DSRAG: A Domain-Specific Retrieval Framework Based on Document-derived Multimodal Knowledge Graph

Chunxiao Xing; David Osei Opoku; Mengzheng Yang; Peng Ren; Ruochang Li; Yanfei Ren

REVIEW 3 major objections 5 minor 1 cited by

DSRAG: A Domain-Specific Retrieval Framework Based on Document-derived Multimodal Knowledge Graph

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A two-layer document-derived knowledge graph lifts domain-specific question answering past plain RAG on every measured metric.

desk verdict A reasonable two-layer multimodal KG retrieval pipeline for domain QA, but the comparative claim rests on a single 100-question evaluation and an undeclared reranker. read the letter →

arxiv 2509.10467 v1 pith:EHWU5FJN submitted 2025-08-22 cs.IR cs.AIcs.CLcs.CVcs.MM

Mengzheng Yang , Yanfei Ren , David Osei Opoku , Ruochang Li , Peng Ren , Chunxiao Xing This is my paper · ORCID

classification cs.IRcs.AIcs.CLcs.CVcs.MM

keywords retrieval-augmentedgenerationmultimodalknowledgegraphdomain-specificquestionansweringconceptinstancesemanticpruninggraph-enhancedretrievallargelanguagemodels

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

DSRAG is a retrieval-augmented generation pipeline built around a document-derived multimodal knowledge graph with two levels: a Concept KG drawn from the document's table of contents and an Instance KG built by extracting entities, relations, tables, and image descriptions. The paper's claim is that this two-level structure lets retrieval first prune the search space semantically along the concept hierarchy and then match vector embeddings within the surviving chapters, and that this combination yields materially better domain-specific question answering than ordinary RAG, a graph-enhanced pipeline, or a multimodal RAG engine. On one 4,500-page database manual with 100 expert-authored questions, DSRAG reports the highest scores on faithfulness, answer relevancy, and contextual precision (0.83, 0.87, 0.85 versus the best baseline at 0.72, 0.67, 0.70). The practical bet is that technical manuals encode knowledge in their structure as well as their prose, and DSRAG tries to put that implicit structure to work.

What carries the argument

DSKG, the domain-specific multimodal knowledge graph, is the load-bearing object. It has a Concept KG—a directed acyclic graph of parts, chapters, and sections derived from the table of contents, with chapter summaries and keyword edges (subTopic, hasKeyword)—and an Instance KG, built by a three-level extraction pipeline that turns text chunks, tables, and image descriptions into entity-relation triples tied to the concept subgraphs. The companion retrieval mechanism, DSKG-Enhanced Retrieval, uses layer-wise pruning to reduce the candidate space before vector search, then feeds graph context plus vector hits into the generator.

What would settle it

Re-score the same 100 questions with human raters or a second judge model and compute per-metric confidence intervals: if DSRAG's faithfulness, answer relevancy, and contextual precision no longer exceed the best baseline outside the error bars, the central claim is falsified. A second decisive check is applying DSRAG to another long technical manual with a standard QA benchmark and seeing whether it still beats the same three baselines on the same metrics.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that organising a domain-specific technical document into a two-layer multimodal knowledge graph—concept nodes from the table of contents and instance triples from text, tables, and images—and using that graph to guide retrieval produces answers that are more faithful to the source, more relevant to the query, and better supported by the retrieved context than vector-only RAG or existing graph-plus-vector RAG systems. The mechanism is a two-stage retrieval: graph-guided focusing first decomposes the query, prunes irrelevant chapters and sections along the Concept KG, and extracts instance-level facts from the focused subgraph; vector retrieval then

Load-bearing premise

The central superiority claim rests on a single 4,500-page database manual, 100 expert-written questions, and LLM-as-judge scores without confidence intervals or significance tests; if that evaluation set is unrepresentative or the judge favors one answer style, the claimed advantage may not hold elsewhere.

