Calibrating Exams Using
3 Different Parameters:
As soon as our platform grows we will use 3 different parameters for an optimal choice of question for your students

difficulty
It is a well known fact that newly created MCQ questions can be extremely hard for students to answer and this often results in very low pass rates upon the first usage of new questions in the exams.
Only after students repeat the exam a few times, can they actually gather the questions they remember and this results in somewhat better pass rates. Difficulty parameter is also to be used carefully with relevance-meter.
Our difficulty-meter aims to better optimize the question difficulty for first time exam takers to elimiate this issue.

Recency
Another issue with MCQ questions is that students oftenrecall and summarizethe questions they can remeber from the previous exams. This results in very high pass rates upon repeated usage of older questions.
Our recency-parameter ensures that students cannot easily pass the exams by simply recalling and writing down the questions from previous exams. Of course this parameter should be used in moderation along with the difficulty-meter.
Our recency-parameter aims to constantly provide students with new questions, these questions however should not be too hard and irrelevant.

relevance
Not a common issue with MCQ type of questions is a lack of relevance to the topics taught in your classes. However, if the questions are found somewhere ready-to-use online, then questions can often be completely irrelevant.
Our questions are tagged by keywords the students should know definitions and relations of, in order to successfully answer a questions. While this can narrow down the choice of questions,it is recommended to check the questions personally before finally giving them to the students.
Our relevance parameter aims to provide your students with relevant questions. Questions should however stick to general student-grade topics.
Use Without Registration:
AI System Prompt
Edit this prompt to customize question generation for this session. This will override once refreshed.
Choose the difficulty level for the questions to be generated.
Runs the separate AI relevance/difficulty scoring request right after generation.
Leaves each question's explanation blank instead of having the AI write one.
Generates higher quality distractors.
LLMs are known to make the correct answer obviously long. Check this box to fix it.
The system message is sent as part of the user message instead of as a separate system message.
Sends ONLY your text from the "AI System Prompt" box above + the file text. The built-in guidelines, the system message and the "Generate EXACTLY N questions" line are left out. The JSON format instructions are a separate option: "Request Json Format".
Asks the model to answer in JSON: sends the JSON format instructions with the JSON example, "Always respond with valid JSON only" and JSON mode (response_format). Untick it if your own prompt asks for a different answer format.
Does not force JSON mode in the request, so the model can answer like "Frage 1: … (A)–(E) … Antwort: B" or "Question 1: … (A)–(E) … Answer: B". An answer that is not JSON is always tried in this text format as well; if it is neither, you get the error "Invalid LLM response format".
Builds a knowledge graph from your text first, then generates questions from it. Produces more rigorously-grounded distractors. Uses significantly more API calls/time.
Sends only your uploaded file + the prompt above to DeepSeek as a single user message (no system message), with thinking enabled at DeepSeek's default effort. Nothing else is added: no difficulty block, no format instructions, no other settings. Turns off every other option and pipeline (including the Knowledge Graph).

Unknown Author
– Internet
Our previous work got recognition
Our work was featured in major national but also regional news outlets and media.
















