An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling

Anh Nguyen-Duc; Anh Quang Nguyen; Bao Anh Tran; Chien Vu Manh; Luan Le Chi; Viet Phuong Ngo

REVIEW 4 major objections 6 minor 14 references

An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling

T0 review · 4 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 deployed multi-agent retrieval-augmented chatbot sustained 92 percent accuracy on 6,000 real admissions queries.

desk verdict Genuine deployment case study with valuable cost data, but the 92% accuracy and 1.45% hallucination claims need independent evaluation before they can be quoted. read the letter →

arxiv 2507.11272 v1 pith:BX44OVGI submitted 2025-07-15 cs.SE cs.IR

Anh Nguyen-Duc , Chien Vu Manh , Bao Anh Tran , Viet Phuong Ngo , Luan Le Chi , Anh Quang Nguyen This is my paper

classification cs.SEcs.IR

keywords retrieval-augmentedgenerationmulti-agentsystemshybridretrievaluniversityadmissionscounselinghallucinationreductionGPT-4ominiVietnameselow-resourcelanguagecasestudy

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

MARAUS is a conversational system for university admissions counseling built on hybrid retrieval (keyword plus semantic vector search), an LLM re-ranking step, and a multi-agent pipeline that routes each question to a specialized handler. The paper's central claim is that this combination can run reliably in real, high-stakes admissions work: across two weeks at a Hanoi university it fielded over 6,000 genuine user questions, achieved an average 92 percent accuracy, cut hallucination from a 15 percent LLM-only baseline to 1.45 percent, and answered in under 4 seconds on average. The paper also reports that the two-week live deployment cost only 11.58 USD in LLM inference charges using GPT-4o mini. If these figures hold, they would show that domain-specific agentic RAG is practical and affordable in low-resource educational settings, not just on benchmarks.

What carries the argument

The central object is the hybrid retrieval–re-ranking pipeline the paper calls Hybrid RAG: BM25 keyword search and FAISS dense-vector search run in parallel, their results are merged, and a GPT-4o mini zero-shot cross-encoder—a prompt that scores each query–passage pair for relevance—keeps only the top two passages. A multi-agent coordinator decides which of four specialized pipelines handles the question, and a post-processor discards any generated answer that lacks a passage citation, regenerating it with penalized decoding. The paper credits this citation-integrity enforcement with the drop from 15 percent to 1.45 percent hallucination, and the re-ranking step with a 38 percent reduction in false positives versus raw FAISS retrieval.

What would settle it

Re-label a random sample of the 6,079 logged question-answer pairs with independent annotators working from a pre-registered rubric and blind to the system's own labels; if the independent accuracy falls well below 92 percent, or inter-annotator agreement is low, the headline result does not survive.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a multi-agent, hybrid-retrieval RAG system can sustain near-92 percent accuracy on real admissions queries while suppressing hallucination to 1.45 percent, in contrast to a plain LLM baseline at 70 percent precision and 15 percent hallucination. The system routes queries to four agent pipelines—information search, score calculation, recommendation, and a general fallback—then retrieves evidence with combined BM25 and FAISS semantic search, re-ranks with GPT-4o mini as a zero-shot cross-encoder, and enforces citation integrity in the final GPT-4o response. On the 100-pair benchmark, hybrid RAG reached 0.985 precision, 0.89 recall, and 3.75 seconds average response time. On the live data, accuracy varied by day from 87 to 94 percent, with the hardest categories being multi-turn personalization and ambiguous subjective questions. The paper interprets these results as evidence that domain orchestration plus hybrid retrieval is what separates a useful admissions bot from an off-the-shelf LLM.

Load-bearing premise

The 92 percent accuracy figure depends on the authors' manual inspection of 6,079 question-answer pairs; if that inspection was lenient or internally inconsistent, the headline accuracy could be substantially overstated.

