Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review

Abhinav Manikantha Sai Cheruvu; Aditya Nagori; Ayush Gautam; Ricardo Accorsi Casonatto; Rishikesan Kamaleswaran

REVIEW 3 major objections 6 minor 1 cited by

Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A self-routing agent that dynamically selects between graph-based and vector-based retrieval, refined by direct preference optimization, outperforms a static hybrid RAG baseline on synthetic scientific literature-review questions.

desk verdict A well-described open-source agentic hybrid RAG system whose headline gains are real only on its self-generated benchmark; the evaluation changes two variables at once, so the central claim about dynamic orchestration is not actually demonstrated. read the letter →

arxiv 2508.05660 v1 pith:YZFHG5W3 submitted 2025-07-30 cs.IR cs.AI

Aditya Nagori , Ricardo Accorsi Casonatto , Ayush Gautam , Abhinav Manikantha Sai Cheruvu , Rishikesan Kamaleswaran This is my paper

classification cs.IRcs.AI

keywords agenticRAGhybridretrievalknowledgegraphvectorsearchdirectpreferenceoptimizationscientificliteraturereviewsyntheticbenchmarkuncertaintyquantification

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 aims to show that wrapping a hybrid retrieval-augmented generation pipeline in an autonomous agent improves scientific literature review. The agent decides per query whether to answer via a citation knowledge graph (GraphRAG, using Cypher queries) or via full-text vector and keyword search (VectorRAG), then passes the retrieved context to a generator. The core claim is that dynamic orchestration plus instruction tuning with direct preference optimization (DPO) outperforms a non-agentic hybrid baseline, with gains of +0.63 in vector-store context recall and +0.56 in overall context precision on the paper's synthetic benchmark. If correct, this offers an open-source, reproducible path to more accurate and verifiable automated literature synthesis.

What carries the argument

The load-bearing mechanism is the agentic router: an LLM acting as an agent that holds two tool functions, a Cypher query generator over a Neo4j knowledge graph and an ensemble retriever that fuses sparse keyword search with dense semantic embeddings and then reranks the candidates. On top of this, the paper applies direct preference optimization to a small instruction-tuned generator using only 15 human preference pairs, which steers answers toward being grounded in retrieved context, and it wraps evaluation in a bootstrap procedure with 12 resamples to attach standard errors to all reported metrics.

What would settle it

Run the same agentic pipeline and the non-agentic baseline on an externally written or pre-existing set of scientific literature questions with ground-truth answers, and check whether the DPO-tuned agent still beats the baseline by the reported margins; if the gains shrink to zero or become negative on such a corpus, the core claim fails to generalize.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that an instruction-tuned agent with direct preference optimization substantially outperforms a static hybrid RAG baseline on a purpose-built synthetic benchmark. The agent's routing decision—choosing GraphRAG for structured, relational questions and VectorRAG for full-text semantic questions—turns out to matter more than simply combining both retrievers in a fixed pipeline. The reported gains include +0.63 in VS Context Recall and +0.56 in overall Context Precision, plus improvements in faithfulness, precision, and answer relevance, with slight declines on KG precision and faithfulness suggesting the agent's graph reasoning still has room to improve.

Load-bearing premise

The headline gains are measured on a synthetic benchmark that the system's own pipeline generates—LLaMA-3.3 writes the VectorRAG questions from the system's own chunks, and the GraphRAG questions are templated from the system's own knowledge graph—so the reported 0.63 and 0.56 improvements are only as trustworthy as this benchmark's resemblance to real researcher queries.

Editorial extensions

If this is right

Dynamic retrieval routing can outperform a static hybrid pipeline that always combines knowledge-graph and vector-store results.
Direct preference optimization with a very small number of preference pairs can meaningfully improve faithfulness and context utilization in a RAG generator.
An open-source, Dockerizable implementation means the agentic hybrid RAG stack is reproducible and adaptable to other corpora.
Bootstrapped uncertainty estimates provide a practical way to flag low-confidence or conflicting outputs during literature exploration.
The framework establishes a foundation on which further refinement—better Cypher translation, OCR support, and learned routing policies—could build toward autonomous literature review.

