Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Multimodal CoLRAG-TF: Triple-Filtered Retrieval for Complex PDFs

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A four-axis retrieval system reaches 99% block recall when knowledge triples are the dominant signal.

desk verdict The engineering is real but the headline recall is internally inconsistent with the paper's own tables, and the self-referential evaluation can't support the causal claims. read the letter →

arxiv 2607.20517 v1 pith:LKU7MH5W submitted 2026-07-07 cs.LG

Takato Yasuno This is my paper
classification cs.LG
keywords retrieval-augmentedgenerationmultimodalretrievalknowledge-graphtriplesopeninformationextractionBM25fusionmulti-hopreasoningPDFdocumentunderstandingBayesianweightoptimization
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

The paper seeks to show that retrieval over heterogeneous, noisy PDFs — text, tables, scanned images, dense jargon — becomes near-perfect when the retriever fuses four signals and lets a structured knowledge-triple axis dominate the score. On a 457-pair benchmark built from 43 Japanese disaster lesson documents, the system reports top-5 block recall of 0.9909, perfect single-hop recall, and a 71.6% higher answer-similarity for multi-hop than single-hop questions. The paper argues that Bayesian weight tuning is what reveals the triple axis must be weighted at 0.44; otherwise keyword matching swamps the graph signal in terminologically dense domains. If this is right, the practical message is that multi-hop reasoning in domain RAG depends less on embedding quality than on structured relation extraction and fusion calibration.

What carries the argument

The central object is the four-axis fusion score s(b)=α_t·s_t + α_b·s_b + α_tr·s_tr + α_i·s_i, where each axis is normalized to [0,1] and weights are found by Bayesian optimization. The triple axis is served by an index of 11,414 subject–relation–object triples extracted from blocks, with sub-second lookup and relevance propagated upward through a three-level volume→chapter→block hierarchy. The work the machinery does: it lets structured facts both override and complement lexical and dense signals, and it turns fusion-weight calibration into the main correctness lever.

What would settle it

Re-run the 457-pair evaluation with the triple axis removed (α_triple=0) while keeping all other components identical; if recall stays near 0.99 and the multi-hop advantage stays near 71%, the paper's central causal claim is false. Alternatively, check whether any of the 40 tuning blocks appear in the benchmark's gold lists — overlap would mean the headline is partly a fitting artifact.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that a four-axis fusion score — dense embedding similarity, BM25 keyword match, triple similarity from open-information extraction, and image/caption similarity — with the triple axis weighted at 0.44 after Bayesian optimization, achieves 0.9909 retrieval recall on a 457-pair QA benchmark and lifts multi-hop answer similarity to 0.5605 versus 0.3267 for single-hop, a 71.6% gap. The paper reports that the flat dense-only baseline reaches only 0.72 recall, that triple scores peak at 0.9265 for numeric fact queries, and that weight calibration changes axis dominance: triple-driven top results rise from 5% to 40%. For image queries, a vision-language descriptio

Load-bearing premise

The headline recall and multi-hop gains rest on the 40 gold-standard blocks used to tune fusion weights being disjoint from the 457-pair benchmark; without that separation, the reported numbers are in-sample fit statistics.

Editorial extensions

If this is right

  • If the headline numbers hold, top-5 evidence coverage is effectively solved for this corpus: 0.9909 overall recall and 1.0000 single-hop recall on the 457-pair set.
  • A 71.6% multi-hop advantage means the triple axis is supplying cross-document, compositional evidence rather than merely re-ranking by keywords.
  • The flat dense baseline's 0.72 recall implies hierarchical narrowing and triple filtering — not embedding choice — account for most of the retrieval gain.
  • Numeric-fact queries benefit most from triples: peak triple score 0.9265 versus 0.58 text score, so exact quantities are carried by structured relations.
  • Weight calibration is not a nicety: shifting triple dominance from 5% to 40% of top results changes retrieval behavior substantially in this domain.