Editorial extensions

If this is right

If the 92 percent live accuracy holds, admissions offices with similar FAQ and regulation corpora can field a first-line chatbot that answers the majority of repetitive queries without a human officer.
The 1.45 percent hallucination rate, achieved by enforcing that every answer cite a retrieved passage, suggests citation-integrity post-processing is a cheap and effective guardrail for Vietnamese-language responses.
Because the two-week inference cost was 11.58 USD on GPT-4o mini, the per-question cost is low enough that the architecture is plausible for other low-budget institutions, not only well-funded ones.
The system's weakest live performance on multi-turn personalization and ambiguous subjective questions implies that FAQ-style retrieval alone is insufficient, and that the agent architecture must add conversation memory and softer reasoning for those categories.
The reported gains over LLM-only baselines quantify the value of retrieval grounding in a domain where rules change yearly and answers depend on applicant-specific metadata.

Reading between the lines

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

The paper's own data point to multi-turn personalization and ambiguous/subjective questions as the source of most wrong answers; a direct next step would be to test whether adding conversation memory and a dedicated intent-disambiguation step closes that gap.
The 11.58 USD figure covers only LLM inference during two weeks; institutions should read it as a lower bound on total cost once indexing, hosting, and evaluation labor are included.
The architecture's components—BM25, FAISS, a re-ranking prompt, citation enforcement—are all standard enough that the same recipe could be applied to other regulation-heavy counseling domains such as financial aid or visa advice.
A blind, pre-registered re-labeling of the 6,079 pairs would turn the headline 92 percent from a self-assessment into a replicable number.

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

Punchline: this is a genuine deployment study of a multi-agent RAG system for university admissions in Vietnam, with real traffic and real cost data. The 92% accuracy and 1.45% hallucination numbers, though, rest on measurement choices that make them hard to trust as reported.

What is new and good: MARAUS was run for two weeks at UTT, processed 6,079 interactions, and cost $11.58 total with GPT-4o mini. That scale and cost accounting is rare in the RAG literature, most of which stops at 100-pair lab tests. The system design is sensible: hybrid retrieval (BM25 + FAISS), GPT-4o-mini re-ranking, multi-agent routing for score calculations, recommendations, and general queries. The paper also documents preprocessing, chunking, and infrastructure choices well enough to reproduce the build.

Where it gets soft. The accuracy figure is based on the coauthors manually inspecting all 6,079 question-answer pairs, with no rubric, no blinding, no second annotator, and no sampling plan. That is a real measurement risk; the paper itself acknowledges it in Section 3.3 but then does not mitigate it. The hallucination reduction is even trickier. Table 2 comes from a separate 100-pair experiment, not the deployed logs, so the 15% to 1.45% comparison is not apples-to-apples. And the post-processor in Section 4.4 discards any answer without a passage citation. If "hallucination" is effectively measured as missing a citation, the improvement is partly by construction. The paper never defines the hallucination metric precisely enough to rule that out. Finally, Section 3.3 promises a public codebase and test dataset, but no link appears anywhere in the text, so the 6,079 pairs cannot currently be re-annotated or audited.

None of this makes the deployment worthless. The cost data, the query taxonomy, and the lessons about operating in Vietnamese with limited infrastructure are useful. It is just that the headline numbers should not be quoted yet.

Who this is for: anyone building or evaluating conversational AI in low-resource settings, and researchers studying RAG deployments. The paper deserves a serious referee; the evaluation needs major revision, with independent annotation, a defined hallucination metric, and a same-setting baseline.

Recommendation: send to peer review. It is a useful case study, but it needs heavy revision before the empirical claims can be trusted.