Reading between the lines

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

The reported gains are measured on a benchmark generated by the system's own pipeline, so real-world questions that do not fit cleanly into either retrieval mode would likely show smaller improvements.
Because DPO worked with only 15 preference pairs, small, targeted alignment datasets could be a cheap and general lever for improving RAG faithfulness in other domains.
The routing policy itself is a natural target for reinforcement learning, which the paper lists as future work; a learned router might outperform the few-shot-prompted agent on ambiguous or mixed queries.
The framework could be extended to non-scientific corpora by swapping the knowledge-graph schema and the embedding model while keeping the agentic orchestration layer intact.

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

You should know two things about this paper before diving in. First, it is a competent engineering report: a fully open-source pipeline that combines a Neo4j knowledge graph, FAISS vector store, a LLaMA-3.3 routing agent, and a Mistral-7B generator with a 15-pair DPO fine-tune. Second, the evaluation does not actually support the paper's central claim that dynamic orchestration improves relevance, because the comparison changes two things at once. The baseline is a joint semantic search over both stores; the agentic VectorRAG adds BM25 and Cohere reranking, and the agentic GraphRAG uses Cypher over a structured graph. Any of those upgrades could explain the gains, and no ablation isolates the routing decision. The stress-test note is correct on this point.

What is genuinely new is narrow: the specific integration of agentic routing between GraphRAG and VectorRAG, with DPO on a small preference set, evaluated on a homemade 40-question benchmark. That is not a new method, and every component is already in the cited literature. The paper does several things well. The system description is clear and mostly reproducible (code is linked). The authors use standard RAG evaluation metrics and give bootstrapped standard errors. They are also honest in Section 6 that the synthetic benchmark may not capture the full complexity of real scientific inquiries. That is more candor than many papers show.

Now the soft spots, in proportion. The benchmark is self-generated: the 20 VectorRAG questions come from LLM-generated questions on the system's own chunks, and the 20 GraphRAG questions are templated from the system's own KG. So the 0.63 and 0.56 gains measure performance on questions derived from the system's data. That is a circularity problem, though not a fatal one for an engineering proof-of-concept. Second, the baseline is weak: it lacks BM25, reranking, and Cypher, so it is a strawman rather than a meaningful comparison. Third, the DPO is described in only a few lines; no hyperparameters, training details, or data are given, and the VS precision actually drops by 0.10 after DPO. That is a red flag that the benefit of DPO is narrower than the abstract implies. Fourth, the stats are thin: 40 questions total, 12 bootstrap resamples, and results are shown only in a figure, with no tables of per-metric confidence intervals. The standard error claim (≤0.10) is not enough to evaluate significance.

Who is this for? Practitioners building RAG systems might get ideas for tooling and orchestration, but researchers looking for evidence about what makes agentic RAG work will be disappointed. The paper deserves a serious referee because the system is real and the evaluation is honest, but it needs a major revision with ablations, an external benchmark, and fuller experimental reporting. Send it to peer review, but push for those changes.

My recommendation: engage with it as a system paper, not as a demonstration of the agency hypothesis.

Referee Report

3 major / 6 minor

Summary. The paper presents an open-source agentic hybrid RAG framework for scientific literature review. The pipeline ingests bibliometric data from PubMed, arXiv, and Google Scholar, builds a Neo4j knowledge graph and a FAISS vector store, and uses a LLaMA-3.3-70B agent to choose between GraphRAG (natural-language-to-Cypher querying) and VectorRAG (BM25 plus dense retrieval with reranking). The response generator is Mistral-7B-Instruct-v0.3, optionally refined with Direct Preference Optimization using 15 preference pairs. Evaluation is performed on a synthetic 40-question benchmark (20 VS questions generated from the system's own chunks and 20 KG questions templated from its own knowledge graph), comparing a non-agentic joint semantic search baseline, the agentic system, and the DPO-tuned agentic system on RAGAS-style metrics. The abstract and Section 5 report large gains for the DPO agent over baseline, notably +0.63 in VS Context Recall and +0.56 in overall Context Precision, with the claim that dynamic orchestration improves relevance, reduces hallucinations, and supports reproducibility.