Reading between the lines

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

  • A direct ablation — setting the triple weight to zero and rerunning the 457-pair benchmark — is the decisive test of the causal claim; the paper reports an unablated system only.
  • If the triple axis is truly what drives multi-hop gains, the same fusion recipe should transfer to other jargon-dense corpora (clinical guidelines, regulatory handbooks), but the Japanese-specific tokenizer and embedding would need to be rebuilt.
  • The benchmark's gold answers and QA pairs were generated and filtered by LLMs without human validation; until a human-annotated subset is checked, the 0.9909 figure should be read as an upper bound on measured performance, not a settled ceiling.
  • For image queries, improving the vision model's disaster-type recognition from 41.7% would likely raise top-1 scores faster than any retrieval-side change, since misclassification propagates directly into the query text.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents Multimodal CoLRAG-TF, a four-axis retrieval system for heterogeneous, multimodal Japanese disaster PDFs. It combines dense text embeddings, BM25, OpenIE knowledge-graph triples, and image similarity in a late-interaction fusion score, with a HippoRAG2-inspired volume→chapter→block hierarchy and Optuna-based Bayesian weight optimization. The empirical evaluation reports an overall Retrieval Recall of 0.9909 on a 457-pair QA benchmark, a 71.6% multi-hop improvement in answer similarity over single-hop queries, and a dominant triple-axis weight (α_triple=0.44) after optimization. The paper also describes a hybrid OCR pipeline, an image-to-lesson interface, and a companion GitHub repository.

Significance. If the reported results were valid, the paper would make a useful engineering contribution: it demonstrates a concrete way to integrate structured triples into a multi-axis RAG retriever for a domain with noisy OCR and dense administrative vocabulary, and it ships reproducible code and configuration details. The hybrid OCR improvement and the Bayesian weight analysis are interesting in isolation. However, the central quantitative claims—near-perfect retrieval recall and a triple-caused multi-hop gain—are undermined by internal arithmetic inconsistencies and by a self-referential evaluation setup. The strengths are architectural and methodological; the empirical evidence as presented does not support the headline conclusions.

major comments (4)
  1. [§5.3, Tables 1 and 2; also Table S5] The reported Overall Retrieval Recall of 0.9909 (Table 1) is arithmetically incompatible with the per-type recalls in Table 2. The benchmark has 169 single-hop and 288 multi-hop pairs; Table 2 reports Recall 1.0000 for single-hop and 0.5000 for multi-hop. Weighted aggregation gives (169×1.0000 + 288×0.5000)/457 ≈ 0.685, not 0.9909. The same discrepancy appears in Table S5 for Naive RAG: 0.72 overall vs. (169×0.89 + 288×0.41)/457 ≈ 0.59. Since §5.1 defines Retrieval Recall as the fraction of gold block IDs in top-5, no variation of the definition reconciles these numbers. The headline near-perfect recall is therefore not supported by the paper's own data.
  2. [§3.6 and §5.7] The fusion weights that define the final system are optimized against 40 gold-standard blocks via Optuna (Eq. 6) and then the same system is evaluated on the 457-pair benchmark in §5.3 with no explicit statement that the 40 blocks were held out. If the 40 blocks belong to the 457-pair benchmark, the reported recall is an in-sample fit statistic. Additionally, the QA pairs and gold answers in §4.2 are generated by Qwen2.5-7B and filtered by an LLM judge, with no human validation. This creates a self-referential evaluation loop: the same family of models generates the queries, the gold answers, and the quality judgments used to tune and measure the system. The paper needs an out-of-sample, human-validated benchmark or a clear statement that the tuning set is disjoint.
  3. [§5.4, Table 2, and Table S5] The claim that the +71.6% multi-hop advantage in answer similarity 'validates triple-augmented compositional reasoning' is not supported. The dense-only Naive RAG baseline already shows a +53.5% multi-hop advantage (Table S5), so much of the gain may reflect multi-hop answer length or coverage rather than triple filtering. Moreover, the comparison v0.7.2 vs. v0.7.3 changes more than one variable: v0.7.2 has no BM25 and different weights, so it cannot isolate the triple axis. A proper ablation would remove the triple axis entirely while holding other components fixed; no such experiment is reported.
  4. [§6.2 and §7] The paper explicitly acknowledges that multi-hop gold annotations are incomplete, calling it a 'multi-hop gold annotation incompleteness' limitation. Yet the same annotations are used to compute the headline single-hop and multi-hop recall values. This undercuts the internal-consistency defense that the 0.5 multi-hop ceiling is an artifact of annotation: if annotations are incomplete, the reported 1.0000 single-hop recall and 0.9909 overall recall cannot be interpreted as retrieval accuracy. The paper should either supply complete annotations or use a metric robust to missing labels; currently the claims overstate what the data can show.
minor comments (5)
  1. [§3.3] Typo: 'BM25Okapi over the same 2,430 blocks' should be 2,403 (the corpus count given elsewhere).
  2. [Figure 5 and Table 4] The mean Top-1 score is reported as 0.6556 in the text and Table 4, but Figure 5's caption lists 0.6631. These should be consistent.
  3. [§5.5] The phrase 'Phase 6 demo' is never defined; it is unclear which evaluation phase generated the representative queries in Table 3.
  4. [References] Reference [24] (RAGAS) is cited in §4.2 as the source of the 'compositional multi-hop generation step', but RAGAS is an evaluation toolkit, not a generation method. This citation appears to be a mismatch.
  5. [Abstract and §5.3] The abstract attributes the optimized weight α_triple=0.44 to the system that achieves the 0.9909 recall, but Table 5 reports those optimized weights for v0.7.4, while the headline recall is reported for v0.7.3. The relationship between the optimized variant and the reported evaluation results should be clarified.

