Source-linked AI summary
Beyond the Chatbot: Co-Learning and Co-Teaching through a Dual-Persona Generative-AI Assistant
Chaido Mizeli, Marina Delianidi, Konstantinos Diamantaras
TL;DR
Educational chatbots need stronger curriculum grounding and role-sensitive support than raw information retrieval provides. This paper introduces S.T.A.R.T. BOT, a Greek educational assistant using textbook-grounded RAG, semantic segmentation, and persona switching, and reports competitive performance from Gemma in a Home Economics pilot. The findings also identify a reliability problem for DeepSeek in the tested assessment tasks.
Problem
Educational chatbots often lack the contextual grounding needed to remain faithful to specific national curricula, while the paper asks whether SLMs can compete with LLMs in teaching applications.
Method
S.T.A.R.T. BOT combines RAG over official textbooks, metadata-enriched semantic segmentation, and persona-based prompting for teacher and student roles.
Results
Gemma achieved a mean score of 0.753, close to Gemini (0.773) and DeepSeek (0.764), while DeepSeek generated no output for the requested assessment tests.
Takeaways & Limitations
The pilot indicates that SLMs can serve as instructional assistants and can rival or exceed larger models in specific creative teaching tasks.
Abstract
from arXiv · showhide
In this paper we present a generative AI application developed to support both teachers and students in secondary education. The system employs two Large Language Models-LLMs, Gemini and DeepSeek, and a Small Language Model-SLM, Gemma, integrated within a Retrieval Augmented Generation - RAG framework, creating a pedagogically grounded, Greek-language assistant capable of adapting its reasoning and communication style to the user role. Unlike conventional chatbots, the assistant introduces pedagogical persona switching, a dual-role mechanism that enables the same AI model to act as both a teaching companion and a learning guide. Utilizing a RAG paradigm tailored to the Greek educational domain, the architecture segments official textbooks into coherent units. Enriched with specific metadata, these units preserve curricular structure and instructional context, demonstrating how generative AI optimizes modern instructional design. The initial case study focuses on home economics in Greek lower secondary education, a cross-disciplinary subject that integrates elements of economics, health education, and social responsibility. The assistant has been developed to support both learners and educators in complementary ways. In future classroom implementations, students will be able to use it to clarify key concepts such as financial literacy, resource management, and healthy living, while teachers could employ it to design authentic instructional materials, formative assessments, and classroom activities aligned with the official curriculum. The study elevates the concept beyond a simple chatbot, proposing a structured, contextually adaptive framework for pedagogical generative assistants that effectively bridge technology, curriculum, and human learning.
1. Introduction
The paper addresses the difficulty of grounding educational AI in national curricula by introducing S.T.A.R.T. BOT, a Greek-language assistant with RAG and pedagogical persona switching. Its pilot study examines whether a Small Language Model can provide curriculum-aligned support alongside larger models.
- RAG retrieves external factual information before generation, helping reduce hallucinations and ground answers in verified data.The paper connects this function to the need for curriculum-faithful educational responses.
- S.T.A.R.T. BOT uses RAG to ground responses in official school textbooks segmented into metadata-enriched units.The architecture preserves curriculum structure and teaching context during retrieval and generation.
- Pedagogical Persona Switching lets one AI act as an educator’s assistant or a student’s learning guide.Teacher support includes assessment items and lesson plans, while student support includes curriculum-based explanations and feedback.
- The pilot focuses on Home Economics in Greek lower-secondary education and evaluates Small Language Models against Large Language Models.The subject integrates economics, health education, and social responsibility.
- The framework combines localized curriculum-specific RAG, semantic educational text segmentation, and adaptive persona-driven prompting.The paper presents this combination as a reproducible foundation for AI integration in local educational contexts.
3. Methodology
The methodology is organized into three phases: data preprocessing, search-index creation, and interaction-mechanism design.
- The system methodology comprises data preprocessing, search-index creation, and interaction-mechanism design.
A. Data Pre-processing and Semantic Chunking
Official student textbooks are divided into focused conceptual units and enriched with metadata to preserve the curriculum’s organizational structure.
- Official student textbooks provide the system’s primary data source.
- Text is divided into conceptual units organized by chapter.
- Each section is limited to 350 words and enriched with module, subsection, and document-type metadata.These constraints keep retrieved information focused while maintaining curriculum structure.
B. Semantic Indexing and Vector Storage
The system converts Greek educational text into semantic vectors for retrieval and provides role-specific interaction through a Streamlit interface supporting both large and small language models.
- A Greek-language sentence encoder converts textbook text into embeddings for semantic searching.The encoder is described as specially trained for Greek.
- Generated vectors are stored in a FAISS IndexFlatL2 index for fast Euclidean-distance search.The retrieval process matches user questions to relevant book passages by meaning rather than keywords.
- The Streamlit application lets users select Student or Teacher roles through an interactive web interface.
- Specialized prompts combine the selected role, behavioral instructions, retrieved RAG data, and response rules.Student Mode provides syllabus-based explanations and feedback, while Tutor Mode creates educational materials such as tests and experiments.
- The evaluation supports Gemini-1.5-Flash-Lite and DeepSeek as LLMs, alongside Gemma 3:4b as an SLM.
4. Experiments
The experiments evaluated S.T.A.R.T. BOT in Home Economics for Greek lower-secondary education across two pedagogical modes and three language models within a RAG framework. Student Mode focused on curriculum-grounded explanations using official educational materials.
- The case study used Home Economics for Greek lower-secondary education because of its cross-disciplinary focus on practical life skills.
- Gemini, DeepSeek, and the resource-efficient SLM Gemma were evaluated to assess how model scale affects educational outcomes.
- All models operated within a RAG framework using a Greek semantic encoder and FAISS for vector retrieval.