Referee Report

4 major / 6 minor

Summary. The paper presents MARAUS, a multi-agent hybrid retrieval-augmented generation (RAG) system for university admissions counseling, deployed at the University of Transport Technology (UTT) in Hanoi. The authors report a two-week real-world deployment with over 6,000 user interactions, claiming 92% average accuracy, a reduction in hallucination rate from 15% (LLM-only) to 1.45% (hybrid pipeline), average response times below 4 seconds, and a two-week operational cost of 11.58 USD using GPT-4o mini. The manuscript describes the system architecture, retrieval and re-ranking pipeline, post-processing with citation enforcement, and a case-study evaluation. The main evidence for the headline claims is a manual inspection of 6,079 question-answer pairs by the coauthors and a separate 100-pair controlled experiment comparing LLM-only, RAG+re-rank, and hybrid configurations.

Significance. If the reported results are valid, the paper provides a valuable real-world data point for deploying agentic RAG systems in low-resource educational settings, with unusually concrete cost and latency figures. The study's strengths include the use of real operational logs, a clearly described multi-agent architecture, explicit retrieval parameters, and a comparison against LLM-only and RAG baselines. However, the central quantitative claims rest on measurement practices that are currently under-specified and partly circular. The accuracy figure depends on unblinded, coauthor-only annotation without a rubric or inter-rater reliability measure, and the hallucination reduction is entangled with a post-processing rule that discards uncited answers. These issues must be addressed before the empirical contribution can be considered reliable.

major comments (4)

[Section 5.1, 5.2, 5.3] The 92% accuracy claim is based on 'manually inspected by coauthors of this paper' (Section 5.1), but the manuscript provides no annotation rubric, no blinding procedure, no inter-rater reliability statistic, and no error bars around the daily accuracy figures reported in Section 5.3 and Figure 2. Because the annotators are the system builders, the accuracy estimate is vulnerable to confirmation bias, and the reported range of 87% to 94% lacks any measure of annotation uncertainty. The paper should specify the rubric used to classify 'correct' versus 'incorrect' answers, use independent annotators, and report inter-rater agreement (e.g., Cohen's kappa).
[Section 4.4, Table 2] The hallucination-rate reduction is partly circular as reported. Section 4.4 states that answers lacking at least one passage citation are discarded and regenerated, while Table 2 reports 'Hallucination Rate' values without defining how hallucination was operationalized in the 100-pair experiment. If hallucination is measured as absence of citation, or if citation presence influences the evaluator's judgment, then the reduction from 15% to 1.45% is at least partly enforced by the pipeline rather than independently measured. The authors must define the hallucination metric explicitly, separate citation integrity from factual correctness, and ideally measure hallucination on the deployed 6,079 interactions using blind human evaluation.
[Section 5.1 versus Section 4.5] The abstract and key observations pair the 92% deployment accuracy with the 1.45% hallucination rate, but these figures come from different evaluations: the accuracy is from manual inspection of 6,079 real interactions (Section 5.1), while the hallucination rate is from a 100-pair controlled experiment (Section 4.5, Table 2). The paper does not report the hallucination rate on the deployed logs, and it does not compare MARAUS against an LLM-only baseline on the 6,079 pairs. To support the 'substantial improvements over LLM-only baselines' claim, the same metrics should be measured on the same corpus, or the scope of each claim should be stated more carefully.
[Section 3.3] The paper claims that 'we have documented all configuration parameters in detail and publicly released our codebase and test dataset,' but no repository link, DOI, or dataset identifier appears anywhere in the manuscript. Without public access to the code and the 6,079 annotated pairs, the reproducibility threat described in Section 3.3 is not mitigated, and the manual annotations cannot be re-audited by other researchers. The authors should provide a working artifact link or state explicitly if the release is pending.

minor comments (6)