Editorial extensions

If this is right

Domain-specific question answering on long technical documentation can be improved by mining the table of contents into a concept graph rather than treating the document as a flat corpus.
A hierarchical knowledge graph with both concept and instance layers makes retrieval more precise because the concept layer constrains where the vector matcher is allowed to look.
Tables and image descriptions can be treated as first-class knowledge chunks during graph construction instead of being flattened into plain text.
The full two-layer system outperforms either layer alone on all three metrics in the ablation, showing that concept-level and instance-level structure carry complementary information.
Strongest answers come from combining structured graph context with unstructured vector context, not from either retrieval signal alone.

Reading between the lines

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

The same two-layer recipe may transfer to any structured technical corpus whose table of contents encodes domain semantics, such as medical handbooks or equipment manuals; the paper only demonstrates it on a database manual.
Graph-guided pruning could be reused as a routing step in other retrieval systems, since it depends only on a document hierarchy and entity extraction rather than on a specific generative model.
A testable extension is to measure how retrieval efficiency (number of chunks remaining after pruning) scales with document length; DSRAG's design suggests sub-linear candidate growth, but the paper does not report latency or coverage statistics.

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.

Desk Editor's Note

read the letter

Quick take: this is a coherent system paper. The DSRAG pipeline — TOC-driven concept KG plus an instance KG built from text, tables, and images, with layer-wise pruning before vector retrieval — is a sensible integration of existing pieces. The method section is clear, and the ablation shows the expected monotonic improvement from adding IKG and then the full DSKG. That internal consistency is worth something.

The problem is the evidence for the headline claim. "Significantly outperforms all baselines" (Section 4.2) is based on one 4,500-page database manual, 100 self-designed questions, and three point estimates from an LLM judge. There are no error bars, no significance tests, no multi-domain validation, and no released code or data to check the numbers. That is thin for such a strong claim.

On top of that, the stress-test concern holds up: Table 1 lists jina-reranker-v2-base as the reranking model, but Section 3.3's retrieval description never mentions a reranking step. If DSRAG used the reranker and the baselines didn't, the +0.11 to +0.20 gains could come from that post-processing rather than from the KG. The paper needs an ablation that isolates the reranker.

Two other omissions: there is no direct comparison with DO-RAG (reference 10, by two of the same authors), which is the most relevant graph-augmented RAG baseline; and the retrieval hyperparameters (top-k, pruning thresholds) are not released, so reproducibility is limited.

None of this suggests the approach is wrong. The concept-plus-instance KG design is a legitimate engineering contribution, and the paper is honest about using expert annotations. But as submitted, the evaluation does not support the claim. I would send this to peer review — a good referee can push for a proper multi-domain evaluation, code release, and a reranker ablation — and I would not expect it to be accepted without those changes. For a reading group, it is a useful case study in how to (and how not to) evaluate a RAG system.

Referee Report

3 major / 5 minor

Summary. The paper proposes DSRAG, a domain-specific retrieval-augmented generation framework built on a document-derived multimodal knowledge graph (DSKG). The framework has two graph layers: a Concept KG constructed from document structure and summaries, and an Instance KG built from fine-grained multimodal entities and relations. Retrieval combines graph-guided semantic focusing (sub-query decomposition, layer-wise pruning, subgraph retrieval) with vector retrieval over chapter-filtered chunks. The authors evaluate on a 4,500-page database manual with 100 expert-designed questions, comparing against Naive RAG, TiDB AutoFlow, and RAGFlow using LLM-as-judge metrics (Faithfulness, Answer Relevancy, Contextual Precision), and report that DSRAG 'significantly outperforms' all baselines on all metrics (Table 2). An ablation study (Table 3) examines RAG, IKG-RAG, and full DSKG-RAG.

