About CHETI
An innovative AI learning platform from the Chemistry Didactics department at RPTU Kaiserslautern-Landau (Campus: Kaiserslautern).
Our Vision
CHETI combines modern AI technology with subject-didactic expertise. Our goal is to support students with the help of an LLM-based learning tutor and make learning chemistry content more engaging and effective. The AI is adapted to promote self-efficacy in chemistry. At the same time, CHETI is a research platform to further investigate the interaction between LLMs and learners.
AI-Powered Tutors
Through the use of LLMs (Large Language Models) like GPT-4 and Claude, we simulate interactive tutoring systems. These enable adaptive feedback and personalized assistance for complex chemistry questions.
Research Project
This project is part of the research at the Chemistry Didactics department at RPTU in Kaiserslautern. We investigate how AI tools can be effectively and safely integrated into educational scenarios to deepen chemical understanding and support teachers.
AI Chat Systems: Technical Architecture
How the two chat types of CHETI work — explained in an understandable way.
The Two Chat Types at a Glance
Both systems use the same technical infrastructure but differ fundamentally in context and knowledge access.
Learning Material Chat
Topic-based Tutor
The student is in a specific learning module (e.g., "Acids and Bases"). The AI knows the learning content, the tasks, and the student's previous answers. It responds with context and provides targeted support.
Free Learning Chat
Untethered Tutor
The student opens a standalone chat page, detached from any specific learning module. They can ask any chemistry questions — e.g., to prepare for an exam or out of interest in a topic.
How AI Responses Are Generated
The steps a student query goes through from typing to response.
Learning Material Chat
8 StepsFree Learning Chat
5 Steps1. Student Message
Security check: no personal data, no injection attempts.
1. Student Message
Identical security check. Images can also be uploaded.
2. Assign Session
One chat session per student and lesson. Messages are counted (quota).
2. New Session
Each chat creates its own session. Sidebar for previous chats.
3. Classify Intent
Automatic detection: task, explanation, small talk, or off-topic.
3. Classify Intent
Same classification, but "task" is never detected (no context).
4. Search Learning Material (RAG)
The question is compared with the learning material. The 5 most relevant text passages are selected (vector + full-text search).
No Access to Learning Material
5. Compile Context
Tutor rules, learning passages, tasks + answers, chat history.
4. Compile Context
Simplified tutor prompt. Chat history as the only context.
6. Generate Response
AI model generates response as a live stream (token by token).
5. Generate Response
Stream-based output, storage, logging. Without solution protection.
7. Solution Protection
Automatic comparison with model solution. On violation: restart with stricter prompt.
No Solution Protection Needed
8. Save & Log
Response in DB, all steps logged for quality analysis.
Save & Log
Response in DB, all steps logged for quality analysis.
Core Differences
No RAG Retrieval
No Solution Protection
Simplified Prompt
Targeted search for text passages from the learning content and use as context (Step 4).
Safety guard checks whether the AI reveals the model solution (Step 7).
Tutor rules, learning passages, task context, and didactic modes.
No access to learning material. Only general knowledge of the AI model.
No tasks available, so no solution protection needed.
General chemistry tutor without material-specific instructions.
How the AI Accesses Learning Material
The Retrieval-Augmented Generation (RAG) method enables the AI to respond specifically from the learning material.
Step 1
Preparation
Learning material is automatically broken down into sections (chunks). Each chunk receives a mathematical representation (embedding) with 3,072 dimensions. Images are described by AI.
Step 2
Search
The student question is converted into a vector. The system searches in parallel: semantic similarity (vector search) and word matching (full-text search). Both results are combined.
Step 3
Integration
The 5 most relevant passages go to the AI along with the question, chat history, and tasks. It "reads" the learning material and formulates a technically correct answer. Images are also sent as context.
Step 1
Preparation
Learning material is broken into sections. Each receives a mathematical representation (3,072 dimensions). Images are described by AI.
Step 2
Search
Student question is converted to a vector. Parallel search: semantic similarity + word matching. Results are combined.
Step 3
Integration
5 most relevant passages + question + tasks to AI. Formulates technically correct answer. Images also as context.
Didactic Control of the AI
How the system ensures the AI acts as a tutor — and not as a solution provider.
Intent Routing
Each message is automatically assigned to a category. The AI responds differently depending on the category:
Help with task → Scaffolding (without solution)
Concept question → Explanation mode with learning material
Check solution → Evaluation mode
Greeting, farewell → Short response
Non-chemistry → Redirect to learning
Solution Protection
Before the AI response is displayed, an automated system checks whether it contains the model solution.
✓ Response Passes Check
Displayed directly. Hints and thought prompts, but not the solution.
⚠ Partial Solution Detected
Second attempt: AI should only state the next thought step.
✗ Complete Solution Leaked
Response is replaced: "I cannot tell you the solution directly."
Direct Comparison
The key differences at a glance.
| Property | Learning Material Chat | Free Learning Chat |
|---|---|---|
| Location | Embedded in learning page | Standalone page |
| Knowledge Base | Learning material of the subtopic | General AI knowledge |
| Task Reference | Yes — tasks + student answers | No |
| Solution Protection | Yes — automatic check | No |
| Didactic Modes | 3: Scaffolding, Explanation, Evaluation | General Tutor |
| Message Limit | Configurable per subtopic | Unlimited |
| Personalization (optional) | Self-assessment, Encouragement style, Interests, Learning | Self-assessment, Encouragement style, Interests, Learning |
| Image Upload | Yes (if enabled) | Yes |
| Sessions | One per student + subtopic | Any number, sidebar |
| AI Model | Per subtopic / class / global | Per class |
Using CHETI for your research? Find the appropriate citation in APA 7 format here.
Fitting, N., & Seibert, J. (2026). CHETI: AI Tutor for Chemistry. Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau. https://cheti-ki.de/