[Abstract] There is a typo in the abstract: '15 precent' should be '15 percent'.
[Throughout] The system name is inconsistently rendered as both 'MARAUS' and 'MARUAS' (e.g., Section 1); please use a single spelling.
[Table 1 and Table 2] The tables are formatted as 'T able 1' and 'T able 2' in the text; the spacing should be removed. Also, Table 1 lists 'GPT-4o mini' for this study while the abstract and Section 4.4 mention GPT-4o as well; clarify which model is used where.
[Figure 2] Figure 2 is referenced as showing daily accuracy, but the figure itself is not included in the manuscript text; the data behind the figure should be provided or the figure should be embedded.
[Section 5.2] The 'Perceived user satisfaction' metric is described as a rating given by university staff members, but the paper does not state the rating scale, the exact questions asked, or how the 4.5/5 average was computed; please elaborate.
[Section 4.3] The re-ranking step is called a 'GPT cross-encoder' but cross-encoders are typically a separate model class; clarify whether this is a prompt-based zero-shot reranker using GPT-4o mini or an actual cross-encoder model.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy and hallucination figures are empirical measurements against external UTT ground truth and manual log inspection, not consequences of the system's own definitions.

full rationale

The paper's central claims are empirical measurements rather than derivations from its own premises. The field accuracy of 92% is based on manual inspection of 6,079 question-answer pairs from real deployments (Section 5.1), with an explicit correctness definition in Section 5.2. The hallucination rate of 1.45% is reported from a 100-pair controlled experiment with ground truth provided by UTT (Section 4.5), not from the deployed logs. Although Section 4.4 describes a post-processor that discards and regenerates answers lacking at least one passage citation, the paper does not define 'hallucination' as 'absence of a citation'; the observed reduction is therefore not a definitional identity. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior work, no load-bearing self-citation appears, and no known result is merely renamed. The main limitations are methodological: the annotation rubric for the 6,079 pairs is not documented, the 100-pair set may have influenced design choices reported in the same table, and the promised public code/test-data link is not present in the manuscript. These are correctness, reproducibility, and validity concerns, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The system has no mathematical derivation, but the empirical claims rest on dataset completeness, hand-tuned thresholds, and self-assessment assumptions.

free parameters (6)

semantic similarity filter threshold = 0.9
Hand-set in Section 4.3 to filter FAISS results; no sensitivity analysis, affects retrieved context and thus answer quality.
FAISS top-k = 15
Section 4.3 retrieves top-k=15 FAQ entries; hand-tuned, no ablation.
passages retained after re-ranking = 2
Section 4.3 retains top-2 passages; affects the final prompt and answer grounding.
chunk size / stride = 500 tokens / 100 tokens
Section 4.2 uses overlapping context windows; chosen empirically, no reported ablation.
LLM sampling parameters = temperature 0.7, top-p 0.9, max tokens 350
Section 4.4; no ablation, but influences output variability and hallucination regeneration.
Jaccard deduplication threshold = < 0.9
Section 4.2; near-duplicate filter threshold set without sensitivity analysis.

assumptions (4)

domain assumption The UTT-provided 100 question-answer pairs are correct ground truth for admission policy at deployment time.
Section 4.5 uses these pairs to measure precision, recall, and hallucination; if policy changed mid-period or answers are stale, the lab numbers misstate performance.
domain assumption The coauthors' manual labeling of 6,079 Q&A pairs is an unbiased correctness measure.
Section 5.1 says pairs were manually inspected by coauthors; no independent annotators or rubric, so the 92% accuracy inherits any labeling bias.
domain assumption The two-week deployment window in summer 2025 is representative of real admission workload.
Sections 3.1 and 5.1 base conclusions on this window; daily volume varies from 180 to 797 questions, so representativeness is not established.
domain assumption The university's internal documents contain all policy and score information needed to answer queries.
The system only retrieves from the curated corpus; if documents are incomplete or outdated, accuracy and hallucination rates will be worse than reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling." pith.science (2026). https://pith.science/paper/BX44OVGI