Significance. If the empirical claim is sound, DSRAG would be a useful engineering contribution: it demonstrates a concrete way to combine multimodal document parsing, two-level knowledge graph construction, and graph-guided retrieval with vector search, and it compares against realistic industrial baselines. The hierarchical Concept/Instance KG distinction and the pruning/subgraph retrieval mechanism are sensible and reasonably detailed. The design does not appear circular: the evaluation metrics are not derived from the framework, and the method has no fitted parameters that are tuned on the test set. However, the current evidence base is too thin to support the advertised claim: one document, 100 questions, no variance or statistical testing, and an unaccounted reranking component listed in Table 1 but absent from the method description. These issues make the central quantitative claim provisional rather than established.

major comments (3)

[Table 1 vs. §3.3] The reranking model jina-reranker-v2-base is listed in the experimental configuration (Table 1) but is never mentioned in the DSRAG retrieval pipeline in §3.3, which describes only graph-guided focusing and vector retrieval. It is therefore unclear whether DSRAG applies a reranking post-processing step and, critically, whether the baselines (Naive RAG, TiDB AutoFlow, RAGFlow) received the same reranking treatment. If only DSRAG used reranking, the reported gains in Table 2 (e.g., +0.11 Faithfulness, +0.20 Answer Relevancy, +0.15 Contextual Precision over the best baseline) could be attributable to reranking rather than to the multimodal knowledge graph. Please clarify whether reranking is part of DSRAG's pipeline, whether it was applied uniformly to all baselines, and ideally add an ablation that isolates the reranker or remove it from the configuration.
[§4.1–4.2, Tables 2–3] The claim that 'DSRAG significantly outperforms all baseline methods' is not supported by the reported evidence. The evaluation uses a single 100-question dataset, a single document, and single point estimates from one run. No standard deviations, confidence intervals, repeated runs, or statistical significance tests are reported, so the word 'significantly' is not justified. The LLM-as-judge setup (GPT-4o as evaluator) is also not validated against human judgments. At minimum, the authors should provide variance estimates across multiple runs and a paired significance test (e.g., bootstrap or Wilcoxon signed-rank on the 100 questions), and should temper the wording to match what the evidence can support. As is, the central comparative claim is load-bearing and currently unsupported.
[Tables 2 and 3] There is an unexplained inconsistency between the main comparison and the ablation study. The 'RAG' row in Table 3 reports Faithfulness 0.68, Answer Relevancy 0.70, Contextual Precision 0.58, but the 'Naive RAG' row in Table 2 reports 0.63, 0.58, 0.55 for the same three metrics. If the RAG baseline is the same in both tables, the numbers should match; if the ablation uses a different RAG configuration (e.g., different chunking or retrieval parameters), that needs to be stated. This discrepancy undermines the interpretability of the ablation and makes it hard to attribute improvements to the KG components.

minor comments (5)

[§4.1] The baseline is called 'NativeRAG' in the text but 'Naive RAG' in Table 2. Please unify the name to avoid confusion.
[Table 1] There is a formatting issue: 'T able 1' has an extra space. Also, if the reranking model is retained, its role in the pipeline should be described in §3.3.
[§4.1] The Langfuse multidimensional scoring mechanism is mentioned as the evaluation tool, but no details are given about how the three metrics are computed, what prompts are used for the LLM judge, or whether judge outputs were manually spot-checked. Please add a short description or reference.
[§4.1] The dataset description would benefit from sample questions and answers, especially because the 100 questions are self-designed. Including a few examples in an appendix would help readers assess the difficulty and domain coverage.
[General] Definition of Contextual Precision is vague ('assesses whether factually relevant entries are ranked higher within the retrieved context'). Please give the precise formula or reference, since this metric is central to the retrieval claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DSRAG's claims are empirical comparisons, not derivations from their own inputs.

full rationale

The paper's central claim is an experimental performance comparison (Section 4.2, Table 2) between DSRAG and three external baselines on a held-out set of 100 expert-designed questions. No parameter is fitted to the evaluation metrics and then renamed as a prediction; the DSKG construction and graph-guided retrieval are described procedurally (Sections 3.2–3.3) and do not define the target metrics in terms of the model's own inputs. The only overlapping-author citation is DO-RAG [10], mentioned in Related Work as a simultaneous domain-specific QA approach; it is not used to justify DSRAG's architecture or to import a uniqueness theorem, so it is not load-bearing. The listed reranking model (jina-reranker-v2-base, Table 1) is never described in the retrieval pipeline of §3.3, and no error bars or significance tests accompany Table 2; these are missing-support/confounding concerns for the comparative claim, not evidence that the result reduces to its inputs by construction. Accordingly, no circular step is exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