Circularity Check

1 steps flagged · score 6.0 of 10

The central multi-hop validation loop is self-referential: Qwen2.5-7B generates the gold answers and also generates the system's answers, so the reported +71.6% answer-similarity advantage is in part a self-agreement score rather than an independent test of triple-augmented reasoning.

  1. self definitional [Section 4.2 (QA benchmark construction) and Section 5.1 (Evaluation Metrics)]
    "We construct a 457-pair QA benchmark through a three-stage pipeline:(1)a small language model (Qwen2.5-7B) generates 200 single-hop question–answer pairs from individual block captions;(2)a compositional multi-hop generation step [24] creates 400 multi-hop pairs requiring evidence from≥2 blocks;(3)an LLM-based quality filter [25] retains 457 high-quality pairs... Answer Similarity: Cosine similarity between the embedding of the generated answer and the gold-standard answer."

    The gold-standard answers are produced by Qwen2.5-7B, and Section 5.1 specifies qwen2.5:7b-instruct-q4km as the answer LLM; the benchmark pairs are also filtered by an LLM judge. Thus the answer-similarity metric compares the evaluated generator with outputs of the same model family, so the reported multi-hop answer-similarity gain (0.5605 vs 0.3267, +71.6%) is partly a self-consistency score. The paper's conclusion that this 'validates triple-augmented compositional reasoning' is therefore not an independent measurement; the evaluation target is defined by the same model used to produce the system's answers.

full rationale

The fusion derivation itself (Eq. 5) is not circular: the four similarity axes are defined independently, and the Bayesian weight optimization (Eq. 6) uses an objective (P@5, diversity D, LLM-judge Q) that is not identical to the headline Retrieval Recall metric; moreover, the 457-pair benchmark results are reported for manual-weight v0.7.3, not for the optimized v0.7.4 weights. There are no load-bearing self-citations: the HippoRAG, GraphRAG, and other prior works cited are external. The main circular step is the evaluation loop: Qwen2.5-7B generates the QA pairs and gold answers, an LLM judge filters them, and Qwen2.5-7B generates the answers being scored, making the flagship multi-hop answer-similarity advantage partly an agreement-with-self artifact. Separately, the paper has serious non-circular validity problems: the reported overall Retrieval Recall 0.9909 is arithmetically incompatible with Table 2's per-type recalls ((169×1.0000 + 288×0.5000)/457 ≈ 0.685), and the +71.6% multi-hop advantage is not a triple-axis ablation since the Naive RAG baseline already shows +53.5%. These are correctness concerns, not circular reductions, so they do not raise the circularity score beyond the self-referential benchmark issue.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

No new physical or formal entities are postulated; 'triple filtering' and 'four-axis fusion' are design patterns assembled from existing components. The load-bearing assumptions are empirical: the benchmark is self-generated, the metric is self-referential, the tuning set is not shown to be disjoint from the test set, and the OCR/triple error rates are accepted without audit.