@misc{pith2026250711272,
  author       = {Pith},
  title        = {Pith review of: An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BX44OVGI}},
  note         = {Machine review of arXiv:2507.11272}
}

read the original abstract

This paper presents MARAUS (Multi-Agent and Retrieval-Augmented University Admission System), a real-world deployment of a conversational AI platform for higher education admissions counseling in Vietnam. While large language models (LLMs) offer potential for automating advisory tasks, most existing solutions remain limited to prototypes or synthetic benchmarks. MARAUS addresses this gap by combining hybrid retrieval, multi-agent orchestration, and LLM-based generation into a system tailored for real-world university admissions. In collaboration with the University of Transport Technology (UTT) in Hanoi, we conducted a two-phase study involving technical development and real-world evaluation. MARAUS processed over 6,000 actual user interactions, spanning six categories of queries. Results show substantial improvements over LLM-only baselines: on average 92 percent accuracy, hallucination rates reduced from 15 precent to 1.45 percent, and average response times below 4 seconds. The system operated cost-effectively, with a two-week deployment cost of 11.58 USD using GPT-4o mini. This work provides actionable insights for the deployment of agentic RAG systems in low-resource educational settings.

Figures

Figures reproduced from arXiv: 2507.11272 by the authors.

**Figure 1.** An overview of the MARAUS system 4.2 Pre-processing All textual data undergoes rigorous preprocessing. We remove boilerplate content, HTML artifacts, and near-duplicates using a Jaccard similarity filter (< 0.9). Text normalization includes lowercasing, diacritic normalization, and tokenization via VnCoreNLP. Personally identifiable information (PII) is redacted using regular expression patterns targeting phone nu… view at source ↗

**Figure 2.** Systems’ accuracy Key Observations KO2 – Domain-specialized orchestration that integrates retrieval, reasoning, and personalized interaction is essential for addressing diverse user intents beyond standard FAQ matching in practical deployments 6 Discussion Recent studies have explored various approaches to enhance the accuracy and relevance of AI chatbots using RAG techniques combined with large language models (LL… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages

[1]

Journal of Chemical Education 100(8), 2965–2971 (Aug 2023)

Alasadi, E.A., Baiz, C.R.: Generative AI in Education and Research: Opportunities, Concerns, and Solutions. Journal of Chemical Education 100(8), 2965–2971 (Aug 2023). https://doi.org/10.1021/acs.jchemed.3c00323, https://doi.org/10.1021/acs.jchemed.3c00323, publisher: American Chemical So- ciety

doi:10.1021/acs.jchemed.3c00323 2023
[2]

In: Proceed- ings of the 1st ACM Workshop on AI-Powered Q&A Systems for Mul- timedia

Bui, T., Tran, O., Nguyen, P., Ho, B., Nguyen, L., Bui, T., Quan, T.: Cross-Data Knowledge Graph Construction for LLM-enabled Educa- tional Question-Answering System: A Case Study at HCMUT. In: Proceed- ings of the 1st ACM Workshop on AI-Powered Q&A Systems for Mul- timedia. pp. 36–43. AIQAM ’24, Association for Computing Machinery, New York, NY, USA (Jun...

arXiv 2024
[3]

Chen, Z., Zou, D., Xie, H., Lou, H., Pang, Z.: Facilitating university ad- mission using a chatbot based on large language models with retrieval- augmented generation. Educational Technology & Society27(4), 454–470 (2024), https://www.jstor.org/stable/48791566, publisher: International Forum of Educa- tional Technology & Society, National Taiwan Normal Un...

work page arXiv 2024
[4]

In: Proceedings of the 34th Inter- national Conference on Neural Information Processing Systems

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive NLP tasks. In: Proceedings of the 34th Inter- national Conference on Neural Information Processing Systems. pp. 9459–9474. NIPS ’20, Curran Associates Inc...

work page 2020
[5]

Computers and Education: Artificial Intelligence 8, 100417 (Jun 2025)

