International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
A Survey of Retrieval-Augmented Generation Techniques for Intelligent Educational Assistants
Lakshya Singh1 , Er. Alok Singh Jadaun2 , Prof. Brajesh Kumar Singh3
1Department of Computer Science
and
Engineering
Raja Balwant Singh Engineering Technical Campus
Bichpuri, Agra, Uttar Pradesh, India
2 Department of Computer Science and Engineering
Raja Balwant Singh Engineering Technical Campus
Bichpuri, Agra, Uttar Pradesh, India
3 Department of Computer Science and Engineering
Raja Balwant Singh Engineering Technical Campus
Bichpuri, Agra, Uttar Pradesh, India
Abstract - Retrieval-Augmented Generation (RAG) has emerged as a transformative approach for improving the factual reliability and contextual groundingoflargelanguage models, particularly in knowledge-intensive domains such as education. Standalone language model-based educational assistants frequently suffer from hallucinated responses, insufficient curriculum alignment, and limited explainability, which collectively impede their deployment informallearning environments. This paper presents a systematic review of recent advances in Retrieval-Augmented Generation techniques, with a focused lens on intelligent educational assistants Core RAG architectures including naive, advanced, and modular designs are examined alongside their educational applicationsspanningtutoring, examination support, and academic assistance. A structured comparative analysis of representative studies highlights critical design trade-offs in retrieval strategies, language model selection, and evaluation methodologies. The paper further delineates key open challenges related to evaluation standardization, knowledge base maintenance, scalability, pedagogicalquality, and responsible deployment. Future research directions are outlined to advance the development of reliable, scalable, and learner-centric RAG-based educational systems.
Key Words: Retrieval-Augmented Generation, Large Language Models, Intelligent Educational Assistants, Artificial Intelligence in Education, Knowledge Grounding, Educational Chatbots, Information Retrieval
1.INTRODUCTION
The rapid advancement of Large Language Models (LLMs)hasfundamentallyreshapednaturallanguage processing, enabling sophisticated capabilities in question answering, text summarization, and conversationalinteraction[2].Withineducation,LLMpoweredassistantshaveattractedgrowinginterestfor theirpotentialtodeliverpersonalizedtutoring,instant
feedback,andscalableacademicsupport.Despitethese capabilities, standalone LLMs exhibit critical limitationsincludinghallucinatedresponses,reliance on outdated training data, and absence of domain grounding allofwhichposeacuterisksinhigh-stakes educationalsettings.
Factual accuracy and curriculum alignment are nonnegotiablerequirementsforeducationalsystems. Unlike general-purpose chatbots, educational assistants must deliver precise, explainable, and context-aware answers grounded in verified instructionalmaterialssuchastextbooksandlecture notes. Since conventional LLMs generate responses from probabilistic patterns alone without runtime access to external knowledge they frequently produceconfidentyetincorrectanswers[7],[18].
Toovercometheselimitations,Retrieval-Augmented Generation(RAG) has emerged as a principled frameworkthatcouplesthegenerativefluencyofLLMs with dynamic retrieval from external knowledge sources. By conditioning response generation on retrieveddocuments suchascourseslides,research articles,andcuratedcurricula RAG-basedsystems achieve substantially improved factual accuracy, transparency, and domain relevance [2], [17]. This paper provides a comprehensive review of RAG techniques with a specific focus on intelligent educational assistants, examining core architectures, retrieval strategies, educational applications, comparative design choices, open challenges, and futuredirections.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
2.BACKGROUND ANDMOTIVATION
Educationalquestion-answeringsystemshaveevolved through three broad phases: rule-based retrieval, statistical machine learning, and deep transformerbased models. Early assistants relied on keyword matching and static knowledge bases, offering high precision within constrained domains but lacking conversationalflexibilityandadaptability[4],[7].
The arrival of large-scale transformer models introduced substantially richer interaction patterns, yet their direct deployment in education introduced newproblems.LLMsgenerateresponsesfromlearned statisticalassociationswithoutverifyingclaimsagainst authoritative sources, producing hallucinated or partiallycorrectanswersthatcanmisleadlearnersand undermine trust [5], [13]. Furthermore, domainspecificcurricula,learningobjectives,andinstructional sequencingconstraintsarenotinherentlyencodedin general-purposelanguagemodels.