Significance. If the claims were properly supported, the framework would be a useful open-source contribution to automated literature review, particularly for combining structured and unstructured retrieval in a transparent, Dockerizable pipeline. The paper's strengths include its reproducible code availability, explicit uncertainty quantification through bootstrapping, and honest acknowledgment of limitations in Section 6. However, the current evaluation does not establish the central causal claim that agentic orchestration is the source of the reported gains, because the baseline differs in retrieval machinery as well as routing, and the benchmark is generated from the system's own data and model family. The paper is therefore a promising systems demonstration, but its headline quantitative claims need substantially stronger evidence before they can be accepted.

major comments (3)

[§4.2, §5] The headline comparison does not isolate agentic routing. The baseline is described as a non-agentic joint semantic search over both the VS and KG, while the agentic VectorRAG adds BM25 sparse retrieval and Cohere reranking (Section 3.3.2) and the agentic GraphRAG adds Cypher-based structured query execution (Section 3.3.1). The reported gains of 0.63 in VS Context Recall and 0.56 in overall Context Precision therefore conflate the routing policy with the retrieval machinery used by each tool. Because no ablation holds the underlying retrievers fixed and varies only the orchestration decision, the paper's central attribution of these gains to dynamic orchestration is not supported. I would ask for an additional baseline that uses the same BM25+rerank and Cypher tools but with fixed or random routing, or per-tool comparisons that match retrieval components.
[§3.5, §4.1] The evaluation benchmark is generated by the system's own pipeline, which makes the reported absolute gains difficult to interpret. The 20 VectorRAG questions are produced by prompting a LLaMA-3.3 instance with 20 chunks sampled from the system's own vector store (Section 4.1.1), and the 20 GraphRAG questions are templated from the system's own knowledge graph (Section 4.1.2). Since the routing agent is also built on LLaMA-3.3-70B, the benchmark tests the system on questions derived from its own data and model family, not on independently curated scientific queries. The authors acknowledge in Section 6 that the synthetic benchmark may not capture the full complexity of real scientific inquiries, but because every headline number in Section 5 rests on this benchmark, this limitation is load-bearing rather than incidental. I would request at least one evaluation on an external QA or literature-review dataset, or a human-annotated sample of real researcher queries.
[§5, §4.3] The statistical support for the reported gains is not fully reported. The results are presented only as a bar chart (Figure 5); there is no table of per-metric means, and the bootstrap procedure (Section 4.3) uses only 12 resamples, yet Section 6 asserts 'Standard Error<=0.10' without giving the corresponding confidence intervals or standard deviations for the individual metrics. Given the small benchmark size (40 questions) and the multiple metrics reported, the reader cannot verify which differences are significant. Please provide a full results table with mean, bootstrap standard error, and confidence interval for every metric and system, and state the resampling rule (with or without replacement) and how stratification by VS/KG is handled.

minor comments (6)

[§2.1, §2.2] References [15] and [16] appear to be used for different works in different sections; for example, [15] is cited for both a scoping review of automation tools and the Iancarelli et al. citation-network study. Please re-check the citation numbering and ensure each reference is matched to its intended source.
[§4.2, Eq. (4)] Equation (4) uses K both as the summation upper limit and inside the definition of CP, and the role of the relevance indicator v_k is not defined clearly. Please clarify the notation and specify how 'total number of relevant items in the top K results' is computed.
[§5, Figure 5] Figure 5 is difficult to use for verifying the numerical claims in the text. Adding a companion table with the exact metric values for each system would markedly improve the paper's transparency.
[§3.4] The paper states that DPO used '15 high-quality, human-annotated preference pairs' but provides no details on the annotation protocol, annotator agreement, or how these pairs were selected. Please include this information to support the reproducibility of the DPO step.
[§4.3] The bootstrap description says '12 resamplings, each consisting of 20 randomly selected questions' but does not state whether sampling is with replacement or how the 10 VS / 10 KG split is enforced in each resample. Please clarify.
[§8, §9] The code availability statement gives a repository URL but no license, version, or commit identifier. For a paper whose stated contribution includes reproducibility, please include a license and a pinned version.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the evaluation is self-generated and confounded, but the headline gains are empirical measurements rather than predictions that reduce to their own inputs by construction.