Li, Z., Wang, Z., Wang, W., Hung, K., Xie, H., Wang, F.L.: Retrieval-augmented generation for educational application: A sys- tematic survey. Computers and Education: Artificial Intelligence 8, 100417 (Jun 2025). https://doi.org/10.1016/j.caeai.2025.100417, https://www.sciencedirect.com/science/article/pii/S2666920X25000578

arXiv 2025
[6]

https://doi.org/10.48550/arXiv.2405.20139, http://arxiv.org/abs/2405.20139, arXiv:2405.20139 [cs]

Mavromatis, C., Karypis, G.: GNN-RAG: Graph Neural Retrieval for Large Lan- guage Model Reasoning (May 2024). https://doi.org/10.48550/arXiv.2405.20139, http://arxiv.org/abs/2405.20139, arXiv:2405.20139 [cs]

arXiv doi:10.48550/arxiv.2405.20139 2024
[7]

In: 2024 IEEE 20th International Con- ference on Automation Science and Engineering (CASE)

Mishra, P., Mahakali, A., Venkataraman, P.S.: SEARCHD - Advanced Retrieval with Text Generation using Large Language Models and Cross Encoding Re-ranking. In: 2024 IEEE 20th International Con- ference on Automation Science and Engineering (CASE). pp. 975– 980 (Aug 2024). https://doi.org/10.1109/CASE59546.2024.10711642, https://ieeexplore.ieee.org/abstract...

arXiv 2024
[8]

IEEE Access 12, 142733–142759 (2024)

Mittal, U., Sai, S., Chamola, V., Sangwan, D.: A Comprehen- sive Review on Generative AI for Education. IEEE Access 12, 142733–142759 (2024). https://doi.org/10.1109/ACCESS.2024.3468368, https://ieeexplore.ieee.org/abstract/document/10695056

arXiv 2024

Show all 14 references

[9]

https://doi.org/10.48550/arXiv.1908.10084, http://arxiv.org/abs/1908.10084, arXiv:1908.10084 [cs]

Reimers, N., Gurevych, I.: Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (Aug 2019). https://doi.org/10.48550/arXiv.1908.10084, http://arxiv.org/abs/1908.10084, arXiv:1908.10084 [cs]

2019 doi arXiv
[10]

Foundations and Trends ® in Information Re- trieval 3(4), 333–389 (Dec 2009)

Robertson, S., Zaragoza, H.: The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends ® in Information Re- trieval 3(4), 333–389 (Dec 2009). https://doi.org/10.1561/1500000019, https://www.nowpublishers.com/article/Details/INR-019, publisher: Now Pub- lis...

2009 doi
[11]

Empirical Software Engineering14(2), 131 (Dec 2008)

Runeson, P., Höst, M.: Guidelines for conducting and reporting case study research in software engineering. Empirical Software Engineering14(2), 131 (Dec 2008). https://doi.org/10.1007/s10664-008-9102-8, https://doi.org/10.1007/s10664-008- 9102-8, number: 2

2008 doi
[12]

In: 2024 4th International Conference on Pervasive Computing and Social Networking (ICPCSN)

Singla, A.D., Tripathi, S., Victoria, A.H.: HICON AI: Higher Ed- ucation Counseling Bot. In: 2024 4th International Conference on Pervasive Computing and Social Networking (ICPCSN). pp. 779– 784 (May 2024). https://doi.org/10.1109/ICPCSN62568.2024.00131, https://ieeexplore.iee...

2024
[13]

In: Advances in Neural Information Processing Systems

Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: MPNet: Masked and Permuted Pre-training for Language Understanding. In: Advances in Neural Information Processing Systems. vol. 33, pp. 16857–16867. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/hash/c3...