RAGdirectlyaddressestheselimitationsbyenabling dynamic, inference-time retrieval from curated knowledge bases, thereby grounding generation in verifiable and curriculum aligned content. This combination preserves the fluency of LLMs while adding factual accountability, making RAG a foundationaltechnologyfortrustworthyeducationalAI [2],[17],[21].
3.RETRIEVAL-AUGMENTED GENERATION(RAG)
Retrieval-Augmented Generation is a hybrid framework that enhances factual accuracy by integrating external knowledge retrieval into the language model generation pipeline. Rather than dependingsolelyonparametricknowledgeacquired during pre-training, RAG systems retrieve relevant documents from an external knowledge base at inference time and condition the generator on this retrievedcontext[2],[18].
A canonical RAG architecture comprises two core components. The retriever identifies and fetches the most semantically relevant passages from indexed sources such as lecture notes, textbooks, or course materials usingdensevectorembeddingsstoredina vector database. Upon receiving a user query, the system encodes it into an embedding, performs similaritysearch,andretrievesthetop-krelevanttext
chunks.Thegenerator,typicallyatransformer-based LLM, then produces a response conditioned on both theoriginalqueryandtheretrievedcontext,yielding outputsthatarelinguisticallycoherentandgrounded intraceablesources[1],[3].
RAG approaches span three levels of architectural sophistication. Naive RAG follows a straightforward pipeline:retrievedocumentsandappendthemdirectly to the prompt. Advanced RAG incorporates query rewriting, re-ranking, and multi-hop retrieval to improve contextual precision. Modular RAG extends this further by enabling flexible composition of specialized retrievers, memory modules, re-rankers, andgenerators,allowingthesystemtobecustomized forspecifictasksordomains[8],[18].Fig.1illustrates thegeneralRAGworkflow.
User Query
Model
Vector Database (Top-k Retrieval)
Query + Retrieved Context
LLM Generator
Grounded, Accurate Response
Fig -1: GeneralworkflowofaRAG-basedintelligent educationalassistant.
4.RAGINEDUCATIONAL ASSISTANTS
The integration of RAG into intelligent educational assistantshasgainedconsiderablemomentumdueto its capacity to fulfill the accuracy and grounding requirementsoflearningenvironments.Byleveraging curatedacademicresources lectureslides,textbooks, research articles, and course-specific content RAGpowered systems deliver context-aware and curriculum-aligned responses across diverse instructionalscenariosincludingpersonalizedtutoring, conceptclarification,doubtresolution,andautomated academicsupport[12],[17],[21].
Embedding
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
Recent empirical work demonstrates RAG effectivenessacrossmultipleeducationaldomains.In intelligenttutoringsystems,RAGenablesstep-by-step explanations by retrieving relevant instructional content prior to response generation [16], [19]. For examination preparation and formative assessment, RAG-based models generate practice questions and constructive feedback directly grounded in course material[9],[11].Virtualteachingassistantsinhigher education employ RAG to handle student queries spanning lectures, assignments, and administrative information[13],[22].
A particularly notable advantage of RAG is that it enablescompact,locallydeployedlanguagemodelsto achieve performance levels comparable to large proprietary systems substantially reducing computationalcostwhilepreservingdataprivacy[1], [11].ThisscalabilityadvantagemakesRAGespecially attractive for resource-constrained educational institutions. Nevertheless, effectiveness depends critically on knowledge base quality, coverage, and organization; poorly curated content can degrade retrieval performance and generation quality [10], [14].
From a student experience perspective, RAG-based systems offer a measurable improvement in transparency and trust. Because responses can be tracedbacktospecificretrieveddocuments,learners canverifythe source of information afeature that distinguishesRAGfromopaquestandaloneLLMsand alignswithacademicintegrityexpectationsinformal education. Studies such as Salminen et al. [23] and Alsafarietal.[21]highlightthatstudentsdemonstrate higher satisfaction and trust when AI responses are linkedtoverifiableinstructionalmaterials,suggesting thatsourceattributionisnotmerelyatechnicalfeature butapedagogicallymeaningfuldesignchoice.