full rationale

The paper's central claim is an empirical engineering result, not a mathematical derivation. The synthetic benchmark is built from the system's own VS chunks and KG (Sections 3.5 and 4.1), and the authors themselves caution in Section 6 that it 'may not capture the full complexity of real scientific inquiries.' That is a legitimate external-validity and benchmark-circularity concern, but it is not the kind of by-construction reduction this pass targets: no parameter is fitted to the reported metrics and then renamed as a prediction, and the DPO model is trained on 15 human preference pairs rather than on the evaluation labels. The main uncontrolled comparison is that the agentic arm changes retrieval machinery (BM25, Cohere reranking, Cypher) together with routing, so the 0.63 VS Context Recall and 0.56 Context Precision gains cannot be cleanly attributed to dynamic orchestration; this is a confounded-experiment defect, not circular reasoning. The only self-citation, Ref. [8] (which includes co-author Casonatto), appears in a background enumeration of prior machine-learning tools for systematic review and is not load-bearing for any central claim. Thus, under the strict reduction standard, no circular step can be exhibited; the score reflects only the minor, non-load-bearing self-citation.

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

The central claims depend on a set of design choices and benchmark assumptions. No new physical or conceptual entities are introduced; the synthetic benchmark is a dataset, not an entity. The ledger counts four hand-chosen parameters and four domain assumptions that are not independently validated in the paper.

free parameters (4)

Cosine similarity cutoff (Q3) = third quartile of the score distribution
Keeps only the top 25% of papers by TF-IDF cosine similarity to the query (Section 3); changes corpus composition and benchmark difficulty.
Chunk size / overlap = 2024 characters, 50-character overlap
Text segmentation hyperparameters in Section 3.2; affects retrieval granularity and downstream metrics.
Retrieval top-k = 5 candidates per retriever
Both BM25 and semantic search return top five chunks before reranking (Section 3.3.2).
DPO preference pair count = 15
The DPO training set size in Section 3.4; effect on measured gains is not characterized.

assumptions (4)

domain assumption Open-access data from PubMed, arXiv, and Google Scholar APIs is sufficient to build a literature review corpus for the target domain.
Underpins the entire data collection pipeline in Section 3.
domain assumption The synthetic benchmark, generated from the system's own KG and chunks, accurately mimics real-world literature review queries.
The paper's central performance claims are evaluated only on this benchmark (Sections 3.5 and 6).
domain assumption all-MiniLM-L6-v2 and Cohere rerank-english-v3.0 give relevance scores reliable enough for the reported conclusions.
Used for vector retrieval and reranking in Sections 3.2 and 3.3.2; no error analysis over embedding model choice is provided.
domain assumption The non-agentic baseline that concatenates vector-search results is a representative baseline for static hybrid RAG.
Baseline definition in Sections 4.2 and 5; a stronger static hybrid could reduce the observed gains.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review." pith.science (2026). https://pith.science/paper/YZFHG5W3

