Learning · Implemented
Retrieval, diagnosis, tutoring, and review
How evidence is selected and converted into a diagnosis, tutoring route, mastery update, and review decision.
Hybrid retrieval
Dense retrieval finds conceptual similarity; sparse retrieval preserves exact terms and identifiers. Candidates are fused, filtered by workspace and document scope, expanded through bounded graph paths, reranked, and checked for sufficiency.
- Every candidate retains a source chunk ID and canonical document reference.
- Graph traversal uses an allow-list, maximum depth, maximum nodes, workspace boundary, and timeout.
- The workflow can return insufficient evidence instead of forcing a diagnosis.
Teach-back diagnosis
Load the concept, source evidence, prior mastery, and graph context.
Detect grounded coverage, missing concepts, misconceptions, unsupported claims, and shallow explanations.
Require each non-unsupported gap to reference a source chunk or concept.
Calculate mastery and review recommendations with deterministic services.
Persist only validated evidence and select the next action.
Socratic tutoring
The tutor chooses one move at a time: clarification, contrast, evidence request, prerequisite recall, or final reconstruction. It asks one bounded question, evaluates the response, and either continues, escalates, or returns the learner to teach-back.
Mastery and review
- Mastery states summarize evidence; they are not presented as precise certainty percentages.
- Severe misconceptions override positive coverage until corrected.
- Review priority reflects the diagnosed reason, severity, prior attempts, and evidence quality.
- Completing or skipping a review records an action; mastery changes only after new evidence.