2020
[14]

ECNU Review of Educa- tion 6(3), 355–366 (Aug 2023)

Su, J., Yang, W.: Unlocking the Power of ChatGPT: A Framework for Applying Generative AI in Education. ECNU Review of Educa- tion 6(3), 355–366 (Aug 2023). https://doi.org/10.1177/20965311231168423, https://doi.org/10.1177/20965311231168423, publisher: SAGE Publications Ltd

2023 doi

Pith tools

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

[1] [1]

Journal of Chemical Education 100(8), 2965–2971 (Aug 2023)

Alasadi, E.A., Baiz, C.R.: Generative AI in Education and Research: Opportunities, Concerns, and Solutions. Journal of Chemical Education 100(8), 2965–2971 (Aug 2023). https://doi.org/10.1021/acs.jchemed.3c00323, https://doi.org/10.1021/acs.jchemed.3c00323, publisher: American Chemical So- ciety

doi:10.1021/acs.jchemed.3c00323 2023

[2] [2]

In: Proceed- ings of the 1st ACM Workshop on AI-Powered Q&A Systems for Mul- timedia

Bui, T., Tran, O., Nguyen, P., Ho, B., Nguyen, L., Bui, T., Quan, T.: Cross-Data Knowledge Graph Construction for LLM-enabled Educa- tional Question-Answering System: A Case Study at HCMUT. In: Proceed- ings of the 1st ACM Workshop on AI-Powered Q&A Systems for Mul- timedia. pp. 36–43. AIQAM ’24, Association for Computing Machinery, New York, NY, USA (Jun...

arXiv 2024

[3] [3]

Chen, Z., Zou, D., Xie, H., Lou, H., Pang, Z.: Facilitating university ad- mission using a chatbot based on large language models with retrieval- augmented generation. Educational Technology & Society27(4), 454–470 (2024), https://www.jstor.org/stable/48791566, publisher: International Forum of Educa- tional Technology & Society, National Taiwan Normal Un...

work page arXiv 2024

[4] [4]

In: Proceedings of the 34th Inter- national Conference on Neural Information Processing Systems

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive NLP tasks. In: Proceedings of the 34th Inter- national Conference on Neural Information Processing Systems. pp. 9459–9474. NIPS ’20, Curran Associates Inc...

work page 2020

[5] [5]

Computers and Education: Artificial Intelligence 8, 100417 (Jun 2025)

Li, Z., Wang, Z., Wang, W., Hung, K., Xie, H., Wang, F.L.: Retrieval-augmented generation for educational application: A sys- tematic survey. Computers and Education: Artificial Intelligence 8, 100417 (Jun 2025). https://doi.org/10.1016/j.caeai.2025.100417, https://www.sciencedirect.com/science/article/pii/S2666920X25000578

arXiv 2025

[6] [6]

https://doi.org/10.48550/arXiv.2405.20139, http://arxiv.org/abs/2405.20139, arXiv:2405.20139 [cs]

Mavromatis, C., Karypis, G.: GNN-RAG: Graph Neural Retrieval for Large Lan- guage Model Reasoning (May 2024). https://doi.org/10.48550/arXiv.2405.20139, http://arxiv.org/abs/2405.20139, arXiv:2405.20139 [cs]

arXiv doi:10.48550/arxiv.2405.20139 2024

[7] [7]

In: 2024 IEEE 20th International Con- ference on Automation Science and Engineering (CASE)

Mishra, P., Mahakali, A., Venkataraman, P.S.: SEARCHD - Advanced Retrieval with Text Generation using Large Language Models and Cross Encoding Re-ranking. In: 2024 IEEE 20th International Con- ference on Automation Science and Engineering (CASE). pp. 975– 980 (Aug 2024). https://doi.org/10.1109/CASE59546.2024.10711642, https://ieeexplore.ieee.org/abstract...

arXiv 2024

[8] [8]

IEEE Access 12, 142733–142759 (2024)

Mittal, U., Sai, S., Chamola, V., Sangwan, D.: A Comprehen- sive Review on Generative AI for Education. IEEE Access 12, 142733–142759 (2024). https://doi.org/10.1109/ACCESS.2024.3468368, https://ieeexplore.ieee.org/abstract/document/10695056

arXiv 2024