5.LITERATURESEARCHMETHODOLOGY
To ensure systematic coverage of the relevant literature, this review followed a structured search methodology consistent with established systematic reviewpractices.AcademicdatabasesincludingGoogle Scholar,IEEEXplore,andScopuswerequeriedusing combinationsof the followingkeywords:“RetrievalAugmented Generation,” “RAG,” “Large Language Models in Education,” “Intelligent Educational Assistants,” “Educational Chatbots,” “LLM tutoring
systems,” and “knowledge-grounded dialogue systems.”
Thesearchwasboundedtopublicationsfrom2022to 2025 to capture recent developments in this rapidly evolving field. An initial corpus of approximately 80 paperswasidentified.Afterremovingduplicatesand applyinginclusioncriteria namely,papersexplicitly addressing RAG architectures or LLM-based educational applications with empirical or architecturalcontributions 25paperswereretained fordetailedanalysis,ofwhich23aredirectlycitedin this review. Papers were excluded if they addressed LLMs without educational context, lacked sufficient architecturalorevaluativedetail,orappearedinnonpeer-reviewed venues without substantial technical contribution. This process is consistent with establishedsystematicliteraturereviewpractices[4].
The retained studies span a range of educational domains including computer science education, statistics, healthcare, physics, and general university support. The distribution of publication venues includes ACM conferences (SIGCSE, IUI, CHI, L@S), Elsevierjournals(Computers&Education,Procedia ComputerScience),andSpringerjournals(Advanced IntelligentSystems, Information), ensuring a diverse and credible evidence base for the comparative analysispresentedinthisreview.
6. COMPARATIVE ANALYSIS OF EXISTING APPROACHES
Existing RAG-based educational systems reveal significantvariationacrossretrievalstrategy,language model selection, domain focus, and evaluation methodology. Early systems paired simple dense retrievalwithlargeproprietarymodels,whilerecent work increasingly prioritizes modularity, cost efficiency, and domain adaptability [2], [18]. Table I provides a structured comparison of representative studiesacrossthesedimensions.
Retrieval strategies range from traditional dense vectorsimilaritysearchtomoresophisticatedhybrid and graph-based retrieval. With respect to language model configurations, a clear trend emerges toward locallydeployedsmallermodels,motivatedbyprivacy constraints and cost reduction [1], [11]. Evaluation
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
TABLE I
Comparative Analysis of RAG-Based Educational Assistants
Study Domain Retrieval Strategy LLM Configuration Evaluation Method
Gaoetal.[18] GeneralNLP DenseVectorRetrieval GPT(Proprietary) ROUGE,Human Evaluation
Lietal.[17] EducationApps HybridRetrieval MultipleLLMs Accuracy,Relevance
Mishraetal.[1] UniversitySystems Dense+LocalRetrieval LocalLLM(Ollama) Accuracy,Latency
Yigcietal.[13] HigherEducation DenseRetrieval GPT-4 UserStudy
Yuetal.[11] CSEducation DenseRetrieval SmallLM(Local) StudentFeedback
Nemethetal.[22] Statistics RAGPipeline GPT-basedTA ExpertAssessment
Khanetal.[12] UniversityVA DenseRetrieval LLM+RAGFramework ResponseQuality
Wangetal.[6] HealthcareEdu. RAG-EnhancedPipeline GPT-4 ClinicalAccuracy
Tyndalletal.[9] ExamSupport RAG+AIAgents MultipleLLMs ExamPerformance
Cao[19] CS1Tutoring Retrieval+ITS LLMTutor StudentOutcomes
methodologies remain inconsistent, spanning automated NLP metrics (BLEU, ROUGE, cosine similarity),humanexpertjudgment,anduserstudies reflecting a lack of standardized pedagogical benchmarks[10].
Several additional observations emerge from the comparative analysis. First, domain-specific systems consistently outperform general-purpose configurations when the knowledge base is wellcurated and aligned with instructional objectives underscoringtheimportanceofknowledgeengineering in educational RAG design. Second, systems that incorporate re-ranking mechanisms report notably improvedretrievalprecisioncomparedtonaivetop-k approaches,suggestingthatpost-retrievalfilteringisa worthwhileinvestmentevenataddedcomputational cost[8],[18].Third,evaluation gapsareparticularlypronouncedinstudiestargeting K-12 education, where learner age, cognitive development stage, and curriculum constraints introduce evaluation dimensions absent from higher educationsettings[4],[14].