free parameters (7)
  • Fusion weights (α_t, α_b, α_tr, α_i) v0.7.4 = (0.2675, 0.2903, 0.4422, 0.10)
    Optimized by Optuna/TPE over 50 trials on 40 gold-standard blocks (Section 3.6); the headline recall and multi-hop numbers are measured with these fitted weights.
  • Manual fusion weights v0.7.3 = (0.4, 0.3, 0.2, 0.1)
    Hand-set defaults (Section 3.6), used as the basis for the Bayesian optimization comparison.
  • Chapter-level triple signal weight γ_C = 0.3
    Hand-chosen constant in Eq. (4), not optimized or varied.
  • Volume-level fusion weights = 0.6 · embedding + 0.4 · keyword
    Hand-chosen constants in Eq. (3).
  • Hierarchy truncation n_V, n_C, top-k = nV = 2, nC = 3, k ∈ {5, 20}
    Hand-chosen search-space limits for the coarse-to-fine retriever (Section 3.5).
  • Figure-boost multiplier = ×1.2
    Hand-chosen boost applied to table-type blocks when figure keywords are detected (Section 3.6).
  • Text chunking parameters = 500-character chunks / 100-character overlap
    Hand-chosen segmentation in Section 3.2, affecting block granularity and indexing.
assumptions (6)
  • domain assumption The 457 QA pairs and gold-standard block IDs are correct and complete, at least for 1-hop queries.
    The benchmark is generated by Qwen2.5-7B and filtered by an LLM judge (Section 4.2); no human validation is described. The paper itself concedes multi-hop annotations are incomplete (Section 7), so the risk extends to the whole benchmark.
  • domain assumption Cosine similarity in the hotchpotch static-embedding-japanese space is a valid measure of answer quality.
    Answer quality is computed by embedding the generated and gold answers with the same model used for retrieval (Section 5.1), making the metric self-referential and unvalidated against human judgments.
  • ad hoc to paper The 40 gold-standard blocks used for Optuna tuning are disjoint from the 457-pair evaluation set.
    Without this holdout, the reported 0.9909 recall and +71.6% improvement are in-sample fits. The paper never states that the 40 blocks are disjoint (Sections 3.6, 5.3).
  • domain assumption OpenIE triples extracted by Qwen2.5-7B are accurate enough to serve as a knowledge graph.
    11,414 triples are extracted at an 86% success rate (Section 3.4) with no human audit of triple correctness; downstream claims about the triple axis inherit this error rate.
  • domain assumption The 43 MLIT Japanese disaster PDFs are representative of 'complex PDFs' generally.
    The abstract and Section 8.1 claim a general framework 'applicable beyond the disaster domain'; no cross-domain experiment is reported.
  • domain assumption OCR and caption failures do not bias the evaluation benchmark.
    13.2% of table blocks fail OCR entirely (Section 7), yet the QA benchmark is built from the resulting blocks; the effect of blind spots on question coverage is unquantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal CoLRAG-TF: Triple-Filtered Retrieval for Complex PDFs." pith.science (2026). https://pith.science/paper/LKU7MH5W

@misc{pith2026260720517,
  author       = {Pith},
  title        = {Pith review of: Multimodal CoLRAG-TF: Triple-Filtered Retrieval for Complex PDFs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKU7MH5W}},
  note         = {Machine review of arXiv:2607.20517}
}
abstract

Retrieval-augmented generation (RAG) over heterogeneous PDF collections remains challenging due to multimodal content, domain-specific terminology, and the need for multi-hop reasoning across dispersed evidence. We present Multimodal CoLRAG-TF, a four-axis fusion architecture that integrates dense text embeddings, BM25 keyword matching, knowledge-graph triple filtering, and image-based similarity for robust retrieval over complex documents. Our system constructs a multimodal index of 2,403 blocks extracted from 43 Japanese disaster lesson PDFs, supported by a hybrid OCR pipeline and LLM-based caption generation. To enhance compositional reasoning, we extract 11,414 OpenIE triples and index them with FAISS, enabling sub-second triple lookup and hierarchical propagation of relevance signals. A HippoRAG2-inspired coarse-to-fine retriever (volume $\to$ chapter $\to$ block) narrows the search space before final fusion scoring. Bayesian optimization over fusion weights reveals that the triple axis must dominate ($\alpha_\text{triple} = 0.44$) to counteract lexical bias and sustain multi-hop retrieval quality. Evaluated on a 457-pair benchmark, Multimodal CoLRAG-TF achieves a Retrieval Recall of 0.9909 and a 71.6$\%$ improvement in multi-hop answer similarity over single-hop queries. An image-to-lesson pipeline using a vision LLM further demonstrates the applicability of the approach to visual inputs. These results show that triple-filtered multimodal fusion is essential for structured reasoning over noisy, heterogeneous PDFs and provides a general framework applicable beyond the disaster domain.

Figures

Figures reproduced from arXiv: 2607.20517 by the authors.