@misc{pith2026250805660,
  author       = {Pith},
  title        = {Pith review of: Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YZFHG5W3}},
  note         = {Machine review of arXiv:2508.05660}
}

read the original abstract

The surge in scientific publications challenges traditional review methods, demanding tools that integrate structured metadata with full-text analysis. Hybrid Retrieval Augmented Generation (RAG) systems, combining graph queries with vector search offer promise but are typically static, rely on proprietary tools, and lack uncertainty estimates. We present an agentic approach that encapsulates the hybrid RAG pipeline within an autonomous agent capable of (1) dynamically selecting between GraphRAG and VectorRAG for each query, (2) adapting instruction-tuned generation in real time to researcher needs, and (3) quantifying uncertainty during inference. This dynamic orchestration improves relevance, reduces hallucinations, and promotes reproducibility. Our pipeline ingests bibliometric open-access data from PubMed, arXiv, and Google Scholar APIs, builds a Neo4j citation-based knowledge graph (KG), and embeds full-text PDFs into a FAISS vector store (VS) using the all-MiniLM-L6-v2 model. A Llama-3.3-70B agent selects GraphRAG (translating queries to Cypher for KG) or VectorRAG (combining sparse and dense retrieval with re-ranking). Instruction tuning refines domain-specific generation, and bootstrapped evaluation yields standard deviation for evaluation metrics. On synthetic benchmarks mimicking real-world queries, the Instruction-Tuned Agent with Direct Preference Optimization (DPO) outperforms the baseline, achieving a gain of 0.63 in VS Context Recall and a 0.56 gain in overall Context Precision. Additional gains include 0.24 in VS Faithfulness, 0.12 in both VS Precision and KG Answer Relevance, 0.11 in overall Faithfulness score, 0.05 in KG Context Recall, and 0.04 in both VS Answer Relevance and overall Precision. These results highlight the system's improved reasoning over heterogeneous sources and establish a scalable framework for autonomous, agentic scientific discovery.

Figures

Figures reproduced from arXiv: 2508.05660 by the authors.

**Figure 2.** KG structure overview highlighting metadata enti [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** VS retrieval workflow combining keyword and se [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Workflow of KG retrieval, translating natural [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

**Figure 5.** Figure 5: Model comparison on faithfulness (F), answer rele [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in 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

HVM-GraphRAG: Holistic-View Multimodal Graph Retrieval-Augmented Generation on Complex Document
cs.IR 2026-07 conditional novelty 5.0 of 10

A holistic-view multimodal GraphRAG that conflict-resolves concept indices and retrieves via compact concept anchors plus modality-grouped evidence, beating strong baselines on three complex-document QA sets.

Reference graph

Works this paper leans on

41 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

Alexandra Bannach-Brown, Piotr Przybyła, James Thomas, Andrew SC Rice, Sophia Ananiadou, Jing Liao, and Malcolm Robert Macleod. 2019. Machine learning algorithms for systematic review: reducing workload in a preclinical review of animal studies and reducing human screening error.Systematic reviews 8 (2019), 1–12

work page 2019
[2]

Taylor Berg-Kirkpatrick, David Burkett, and Dan Klein. 2012. An empirical investigation of statistical significance in NLP. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computa- tional Natural Language Learning . Association for Computational Linguistics, 995–1005

work page 2012
[3]

Lutz Bornmann, Robin Haunschild, and Rüdiger Mutz. 2021. Growth rates of modern science: a latent piecewise growth curve approach to model publication numbers from established and new literature databases. Humanities and Social Sciences Communications 8, 1 (2021), 1–15

work page 2021
[4]

Xiaojun Chen, Shengbin Jia, and Yang Xiang. 2020. A review: Knowledge reason- ing over knowledge graph. Expert Systems with Applications 141 (March 2020), 112948. https://doi.org/10.1016/j.eswa.2019.112948

work page arXiv 2020
[5]

Rong Cheng, Jinyi Liu, Yan Zheng, Fei Ni, Jiazhen Du, Hangyu Mao, Fuzheng Zhang, Bo Wang, and Jianye Hao. 2025. DualRAG: A Dual-Process Approach to Integrate Reasoning and Retrieval for Multi-Hop Question Answering. arXiv preprint arXiv:2504.18243 (2025). Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review

arXiv 2025
[6]