7.RESEARCHGAPSAND CHALLENGES
Despitenotableprogress,severalcriticalgapsconstrain thematurityofRAG-basededucationalsystems.First, thereisastrikingabsenceofstandardizedevaluation frameworks tailored to educational contexts. Most studiesemploygenericNLPmetricsthatfailtocapture pedagogicaleffectiveness,learninggains,orlong-term engagement[10].Withoutunifiedbenchmarks,crossstudycomparisonremainsunreliable.Metricssuchas BLEUandROUGEmeasuresurface-leveltextsimilarity but cannot assess whether a response genuinely supports conceptual understanding, motivates the learner,oralignswiththeinstructionalgoalofagiven lesson.
Second,knowledgebaseconstructionandmaintenance presentpersistentchallenges.Static,manuallycurated repositories do not scale across evolving curricula, diverse subjects, or multiple institutions [12], [22]. Inconsistentcontentstructuring,missingpedagogical metadata, and inadequate alignment with learning objectivesfurtherdegraderetrievalquality.Automated
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
knowledge base construction from raw course materials includingslides,transcripts,andproblem sets remains an open research problem requiring advances in both document parsing and semantic indexing.
Third,scalabilityremainsanopenconcern.Retrieval latency over large corpora can disrupt real-time classroom interactions [9], while context window limitationsrestricthowmuchretrievedcontentcanbe effectively utilized. Balancing retrieval depth with response speed is particularly challenging for institutionswithlimitedcomputationalinfrastructure [1], [11]. On-device or edge deployment of RAG systems, which would benefit privacy-sensitive educational environments, remains technically demandingandlargelyunexplored.
Fourth, pedagogical and ethical dimensions remain underexplored. Many systems optimize for factual accuracy while neglecting instructional quality specifically, clarity of explanation, adaptability to learner proficiency, and support for critical thinking [14],[15].Issuesofbiasintrainingdata,overreliance onautomatedfeedback,andtransparencyinretrieval decisions further underscore the need for human-inthe-loop RAG designs [5], [14]. The risk of learners uncritically accepting AI-generated content without developing independent reasoning skills is a particularly pressing concern that current RAG architecturesdonotadequatelyaddress.
8.FUTURERESEARCHDIRECTIONS
Future research should advance RAG along three converging fronts. Technically, hybrid dense-sparse retrieval,graph-basedknowledgerepresentations[8], and multi-hop retrieval hold promise for improving contextualprecisionandhandlingcomplexeducational queries. Graph-based retrieval, in particular, is wellsuited for educational knowledge domains where conceptsareinherentlyinterconnected forexample, understanding calculus presupposes knowledge of algebraandlimits,relationshipsthatagraphstructure can explicitly encode and exploit during retrieval. Domain-adapted embedding models trained on educational corpora can better align retrieval with curricularcontent[17].
Pedagogically,futureRAGsystemsmustembedlearner centered design by adapting retrieved content and explanation style to individual learner profiles, prior knowledge levels, and learning objectives [15], [16]. Frameworks that incorporate scaffolding, formative feedback,andadaptivedifficultyhavestrongpotential to enhance learning outcomes [19], [20]. Human-inthe-loopmechanisms allowingeducatorstoguideand validate system outputs can further balance automationwithinstructionalaccountability[21],[23]. Theintegrationoflearnermodelingtechniques,which dynamically track student knowledge state and misconceptions,intoRAGretrievalpipelinesrepresents a particularly promising direction for personalized educationalAI.
Evaluatively, the field urgently needs standardized benchmarks that incorporate pedagogical indicators, learnerperformanceoutcomes,andlongitudinaluser studies,movingbeyondconventionalNLPmetrics[4], [10].Explainabilityfeaturesthatlinkretrievedsources to generated responses will be essential for building trustwithbothlearnersandeducators[5].Responsible deploymentprinciples includingbiasmitigation,data governance,andethicaloversight mustbeintegrated into system design from the outset [5], [14]. Collaborative efforts between AI researchers, educational technologists, curriculum designers, and ethicists will be necessaryto translate technical RAG advances into pedagogically sound, equitable, and sustainableeducationaltools.