Figure 1
Figure 1. Multi-modal CoLRAG-TF pipeline. (a) Offline: disaster PDFs are processed through layout analysis, hybrid OCR (PyMuPDF + Tesseract), and caption generation, producing 2,403 multimodal blocks indexed as text embeddings (FAISS 1024-dim), a BM25 keyword index, and 11,414 knowledge triples. (b) On￾line: a user query is analysed for figure intent; four-axis fusion combines text-embedding, BM25, triple, and image-embedding… view at source ↗
Figure 2
Figure 2. Hybrid OCR pipeline coverage over 38 table [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. visualises the full breakdown, including median similarity which shows the strongest BM25 effect (+11.9% for 1-hop). 1-hop (n=169) Multi-hop (n=288) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Answer Similarity (cosine) +71.6% improvement over 1-hop +11.9% (BM25 effect) Answer Similarity: 1-hop vs Multi-hop Queries v0.7.2 Mean v0.7.3 Mean v0.7.2 Median v0.7.3 Median [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Text score, triple score, and Top-1 fusion [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Per-image retrieval scores for 12 disaster photographs. Solid line: Top-1 fusion score (mean [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Fusion weight evolution from manual configuration (v0.7.3) to Bayesian-optimized values (v0.7.4). The triple weight more than doubles (αtr : 0.20 → 0.44, +121%), becoming the dominant axis. if a flood image is described as “earthquake dam￾age”, the retrieved lesson blo…
Figure 8
Figure 8. Figure 8: Left: Retrieval Recall@5 by question type (v0.7.3). Overall recall is 0.9909; multi-hop recall is capped at 0.5 due to hypothesised annotation incompleteness. Right: Optuna TPE convergence over 50 trials; best-score plateau is reached at trial 14 (objective= 0.41). for…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 8 linked inside Pith

  1. [1]

    Disaster prevention technology basic plan

    Ministry of Land, Infrastructure, Transport and Tourism (MLIT), Japan. Disaster prevention technology basic plan. Technical report, Ministry of Land, Infrastructure, Transport and Tourism (MLIT), River and Civil Engineering Division,

  2. [2]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Pik- tus, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. InAd- vances in Neural Information Processing Sys- tems (NeurIPS), volume 33, pages 9459–9474, 2020

    2020

  3. [3]

    Sentence- BERT: Sentence embeddings using siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence- BERT: Sentence embeddings using siamese BERT-networks. InProceedings of the 2019 Con- ference on Empirical Methods in Natural Lan- guage Processing (EMNLP), pages 3982–3992, 2019

    2019

  4. [4]

    ColBERT: Efficient and effective passage search via contex- tualized late interaction over bert

    Omar Khattab and Matei Zaharia. ColBERT: Efficient and effective passage search via contex- tualized late interaction over bert. InProceedings of the 43rd International ACM SIGIR Confer- ence on Research and Development in Informa- tion Retrieval, pages 39–48, 2020

    2020

  5. [5]

    ColBERTv2: Effective and effi- cient retrieval via lightweight late interaction

    Keshav Santhanam, Omar Khattab, Jon Saad- Falcon, et al. ColBERTv2: Effective and effi- cient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Compu- tational Linguistics (NAACL), pages 3715–3734, 2022

    2022

  6. [6]

    The probabilistic relevance framework: BM25 and beyond

    Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond. Technical Report 4, Foundations and Trends in Information Retrieval, 2009

    2009

  7. [7]

    Retrieval-augmented generation for large lan- guage models: A survey.arXiv preprint arXiv:2312.10997, 2024

    Yunfan Gao, Yun Xiong, Xinyu Gao, et al. Retrieval-augmented generation for large lan- guage models: A survey.arXiv preprint arXiv:2312.10997, 2024

    arXiv 2024

  8. [8]

    Hipporag: Neurobiologically inspired long- term memory for large language models.arXiv preprint arXiv:2405.14831, 2024

    Bernal Jim´ enez Gutierrez, Yiheng Shu, Yu Gu, et al. Hipporag: Neurobiologically inspired long- term memory for large language models.arXiv preprint arXiv:2405.14831, 2024

    arXiv 2024

Show all 26 references
  1. [9]

    From RAG to mem- ory: Non-parametric continual learning for large language models

    Bernal Jim´ enez Guti´ errez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. From RAG to mem- ory: Non-parametric continual learning for large language models. InProceedings of the 42nd 12 International Conference on Machine Learning (ICML), 2025. arXiv:2502.14802