Paul F Christiano et al. 2017. Deep reinforcement learning from human prefer- ences. Advances in Neural Information Processing Systems (2017)

work page 2017
[7]

Yung-Sung Chuang, Wei Fang, Shang-Wen Li, Wen-tau Yih, and James Glass. 2023. Expand, Rerank, and Retrieve: Query Reranking for Open-Domain Question Answering. https://doi.org/10.48550/arXiv.2305.17080 arXiv:2305.17080 [cs]

arXiv doi:10.48550/arxiv.2305.17080 2023
[8]

da Costa

Jaqueline Coelho, Guilherme Bispo, Guilherme Vergara, Gabriela Saiki, An- dré Serrano, Li Weigang, Clovis Neumann, Patricia Martins, Welber Santos de Oliveira, Angela Albarello, Ricardo Casonatto, Patrícia Missel, Roberto Medeiros Junior, Jefferson Gomes, Carlos Rosano-Peña, and Caroline F. da Costa

work page

Show all 41 references

[9]

Ngo Cong-Lem, Ali Soyoof, and Diki Tsering. 2025. A systematic review of the limitations and associated opportunities of ChatGPT. International Journal of Human–Computer Interaction 41, 7 (2025), 3851–3866

2025
[10]

Vimal Kumar Dubey and Amit Kumar Saxena. 2017. A Cosine-Similarity Mutual- Information Approach for Feature Selection on High Dimensional Datasets:. Journal of Information Technology Research 10, 1 (Jan. 2017), 15–28. https: //doi.org/10.4018/JITR.2017010102

2017 doi
[11]

Tibshirani

Bradley Efron and Robert J. Tibshirani. 1993. An Introduction to the Bootstrap . Chapman & Hall/CRC, New York

1993
[12]

Shahul Es, Jithin James, Luis Espinosa-Anke, and Steven Schockaert. 2025. Ragas: Automated Evaluation of Retrieval Augmented Generation. https://doi.org/10. 48550/arXiv.2309.15217 arXiv:2309.15217 [cs]

2025 doi arXiv
[13]

D Gunawan, C A Sembiring, and M A Budiman. 2018. The Implementation of Cosine Similarity to Calculate Text Relevance between Two Documents. Journal of Physics: Conference Series 978 (March 2018), 012120. https://doi.org/10.1088/ 1742-6596/978/1/012120

2018
[14]

Binglan Han, Thomas Susnjak, and Anuj Mathrani. 2024. Automating System- atic Literature Reviews with Retrieval-Augmented Generation: A Comprehen- sive Overview. Applied Sciences 14, 19 (2024), 9103. https://doi.org/10.3390/ app14199103

2024
[15]

Alessia Iancarelli, Thomas F Denson, Chun-An Chou, and Ajay B Satpute. 2022. Using citation network analysis to enhance scholarship in psychological science: A case study of the human aggression literature.PLOS ONE 17, 4 (2022), e0266513. https://doi.org/10.1371/journal.pone.0266513

2022 doi
[16]

Mohamad Yaser Jaradeh, Allard Oelen, Manuel Prinz, Markus Stocker, and Sören Auer. 2022. Open Research Knowledge Graph: A System Walkthrough. arXiv preprint arXiv:2206.01439 (2022). https://arxiv.org/abs/2206.01439

2022 arXiv
[17]

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys 55, 12 (2023), 1–38

2023
[18]

Emily Kenul, Margaret Black, Drew Massey, Zachary Havelka, Mawia Henkai, Kyle Gavin, and Luke Shellhorn. 2024. Replicant framework for synthetic data generation. In Synthetic Data for Artificial Intelligence and Machine Learning: Tools, Techniques, and Applications II, Kimberl...