- A unified prompt combined role instructions, retrieved context, rules, and the user question to maintain each model’s assigned pedagogical persona.
- Student Mode Experiments: Student Mode assessed curriculum-grounded explanations through ten curriculum-based questions compared with official teacher-guide answers using semantic similarity.
Tutor Mode Experiments
Tutor Mode evaluated the assistant as an instructional partner generating structured assessment tests from Home Economics curriculum chapters. The evaluation emphasized coverage, question quality, and relevance to the educational material.
- Tutor Mode focused on generating complex pedagogical artifacts as an instructional partner.
- Each model generated fully structured assessment tests for two Home Economics chapters, including Family Budget and Modern Technology and Food.
- The tests were evaluated for instructional-unit coverage, question formulation quality for junior high school students, and relevance to the provided material.
5. Evaluation and Results
Student Mode evaluation measured how accurately and consistently generated responses matched the official Home Economics curriculum. The study used teacher-guide answers as ground truth and semantic similarity across three models.
- Official teacher-guide answers served as ground truth for evaluating generated responses against the curriculum.
- Semantic Similarity metrics were calculated from ten curriculum-based questions submitted to Gemini, DeepSeek, and Gemma.
Key Findings
Gemma showed competitive Student Mode performance relative to Gemini and DeepSeek, while qualitative analysis found high fidelity to official content. Tutor Mode used an LLM judge to compare generated assessment tests across pedagogical criteria.
- Student Mode: Gemma achieved a mean semantic similarity score of 0.753, close to Gemini at 0.773 and DeepSeek at 0.764.
- Student Mode: Gemma outperformed the larger models on some questions, reaching similarity scores as high as 0.914.
- Student Mode: A qualitative example showed Gemma covering purchasing stages and providing precise citations to source textbooks.
- Tutor Mode: Tutor Mode assessment tests from Gemini, DeepSeek, and Gemma were evaluated by GPT-5.1 pro using an LLM-as-a-Judge framework.
- Tutor Mode: The judge assessed coverage of instructional-unit concepts, relevance to RAG context, artifact format, and question formulation.
- Tutor Mode: The standardized judging procedure produced qualitative feedback and quantitative scores for comparing models as instructional partners.
Performance Comparison and Model Reliability
Gemini and Gemma were compared on curriculum-based assessment generation, while DeepSeek failed to produce outputs for either selected chapter. The active models showed generally strong but differentiated linguistic and curricular alignment.
- Model Reliability: DeepSeek generated no assessment-test output for either selected chapter, so comparisons focused on Gemini and Gemma.The authors identify this as a potential reliability limitation in the Greek educational, persona-driven setting.
- Question Quality: Gemini questions were generally well-written, clear, and linguistically appropriate for junior-high students.The reported evaluation places Gemini’s quality at 65–75%, with one item requiring stronger connection to the source passage.
- Curricular Relevance: Gemini’s questions were mostly relevant to the educational text, with minor leakage into another subsection and one more general question.The cited evaluation identifies Γ4 and Α3 as the main deviations.
- Question Quality: Gemma questions were also clear and student-friendly, but two or three items needed closer connection to textbook phrasing.Its reported quality range was 60–70%.
- Curricular Relevance: Gemma’s questions were generally aligned, although some relied more on logical or experiential knowledge than on explicit textbook content.The main examples identified are questions Α2 and Α4.
Assessment Test 1: Family Budget
In the two assessment tests, Gemini led on the Family Budget unit while Gemma performed better on technical food-preservation content. Together, the findings indicate a trade-off between academic alignment and creative, accessible pedagogy.
- Assessment Test 1: Family Budget: Gemini achieved 65%–75% coverage in Family Budget, exceeding Gemma’s 60%–70% coverage and producing clear, age-appropriate questions.Gemini showed stronger coverage of budgeting, income, and planning sub-units.
- Assessment Test 1: Family Budget: Gemini was the top-performing AI model for Family Budget because of structured academic phrasing and minimal curriculum deviations.The textbook-derived BookTest1 remained the 100% compatibility gold standard.
- Assessment Test 2: Modern Technology and Food: Gemma achieved 60%–70% coverage in the technical food-preservation test, surpassing Gemini’s 50%–60% coverage.Gemma was praised for simple language, structured experiments, and translating theory into practical activities.
- Summary of Findings: The Tutor Mode results describe Gemini as more academic and strictly aligned for economic concepts, while Gemma showed greater creativity and a friendlier tone.The contrast was especially visible in classroom-experiment design.
6. Conclusion
The conclusion presents S.T.A.R.T. BOT as a RAG-based, dual-persona educational assistant whose small language model can compete with larger models in specialized tasks. It further emphasizes local deployment, privacy, and lower computational demands as practical advantages for educational integration.
- Conclusion: S.T.A.R.T. BOT combines RAG with Pedagogical Persona Switching so one AI model can serve students and teachers in different roles.The conclusion reports high curriculum alignment for Gemma, often matching or exceeding larger models in creativity and linguistic clarity.
- Conclusion: Gemma can compete with Gemini and DeepSeek in specialized educational tasks, including Student and Tutor modes.The reported strengths include semantic curriculum alignment, pedagogical creativity, and linguistic clarity.
- Conclusion: SLMs are suited to closed educational environments because they can run on-premise or on local school infrastructure.This keeps sensitive educational data and student interactions local and aligns with privacy and GDPR requirements.
- Conclusion: SLMs require fewer computational resources than LLMs, supporting their proposed use as sustainable and cost-effective systems for national education deployment.The conclusion links this lower hardware requirement to settings with constrained hardware budgets.
- Conclusion: The framework is characterized as reliable, secure, pedagogically grounded, private, localized, and technologically accessible.These properties are presented as the foundation for future AI assistants supporting teachers and students.