🎯 About this course
Most RAG demos fall apart on real documents and real users. This course teaches the engineering that fixes that. You will design chunking that preserves context, build a vector store from first principles, add BM25 and rank fusion, re-rank and filter by permission, and — most importantly — measure everything with recall@k, MRR and nDCG.
The final unit teaches grounding: numbered sources, citation verification, confidence thresholds and prompt-injection defences. Labs are dependency-light and auto-graded, and the applied studio has you improve a real handbook-QA pipeline with an experiment log.
👤Who it is for
Engineers who already call LLM APIs and now need retrieval that works at scale.
✨What makes it different
Every technique is paired with a metric and a failure analysis, not just a code recipe.
By the end you will be able to
- Design chunking strategies that preserve context and metadata, and justify chunk size with measurements.
- Explain embeddings, similarity metrics and exact versus approximate search, and build a working vector store with metadata filtering.
- Implement hybrid retrieval combining BM25 and vector search with rank fusion.
- Apply re-ranking, MMR and retrieval-time access control to raise precision without leaking restricted content.
- Build retrieval evaluation sets and compute recall@k, MRR and nDCG to compare and tune retrievers.
- Ground answers with verified citations and calibrated "I don't know" behaviour, and defend against prompt injection in retrieved text.
- Deliver a measured, permission-aware RAG pipeline with an experiment log documenting each improvement.