9.CONCLUSION
ThispaperpresentedasystematicreviewofRetrievalAugmented Generation techniques for intelligent educationalassistants,synthesizingevidencefrom23 recentstudiesspanningtheperiod2022to2025.The review demonstrated that RAG effectively addresses core limitations of standalone LLMs particularly hallucination,domainmisalignment,andlackofsource transparency makingitafoundationalparadigmfor trustworthyeducationalAI.Corearchitectures(naive, advanced,andmodularRAG),educationalapplications across tutoring, examination support, and virtual assistance, and comparative design trade-offs were
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
examined alongside a structured literature search methodology.
Open challenges in evaluation standardization, knowledge base management, scalability, and pedagogical quality were identified, with corresponding future research directions outlined spanning technical, pedagogical, and evaluative dimensions.Alimitationofthisreviewisitsfocuson English language publications from 2022 to 2025; multilingualandearlierfoundationalworkmayoffer additionalinsightsnotcapturedhere.Furthermore,the rapidly evolving nature of both LLM capabilities and RAG architectures means that findings may require periodicrevisionasthefieldadvances.
Overall, RAG represents a promising and rapidly maturing technology for building reliable, learnercentric educational assistants. Its ability to ground responses in verifiable sources, adapt to domainspecific knowledge bases, and enable cost effective local deployment positions it as a particularly viable solution for diverse educational institutions worldwide provided that technical advances are complemented by pedagogically informed design, rigorous evaluation, and responsible deployment practices.
REFERENCES
[1] A. Mishra and N. Brahmanapally, "A Comparative Performance Analysis of Locally Deployed Large Language Models Through a Retrieval-Augmented GenerationEducationalAssistantApplicationforTextual DataExtraction,"AI,vol.6,no.6,p.119,Jun.2025,doi: 10.3390/ai6060119.
[2]W.Fanetal.,"ASurveyonRAGMeetingLLMs:Towards Retrieval-AugmentedLargeLanguageModels,"inProc. 30thACMSIGKDD,Barcelona,Spain:ACM,Aug.2024,pp. 6491–6501,doi:10.1145/3637528.3671470.
[3] H. Li, Y. Su, D. Cai, Y. Wang, and L. Liu, "A Survey on Retrieval-AugmentedTextGeneration,"arXiv preprint arXiv:2202.01110,2022.
[4]V.Liu,E.Latif,andX.Zhai,"AdvancingEducationthrough Tutoring Systems: A Systematic Literature Review," arXiv:2503.09748, Mar. 2025, doi: 10.48550/arXiv.2503.09748.
[5]D.-M.Cordova-Esparza,"AI-PoweredEducationalAgents: Opportunities, Innovations, and Ethical Challenges," Information, vol. 16, no. 6, p. 469, May 2025, doi: 10.3390/info16060469.
[6]D.Wangetal.,"EnhancementofthePerformanceofLarge Language Models in Diabetes Education through Retrieval-AugmentedGeneration:ComparativeStudy,"J Med Internet Res, vol. 26, p. e58041, Nov. 2024, doi: 10.2196/58041.
[7] B. D. Nye, D. Mee, and M. G. Core, "Generative Large Language Models for Dialog-Based Tutoring: An Early ConsiderationofOpportunitiesandConcerns,"inProc. AIEDWorkshoponEmpoweringEducationwithLLMs, CEUR-WS,vol.3487,pp.78–88,2023.
[8]B.Pengetal.,"GraphRetrieval-AugmentedGeneration:A Survey," ACM Trans. Inf. Syst., vol. 44, no. 2, pp. 1–52, Feb.2026,doi:10.1145/3777378.
[9]E.Tyndall,C.Gayheart,A.Some,J.Genz,T.Wagner,andB. Langhals,"Impactofretrievalaugmentedgenerationand large language model complexity on undergraduate examscreatedand taken by AIagents,"Data &Policy, vol.7,p.e57,2025,doi:10.1017/dap.2025.10024.
