Zahan Ed
Offline RAG pipeline: textbook → chunk with page metadata → Sentence Transformers embeddings → FAISS retrieval → citation-enforced generation → hard refusal on out-of-scope queries. Runs on CPU. LAN-deployable. 4-8GB RAM.
Citation accuracy
85%
Local portfolio doc
Correct refusal
100%
20-question test set
Retrieval
under 100ms
Local portfolio doc
Layer
Query
Student question
Layer
FAISS
Vector retrieval
Layer
LLM
Citation-grounded gen
Layer
Answer
Source + page ref
Refusal path
Out-of-scope query → hard refusal. No hallucination fallback.
Problem
Education AI needs a trust layer. A wrong confident answer is worse than no answer.
- Every answer needs source traceability.
- Out-of-scope questions need refusal.
- Target environments may have no reliable internet.
System
The stack uses source-aware retrieval before generation.
- Textbook chunks carry page metadata.
- Sentence Transformers create embeddings.
- FAISS retrieves relevant chunks for answer generation.
Shipped proof
The evaluation focuses on citation quality, refusal behavior, and latency.
- 85% citation accuracy.
- 100% correct refusal.
- Under 100ms retrieval latency.
Lesson
RAG quality is not a vibe. It needs metrics that map to user trust.
- Measure refusal, not just answer fluency.
- Keep page citations explicit.
- Design for hardware constraints early.
Evidence links