Skip to content

Evaluation

How well the tutor finds sources, answers, refuses and personalizes. Retrieval run none yet, answer quality run none yet, simulation run none yet.

Loading results…

RetrievalHit rate by question type, rank of the first right source, refusal gate calibration
Answer qualityFaithfulness, relevancy, context precision and recall, with the DeepEval cross-check
Simulated studentsMastery curves and gain per persona, adaptive against a fixed order

How to read these numbers

  • The golden questions were generated from the indexed sources by an LLM and have not all been checked by a person. Gold locators come from stored page, slide and timestamp fields.
  • Judge metrics use a different model (Groq or Gemini) from the answer generator (DeepSeek), and are computed on a limited sample of answers.
  • Simulated students follow an assumed learning model, so the gains show the effect of the question-picking policy under that model, not results with real learners.