[10] Z. F. Han et al., "Improving Assessment of Tutoring Practices using Retrieval-Augmented Generation," in Proc.AAAIWorkshoponAIforEducation,2024.
[11] Z. Yu, S. Liu, P. Denny, A. Bergen, and M. Liut, "Integrating Small Language Models with RetrievalAugmented Generation in Computing Education: Key Takeaways,Setup,andPracticalInsights,"inProc.56th ACMSIGCSE,Pittsburgh,PA:ACM,Feb.2025,pp.1302–1308,doi:10.1145/3641554.3701844.
[12] U. H. Khan, M. H. Khan, and R. Ali, "Large Language Model based Educational Virtual Assistant using RAG Framework,"ProcediaComputerScience,vol.252,pp. 905–911,2025,doi:10.1016/j.procs.2025.01.051.
[13] D. Yigci, M. Eryilmaz, A. K. Yetisen, S. Tasoglu, and A. Ozcan,"LargeLanguageModel-BasedChatbotsinHigher Education,"AdvancedIntelligentSystems,vol.7,no.3,p. 2400429,Mar.2025,doi:10.1002/aisy.202400429.
[14] D. Hooshyar et al., "Problems With Large Language Models for Learner Modelling: Why LLMs Alone Fall Short for Responsible Tutoring in K-12 Education," arXiv:2512.23036, Dec. 2025, doi: 10.48550/arXiv.2512.23036.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
[15]Y.Zhang,E.L.Ouh,A.Ho,S.L.Lo,K.W.Tan,andF.Lin, "PromptTutor: Effects of an LLM-Based Chatbot on Learning Outcomes and Motivation in Flipped Classrooms," in Proc. 30th ACM ITiCSE, Nijmegen, Netherlands: ACM, Jun. 2025, pp. 445–451, doi: 10.1145/3724363.3729095.
[16]C.Xia,"ResearchontheApplicationofLargeLanguage ModelsinIntelligentTutoringSystem,"inProc.ICIAAI 2025,vol.122,Dordrecht:AtlantisPress,2025,pp.993–1003,doi:10.2991/978-94-6463-823-3_97.
[17]Z.Li,Z.Wang,W.Wang,K.Hung,H.Xie,andF.L.Wang, "Retrieval-augmented generation for educational application: A systematic survey," Computers and Education:ArtificialIntelligence,vol.8,p.100417,Jun. 2025,doi:10.1016/j.caeai.2025.100417.
[18]Y.Gaoetal.,"Retrieval-AugmentedGenerationforLarge Language Models: A Survey," arXiv preprint arXiv:2312.10997,2023.
[19]C.Cao,"ScaffoldingCS1CourseswithaLargeLanguage Model-Powered Intelligent Tutoring System," in Proc. 28th ACM IUI, Sydney, Australia: ACM, Mar. 2023, pp. 229–232,doi:10.1145/3581754.3584111.
[20]A.LiebandT.Goel,"StudentInteractionwithNewtBot: An LLM-as-tutor Chatbot for Secondary Physics Education," in Ext. Abstracts CHI 2024, Honolulu, HI: ACM, May 2024, pp. 1–8, doi: 10.1145/3613905.3647957.
[21] B. Alsafari, E. Atwell, A. Walker, and M. Callaghan, "Towards effective teaching assistants: From intentbased chatbots to LLM-powered teaching assistants," NaturalLanguageProcessingJournal,vol.8,p.100101, Sep.2024,doi:10.1016/j.nlp.2024.100101.
[22]R.Németh,A.Tátrai,M.Szabó,andÁ.Tamási,"Usinga RAG-enhanced large language model in a virtual teachingassistantrole:Experiencesfromapilotproject in statistics education," Hungarian Statistical Review, vol. 7, no. 2, pp. 3–27, 2024, doi: 10.35618/hsr2024.02.en003.
[23] J. Salminen et al., "Using Cipherbot: An Exploratory Analysis of Student Interaction with an LLM-Based EducationalChatbot,"inProc.ACML@S2024,Atlanta, GA: ACM, Jul. 2024, pp. 279–283, doi: 10.1145/3657604.3664690.