Research prototype
Vidhya-Setu
Learning has a missing step. Find it, then move forward.
Phi-3 Mini · GBNF · BM25
Recreated interface animation based on captured prototype behaviour. Edited timing and camera movement; not a live application recording.
The route before the answer.
A bounded prerequisite path makes the sequence of concepts visible.
View detail ↗
A reply changes the direction.
This recorded status shows a confused response routing the session back one step.
View detail ↗My contribution, evidence and current limitations +
I built and later audited a textbook-derived concept graph. Hybrid retrieval combines BGE-small embeddings and BM25. A deterministic path planner selects up to eight concepts, while a local Phi-3 Mini model explains them and rules route student replies.
The saved graph contains 175 concepts and 197 prerequisite links, with no cycles. The current corpus contains 301 passages from two NCERT Class 9 chapters. Project documentation reports 46 regression checks.
This is a two-chapter prototype, not a validated learning intervention. Recorded explanations have factual weaknesses, and a documented backtracking issue remains. An 8 GB RAM recommendation is more accurate than a 4 GB claim.
Future investigation could compare learning-path quality with a plain retrieval baseline and review teaching content with subject experts.
