2024. 12

    Pith/arXiv arXiv 2024

  30. [30]

    A-RAG: scaling agentic retrieval- augmented generation via hierarchical retrieval interfaces,

    M. Du, B. Xu, C. Zhu, S. Wang, P. Wang, X. Wang, and Z. Mao, “A-RAG: scaling agentic retrieval- augmented generation via hierarchical retrieval interfaces,”CoRR, vol. abs/2602.03442, 2026

    arXiv 2026

  31. [31]

    SCOUT-RAG: scalable and cost-efficient unifying traversal for agentic Graph-RAG over distributed domains,

    L. Li, Y . Zou, J. Wu, Y . Wen, J. Li, H. Qian, and I. W. Tsang, “SCOUT-RAG: scalable and cost-efficient unifying traversal for agentic Graph-RAG over distributed domains,”CoRR, vol. abs/2602.08400, 2026

    arXiv 2026

  32. [32]

    Agentic GraphRAG: Navigating unstructured financial data with collaborative AI,

    A. Capozzi and D. Helbing, “Agentic GraphRAG: Navigating unstructured financial data with collaborative AI,”CoRR, vol. abs/2605.18770, 2026

    Pith/arXiv arXiv 2026

  33. [33]

    Agentic RAG with knowledge graphs for complex multi-hop reasoning in real-world applications,

    J. Lelong, A. Errazine, and A. Blangero, “Agentic RAG with knowledge graphs for complex multi-hop reasoning in real-world applications,” inECAI, vol. 413, 2025, pp. 5163–5166

    2025

  34. [34]

    Multi-agent RAG framework for entity resolution: Advancing beyond single-LLM approaches with specialized agent coor- dination,

    A. M. Althaf, M. A. Mohammed, M. G. Milanova, J. R. Talburt, and M. C. Cakmak, “Multi-agent RAG framework for entity resolution: Advancing beyond single-LLM approaches with specialized agent coor- dination,”Comput., vol. 14, no. 12, p. 525, 2025

    2025

  35. [35]

    MCTS-RAG: enhancing retrieval-augmented generation with Monte Carlo tree search,

    Y . Hu, Y . Zhao, C. Zhao, and A. Cohan, “MCTS-RAG: enhancing retrieval-augmented generation with Monte Carlo tree search,” inEMNLP (Findings), 2025, pp. 12 581–12 597

    2025

  36. [36]

    MA-RAG: multi-agent retrieval-augmented generation via collaborative chain-of-thought reasoning,

    T. Nguyen, P. Chin, and Y . Tai, “MA-RAG: multi-agent retrieval-augmented generation via collaborative chain-of-thought reasoning,”CoRR, vol. abs/2505.20096, 2025

    arXiv 2025

  37. [37]

    Multi-agent collaboration via evolving orchestration,

    Y . Dang, C. Qian, X. Luo, J. Fan, Z. Xie, R. Shi, W. Chen, C. Yang, X. Che, Y . Tian, X. Xiong, L. Han, Z. Liu, and M. Sun, “Multi-agent collaboration via evolving orchestration,” inNeurIPS, 2025

    2025

  38. [38]

    Optimized batch prompting for cost-effective LLMs,

    Z. Ji, X. Wang, Z. Luo, Z. Xie, and M. Zhang, “Optimized batch prompting for cost-effective LLMs,” Proc. VLDB Endow., vol. 18, no. 7, p. 2172–2184, 2025

    2025

  39. [39]

    Cost-effective in-context learning for entity resolution: A design space exploration,

    M. Fan, X. Han, J. Fan, C. Chai, N. Tang, G. Li, and X. Du, “Cost-effective in-context learning for entity resolution: A design space exploration,” inICDE, 2024, pp. 3696–3709

    2024

  40. [40]

    RAGCache: Efficient knowledge caching for retrieval-augmented generation,

    C. Jin, Z. Zhang, X. Jiang, F. Liu, S. Liu, X. Liu, and X. Jin, “RAGCache: Efficient knowledge caching for retrieval-augmented generation,”ACM Trans. Comput. Syst., vol. 44, no. 1, pp. 1–27, Nov. 2025

    2025

  41. [41]

    Cache-Craft: Managing chunk-caches for efficient retrieval-augmented generation,

    S. Agarwal, S. Sundaresan, S. Mitra, D. Mahapatra, A. Gupta, R. Sharma, N. J. Kapu, T. Yu, and S. K. Saini, “Cache-Craft: Managing chunk-caches for efficient retrieval-augmented generation,”Proc. ACM Manag. Data, vol. 3, no. 3, pp. 1–28, 2025

    2025

  42. [42]

    Cache mechanism for agent RAG systems,

    S. Lin, Z. Peng, L. Li, X. Lin, X. Zhu, and Y . Zhang, “Cache mechanism for agent RAG systems,”CoRR, vol. abs/2511.02919, 2025

    arXiv 2025

  43. [43]

    MemGraphRAG: Memory-based multi-agent system for graph retrieval-augmented generation,

    C. Wu, Z. Xiang, Y . Tang, Z. Chen, Q. Zhang, and J. Su, “MemGraphRAG: Memory-based multi-agent system for graph retrieval-augmented generation,” inKDD, 2026, pp. 1–20

    2026

  44. [44]

    MAGMA: A multi-graph based agentic memory architecture for AI agents,

    D. Jiang, Y . Li, G. Li, and B. Li, “MAGMA: A multi-graph based agentic memory architecture for AI agents,”CoRR, vol. abs/2601.03236, 2026

    Pith/arXiv arXiv 2026

  45. [45]

    Context-aware intelligent data integration approach: A practical hybrid retrieval agent,

    R. M. Hamza and G. E. Ajam, “Context-aware intelligent data integration approach: A practical hybrid retrieval agent,” inICCES, 2026, p. 121–127

    2026

  46. [46]

    Agent-OM: Leveraging LLM agents for ontology matching,

    Z. Qiang, W. Wang, and K. Taylor, “Agent-OM: Leveraging LLM agents for ontology matching,”Proc. VLDB Endow., vol. 18, no. 3, pp. 516–529, 2024

    2024

  47. [47]

    Tool-planner: Task planning with clusters across multiple tools,

    Y . Liu, X. Peng, J. Cao, S. Bo, Y . Zhang, X. Zhang, S. Cheng, X. Wang, J. Yin, and T. Du, “Tool-planner: Task planning with clusters across multiple tools,” inICLR, 2025. 13

    2025