RAG Class Assistant
A production RAG assistant for Calculus II, built and deployed inside OneNote, with slide-level citations for every answer.
600
engineering students served
94%
accuracy, 47/50 on a validated question bank
1,183
production chunks indexed
An end-to-end retrieval-augmented generation system, engineered and deployed as university infrastructure for 600 engineering students in Al Akhawayn University's Calculus II course. It answers with exact chapter, section, and slide citations at 94% accuracy (47/50 on a validated question bank) and runs inside Microsoft OneNote as a native task-pane add-in. Built under Prof. Sofiane Grira as both a production deployment and an active research project: a paper on hybrid retrieval and cross-encoder reranking for mathematical educational content is in progress, alongside a PIF Spring 2027 grant proposal (72,000 MAD).
Students navigating 1,000+ lecture slides across multiple OneNote class notebooks had no way to instantly locate, cross-reference, and understand specific calculus concepts, from integration by parts to p-series convergence, with precise slide-level citations.
A production-grade RAG pipeline that ingests, indexes, and retrieves mathematical content, delivering answers with exact chapter, section, and slide references, deployed directly inside Microsoft OneNote as a native task-pane add-in.
Built end to end, from the dual OCR math extraction pipeline to the hybrid retrieval engine, the OneNote add-in, and the security-hardened API, and deployed as real university infrastructure rather than a class project. Now scaling from free-tier prototyping to a paid Groq tier for the full engineering department. The paper with Dr. Grira will formalize the retrieval methodology, evaluation framework, and findings.
Hybrid BM25 + dense retrieval fused with Reciprocal Rank Fusion
Cross-encoder reranking and HyDE query rewriting
Dual-engine math OCR; every answer cites chapter, section, and slide
Multi-provider LLM failover with per-provider circuit breakers