DSRAG introduces no mathematically fitted constants, but it relies on several domain assumptions: the document hierarchy is semantically meaningful, expert annotations and the LLM-based extractor produce trustworthy triples, the LLM judge is a valid evaluator, and one manual represents the target domain. The framework's own graph structures are inventions of the paper, not independently verifiable entities.

free parameters (1)

Retrieval hyperparameters (number of sub-queries, vector top-k, pruning similarity thresholds) = not reported
Chosen by hand without sensitivity analysis; directly affect all reported metric values (§3.3).

assumptions (4)

domain assumption Document table of contents reflects domain knowledge organization.
Concept KG construction treats TOC titles as concept nodes (§3.2).
domain assumption Expert annotations and GPT-4o-mini extractions are accurate enough for KG construction.
No manual quality check or inter-annotator agreement is reported; used for Concept KG and Instance KG (§3.2).
domain assumption GPT-4o's Langfuse-based scoring correctly measures faithfulness, answer relevancy, and contextual precision.
All reported numbers come from LLM-as-judge; no human correlation or calibration is reported (§4.1).
domain assumption The single 4,500-page database manual is representative of domain-specific documents.
Evaluation is performed on one corpus; no cross-domain validation is provided (§4.1).

invented entities (1)

DSKG (hierarchical multimodal knowledge graph: Concept KG + Instance KG)
purpose: Unified representation of document concepts, entities, relations, tables, and images for retrieval
Internal framework artifact; no external falsifiable prediction, only downstream QA scores that depend on many other components.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DSRAG: A Domain-Specific Retrieval Framework Based on Document-derived Multimodal Knowledge Graph." pith.science (2026). https://pith.science/paper/EHWU5FJN

@misc{pith2026250910467,
  author       = {Pith},
  title        = {Pith review of: DSRAG: A Domain-Specific Retrieval Framework Based on Document-derived Multimodal Knowledge Graph},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHWU5FJN}},
  note         = {Machine review of arXiv:2509.10467}
}

read the original abstract

Current general-purpose large language models (LLMs) commonly exhibit knowledge hallucination and insufficient domain-specific adaptability in domain-specific tasks, limiting their effectiveness in specialized question answering scenarios. Retrieval-augmented generation (RAG) effectively tackles these challenges by integrating external knowledge to enhance accuracy and relevance. However, traditional RAG still faces limitations in domain knowledge accuracy and context modeling.To enhance domain-specific question answering performance, this work focuses on a graph-based RAG framework, emphasizing the critical role of knowledge graph quality during the generation process. We propose DSRAG (Domain-Specific RAG), a multimodal knowledge graph-driven retrieval-augmented generation framework designed for domain-specific applications. Our approach leverages domain-specific documents as the primary knowledge source, integrating heterogeneous information such as text, images, and tables to construct a multimodal knowledge graph covering both conceptual and instance layers. Building on this foundation, we introduce semantic pruning and structured subgraph retrieval mechanisms, combining knowledge graph context and vector retrieval results to guide the language model towards producing more reliable responses. Evaluations using the Langfuse multidimensional scoring mechanism show that our method excels in domain-specific question answering, validating the efficacy of integrating multimodal knowledge graphs with retrieval-augmented generation.

Figures

Figures reproduced from arXiv: 2509.10467 by the authors.