2024 doi
[19]

Hanan Khalil, Daniel Ameen, and Armita Zarnegar. 2022. Tools to support the automation of systematic reviews: a scoping review. Journal of Clinical Epidemiology 144 (2022), 22–42. https://doi.org/10.1016/j.jclinepi.2021.12.005

2022 doi
[20]

Vimal Kumar Dubey and Amit Kumar Saxena. 2016. Cosine similarity based filter technique for feature selection. In 2016 International Conference on Control, Computing, Communication and Materials (ICCCCM) . IEEE, Allahbad, India, 1–6. https://doi.org/10.1109/ICCCCM.2016.7918222

2016
[21]

Jakub Lála, Odhran O’Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G Rodriques, and Andrew D White. 2023. PaperQA: Retrieval-Augmented Gen- erative Agent for Scientific Research. arXiv preprint arXiv:2312.07559 (2023). https://arxiv.org/abs/2312.07559

2023 arXiv
[22]

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and Sebastian Riedel. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv preprint arXiv:2005.11...

2020 arXiv
[23]

Linyu Liu, Yu Pan, Xiaocheng Li, and Guanting Chen. 2024. Uncertainty Estima- tion and Quantification for LLMs: A Simple Supervised Approach. arXiv preprint arXiv:2404.15993 (2024). https://arxiv.org/abs/2404.15993

2024 arXiv
[24]

Renjie Liu, Haitian Jiang, Xiao Yan, Bo Tang, and Jinyang Li. 2025. PolyG: Effective and Efficient GraphRAG with Adaptive Graph Traversal. https: //doi.org/10.48550/arXiv.2504.02112 arXiv:2504.02112 [cs]

2025 doi
[25]

André Luiz Marques Serrano, Gabriela Mayumi Saiki, Carlos Rosano-Penã, Gabriel Arquelau Pimenta Rodrigues, Robson De Oliveira Albuquerque, and Luis Javier García Villalba. 2024. Bootstrap Method of Eco-Efficiency in the Brazilian Agricultural Industry. Systems 12, 4 (April 202...

2024 doi
[26]

Mehrnaz Mostafapour, Jacqueline H Fortier, Karen Pacheco, Heather Murray, and Gary Garber. 2024. Evaluating Literature Reviews Conducted by Humans Versus ChatGPT: Comparative Study. Jmir ai 3 (2024), e56537

2024
[27]

Long Ouyang et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155 (2022)

2022 arXiv
[28]

James Jie Pan, Jianguo Wang, and Guoliang Li. 2024. Vector Database Man- agement Techniques and Systems. In Companion of the 2024 International Con- ference on Management of Data . ACM, Santiago AA Chile, 597–604. https: //doi.org/10.1145/3626246.3654691

2024
[29]

Nicholas Rossi, Juexin Lin, Feng Liu, Zhen Yang, Tony Lee, Alessandro Magnani, and Ciya Liao. 2024. Relevance Filtering for Embedding-based Retrieval. In Pro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management. ACM, Boise ID USA, 4828–483...

2024 doi
[30]

Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Naya...

2021 arXiv
[31]

Bhaskarjit Sarmah, Benika Hall, Rohan Rao, Sunil Patel, Stefano Pasquali, and Dhagash Mehta. 2024. HybridRAG: Integrating Knowledge Graphs and Vector Retrieval Augmented Generation for Efficient Information Extraction. https: //doi.org/10.48550/arXiv.2408.04948 arXiv:2408.04948 [cs]

2024 doi arXiv
[32]

Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. https://doi.org/ 10.48550/arXiv.2501.09136 arXiv:2501.09136 [cs]

2025 doi arXiv
[33]

Karan Singhal, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, Perry Payne, Martin Seneviratne, Paul Gamble, Chris Kelly, Nathaneal Scharli, Aakanksha Chowdhery, Philip Mansfield, Blaise Ague...