    2025 arXiv

  2. [10]

    From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    Darren Edge, Ha Trinh, Newman Cheng, et al. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    2024 arXiv

  3. [11]

    Pubtables-1m: Towards comprehen- sive table extraction from unstructured docu- ments

    Brandon Smock, Rohith Pesala, and Robin Abraham. Pubtables-1m: Towards comprehen- sive table extraction from unstructured docu- ments. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recog- nition (CVPR), pages 4634–4642, 2022

    2022

  4. [12]

    Lay- outLMv3: Pre-training for document ai with uni- fied text and image masking

    Yupan Huang, Tengchao Lv, Lei Cui, et al. Lay- outLMv3: Pre-training for document ai with uni- fied text and image masking. InProceedings of the 30th ACM International Conference on Mul- timedia, pages 4083–4091, 2022

    2022

  5. [13]

    MuRAG: Multimodal retrieval-augmented gen- erator for open question answering over images and text.arXiv preprint arXiv:2210.02928, 2022

    Wenhu Chen, Hexiang Hu, Xi Chen, et al. MuRAG: Multimodal retrieval-augmented gen- erator for open question answering over images and text.arXiv preprint arXiv:2210.02928, 2022

    2022 arXiv

  6. [14]

    Muhammad Imran, Ferda Ofli, Doina Caragea, and Antonio Torralba. Using artificial intelli- gence and social media multimodal content for disaster response and management: Opportu- nities, challenges, and future directions.Infor- mation Processing & Management, 57(5):102261, 2020

    2020

  7. [15]

    Humaid: Human-annotated dis- aster incidents data from twitter with deep learn- ing benchmarks

    Firoj Alam, Hassan Sajjad, Muhammad Imran, and Ferda Ofli. Humaid: Human-annotated dis- aster incidents data from twitter with deep learn- ing benchmarks. InProceedings of the Interna- tional AAAI Conference on Web and Social Me- dia (ICWSM), volume 15, pages 933–944, 2021

    2021

  8. [16]

    An overview of the tesseract OCR engine

    Ray Smith. An overview of the tesseract OCR engine. InProceedings of the 9th International Conference on Document Analysis and Recogni- tion (ICDAR), pages 629–633, 2007

    2007

  9. [17]

    PyMuPDF: Python bindings for MuPDF

    Artifex Software. PyMuPDF: Python bindings for MuPDF. Software library, 2023. Version 1.23.x,https://pymupdf.readthedocs.io/

    2023

  10. [18]

    Qwen2.5: A party of foundation models.arXiv preprint arXiv:2412.15115, 2024

    Qwen Team. Qwen2.5: A party of foundation models.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  11. [19]

    static-embedding-japanese: A fast CPU-friendly japanese static embedding model

    hotchpotch. static-embedding-japanese: A fast CPU-friendly japanese static embedding model. Hugging Face model repository, 2025.https://huggingface.co/hotchpotch/ static-embedding-japanese

    2025

  12. [20]

    Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2021

    Jeff Johnson, Matthijs Douze, and Herv´ e J´ egou. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2021

    2021

  13. [21]

    Optuna: A next-generation hyperparam- eter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, et al. Optuna: A next-generation hyperparam- eter optimization framework. InProceedings of the 25th ACM SIGKDD International Confer- ence on Knowledge Discovery & Data Mining, pages 2623–2631, 2019

    2019

  14. [22]

    Algorithms for hyper- parameter optimization

    James Bergstra, R´ emi Bardenet, Yoshua Ben- gio, and Bal´ azs K´ egl. Algorithms for hyper- parameter optimization. InAdvances in Neural Information Processing Systems (NeurIPS), vol- ume 24, 2011

    2011

  15. [23]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Sys- tems (NeurIPS), volume 36, 2023

    2023

  16. [24]

    RAGAS: Automated evaluation of retrieval augmented generation

    Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. RAGAS: Automated evaluation of retrieval augmented generation. arXiv preprint arXiv:2309.15217, 2023

    2023 arXiv

  17. [25]

    subject”, “relation

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-judge with MT- bench and chatbot arena.arXiv preprint arXiv:2306.05685, 2023. 13 A Implementation Details A1. OCR Pipeline Configuration Table Transformer is run with con- fidence threshold≥0.7 using the micros...

    2023 arXiv

  18. [2024]

    Available at:https://www.mlit.go.jp/

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.