**Figure 1.** Diagram of the example of factual hallucination To address the knowledge gap in domain-specific domains, the academic community primarily explores two technical approaches: unsupervised incremental pre-training [15] and retrieval-augmented generation (RAG) [9] mechanisms. The former enhances knowledge by injecting domain-specific corpora, but the high training costs and delayed knowledge update efficiency limit its… view at source ↗

**Figure 2.** The overall framework of DSRAG The DSKG-Enhanced Retrieval process further improves accuracy through a two-stage retrieval strategy: the first stage retrieves structured semantic information via graph traversal, while the second stage retrieves contextually relevant information through vector similarity search. Finally, the graph-based context, vector retrieval context, and historical dialogue context are integrate… view at source ↗

**Figure 3.** Flow Diagram of DSKG Construction from Original Documents to MMKG Concept KG Construction. In domain-specific technical documents, the table of contents serves as both an organizational tool and a reflection of implicit knowledge and engineering semantics. This study proposes a method for constructing a concept knowledge graph driven by the document’s structural hierarchy, leveraging existing semantic information… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visualization of the DSKG 3.3 DSKG-Enhanced Retrieval As shown in Fig.5, the proposed DSKG-enhanced retrieval combines structured graph semantic modeling with unstructured corpus vector search, forming a multi-source retrieval-augmented generation framework that builds…

**Figure 5.** Figure 5: DSKG-enhanced Retrieval Process the contextual semantics. The process first performs semantic decomposition of the original query and generates sub-queries, each corresponding to a different semantic sub-dimension. Based on this semantic modeling, a layer-wise pruning …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Panning for Gold: Expanding Domain-Specific Knowledge Graphs with General Knowledge
cs.AI 2026-01 reject novelty 6.0 of 10

ExeFuse uses learned 'logic' transformations and density checks to fuse general-graph facts into domain knowledge graphs, but the benchmark labels and baseline comparisons are too underspecified to support the claimed gains.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages · cited by 1 Pith paper

[1]

Journal of Network and Computer Applications185, 103076 (2021)

Abu-Salih, B.: Domain-specific knowledge graphs: A survey. Journal of Network and Computer Applications185, 103076 (2021)

work page 2021
[2]

In: Proceedings of the IEEE/ACM 12 M

Barnett,S.,Kurniawan,S.,Thudumu,S.,etal.:Sevenfailurepointswhenengineer- ing a retrieval augmented generation system. In: Proceedings of the IEEE/ACM 12 M. Yang et al. 3rd International Conference on AI Engineering-Software Engineering for AI. pp. 194–199 (2024)

work page 2024
[3]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Ding, Y., Yu, J., Liu, B., et al.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5089– 5098 (2022)

work page 2022
[4]

ACM Transactions on Computing for Healthcare (HEALTH)3(1), 1–23 (2021)

Gu, Y., Tinn, R., Cheng, H., et al.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH)3(1), 1–23 (2021)

work page 2021
[5]

IEEE Transactions on Knowledge and Data Engineering 36(4), 1413–1430 (2023)

Hu, L., Liu, Z., Zhao, Z., et al.: A survey of knowledge enhanced pre-trained lan- guage models. IEEE Transactions on Knowledge and Data Engineering 36(4), 1413–1430 (2023)

work page 2023
[6]

ACM Transactions on Information Systems43(2), 1–55 (2025)

Huang, L., Yu, W., Ma, W., et al.: A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems43(2), 1–55 (2025)

work page 2025
[7]

https://www.ragflow.io/ (2024), accessed: 2025-04-28

InfiniFlow: Ragflow. https://www.ragflow.io/ (2024), accessed: 2025-04-28

work page 2024
[8]

ACM computing surveys55(12), 1–38 (2023)

Ji, Z., Lee, N., Frieske, R., et al.: Survey of hallucination in natural language generation. ACM computing surveys55(12), 1–38 (2023)

work page 2023

Show all 18 references

[9]

Advances in neural information processing systems 33, 9459–9474 (2020)

Lewis, P., Perez, E., Piktus, A., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020)

2020
[10]

arXiv preprint arXiv:2505.17058 (2025)

Opoku, D.O., Sheng, M., Zhang, Y.: Do-rag: A domain-specific qa framework using knowledge graph-enhanced retrieval-augmented generation. arXiv preprint arXiv:2505.17058 (2025)

2025 arXiv
[11]

arXiv preprint arXiv:2408.08921 (2024)

Peng, B., Zhu, Y., Liu, Y., et al.: Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921 (2024)