2022 doi arXiv
[34]

Ray Smith. 2007. An overview of the Tesseract OCR engine. Proceedings of the Ninth International Conference on Document Analysis and Recognition (ICDAR

2007
[35]

Nisan Stiennon et al. 2020. Learning to summarize with human feedback. In Advances in Neural Information Processing Systems

2020
[36]

Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, and Xuechen Li
[37]

Guy Tsafnat, Paul Glasziou, George Karystianis, and Enrico Coiera. 2018. Auto- mated screening of research studies for systematic reviews using study charac- teristics. Systematic reviews 7 (2018), 1–9

2018
[38]

https://crfm

Alpaca: A Strong, Replicable Instruction-Following Model. https://crfm. stanford.edu/2023/03/13/alpaca.html

2023
[39]

Zongmeng Zhang, Wengang Zhou, Jiaxin Shi, and Houqiang Li. 2023. Hybrid and Collaborative Passage Reranking. InFindings of the Association for Computational Linguistics: ACL 2023. 14003–14021. https://doi.org/10.18653/v1/2023.findings- acl.880 arXiv:2305.09313 [cs]

2023 arXiv
[40]

Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2022. Finetuned Language Models Are Zero-Shot Learners. In International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2109.01652 arXiv...

2022 arXiv
[2023]

In Proceedings of the 19th International Conference on Web Information Systems and Technologies - WEBIST

Enhancing Industrial Productivity Through AI-Driven Systematic Lit- erature Reviews. In Proceedings of the 19th International Conference on Web Information Systems and Technologies - WEBIST . INSTICC, SciTePress, 472–479. https://doi.org/10.5220/0012235000003584

doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Alexandra Bannach-Brown, Piotr Przybyła, James Thomas, Andrew SC Rice, Sophia Ananiadou, Jing Liao, and Malcolm Robert Macleod. 2019. Machine learning algorithms for systematic review: reducing workload in a preclinical review of animal studies and reducing human screening error.Systematic reviews 8 (2019), 1–12

work page 2019

[2] [2]

Taylor Berg-Kirkpatrick, David Burkett, and Dan Klein. 2012. An empirical investigation of statistical significance in NLP. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computa- tional Natural Language Learning . Association for Computational Linguistics, 995–1005

work page 2012

[3] [3]

Lutz Bornmann, Robin Haunschild, and Rüdiger Mutz. 2021. Growth rates of modern science: a latent piecewise growth curve approach to model publication numbers from established and new literature databases. Humanities and Social Sciences Communications 8, 1 (2021), 1–15

work page 2021

[4] [4]

Xiaojun Chen, Shengbin Jia, and Yang Xiang. 2020. A review: Knowledge reason- ing over knowledge graph. Expert Systems with Applications 141 (March 2020), 112948. https://doi.org/10.1016/j.eswa.2019.112948

work page arXiv 2020

[5] [5]

Rong Cheng, Jinyi Liu, Yan Zheng, Fei Ni, Jiazhen Du, Hangyu Mao, Fuzheng Zhang, Bo Wang, and Jianye Hao. 2025. DualRAG: A Dual-Process Approach to Integrate Reasoning and Retrieval for Multi-Hop Question Answering. arXiv preprint arXiv:2504.18243 (2025). Open-Source Agentic Hybrid RAG Framework for Scientific Literature Review

arXiv 2025

[6] [6]

Paul F Christiano et al. 2017. Deep reinforcement learning from human prefer- ences. Advances in Neural Information Processing Systems (2017)

work page 2017

[7] [7]

Yung-Sung Chuang, Wei Fang, Shang-Wen Li, Wen-tau Yih, and James Glass. 2023. Expand, Rerank, and Retrieve: Query Reranking for Open-Domain Question Answering. https://doi.org/10.48550/arXiv.2305.17080 arXiv:2305.17080 [cs]

arXiv doi:10.48550/arxiv.2305.17080 2023

[8] [8]

da Costa

Jaqueline Coelho, Guilherme Bispo, Guilherme Vergara, Gabriela Saiki, An- dré Serrano, Li Weigang, Clovis Neumann, Patricia Martins, Welber Santos de Oliveira, Angela Albarello, Ricardo Casonatto, Patrícia Missel, Roberto Medeiros Junior, Jefferson Gomes, Carlos Rosano-Peña, and Caroline F. da Costa

work page