2024 arXiv
[12]

https://github.com/pingcap/autoflow (2023), ac- cessed: 2025-04-28

PingCAP: Tidb autoflow. https://github.com/pingcap/autoflow (2023), ac- cessed: 2025-04-28

2023
[13]

In: Proceedings of the 5th ACM International Conference on AI in Finance

Sarmah, B., Mehta, D., Hall, B., et al.: Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In: Proceedings of the 5th ACM International Conference on AI in Finance. pp. 608– 616 (2024)

2024
[14]

Information Processing & Management62(4), 104101 (2025)

Sheng,M.,Wang,S.,Zhang,Y.,etal.:Mqrld:Amultimodaldataretrievalplatform with query-aware feature representation and learned index based on data lake. Information Processing & Management62(4), 104101 (2025)

2025
[15]

Nature 620(7972), 172–180 (2023)

Singhal, K., Azizi, S., Tu, T., et al.: Large language models encode clinical knowl- edge. Nature 620(7972), 172–180 (2023)

2023
[16]

In: Proceedings of the 16th international conference on World Wide Web

Suchanek, F.M., Kasneci, G., Weikum, G.: Yago: a core of semantic knowledge. In: Proceedings of the 16th international conference on World Wide Web. pp. 697–706 (2007)

2007
[17]

In: Proceedings of the 26th ACM SIGKDD interna- tional conference on knowledge discovery & data mining

Xu, Y., Li, M., Cui, L., et al.: Layoutlm: Pre-training of text and layout for doc- ument image understanding. In: Proceedings of the 26th ACM SIGKDD interna- tional conference on knowledge discovery & data mining. pp. 1192–1200 (2020)

2020
[18]

IEEE Transactions on Knowledge and Data Engineering 36(2), 715–735 (2022)

Zhu, X., Li, Z., Wang, X., et al.: Multi-modal knowledge graph construction and application: A survey. IEEE Transactions on Knowledge and Data Engineering 36(2), 715–735 (2022)

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Journal of Network and Computer Applications185, 103076 (2021)

Abu-Salih, B.: Domain-specific knowledge graphs: A survey. Journal of Network and Computer Applications185, 103076 (2021)

work page 2021

[2] [2]

In: Proceedings of the IEEE/ACM 12 M

Barnett,S.,Kurniawan,S.,Thudumu,S.,etal.:Sevenfailurepointswhenengineer- ing a retrieval augmented generation system. In: Proceedings of the IEEE/ACM 12 M. Yang et al. 3rd International Conference on AI Engineering-Software Engineering for AI. pp. 194–199 (2024)

work page 2024

[3] [3]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Ding, Y., Yu, J., Liu, B., et al.: Mukea: Multimodal knowledge extraction and accumulation for knowledge-based visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5089– 5098 (2022)

work page 2022

[4] [4]

ACM Transactions on Computing for Healthcare (HEALTH)3(1), 1–23 (2021)

Gu, Y., Tinn, R., Cheng, H., et al.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH)3(1), 1–23 (2021)

work page 2021

[5] [5]

IEEE Transactions on Knowledge and Data Engineering 36(4), 1413–1430 (2023)

Hu, L., Liu, Z., Zhao, Z., et al.: A survey of knowledge enhanced pre-trained lan- guage models. IEEE Transactions on Knowledge and Data Engineering 36(4), 1413–1430 (2023)

work page 2023

[6] [6]

ACM Transactions on Information Systems43(2), 1–55 (2025)

Huang, L., Yu, W., Ma, W., et al.: A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems43(2), 1–55 (2025)

work page 2025

[7] [7]

https://www.ragflow.io/ (2024), accessed: 2025-04-28

InfiniFlow: Ragflow. https://www.ragflow.io/ (2024), accessed: 2025-04-28

work page 2024

[8] [8]

ACM computing surveys55(12), 1–38 (2023)

Ji, Z., Lee, N., Frieske, R., et al.: Survey of hallucination in natural language generation. ACM computing surveys55(12), 1–38 (2023)

work page 2023