Week 06 — Retrieval-Augmented Generation

Don't make the model memorize everything — let it look things up. The dominant production pattern for grounded LLM applications.

Lecture

The RAG architecture (Lewis et al. 2020) · chunking strategies · embedding models (dense, sparse, hybrid) · vector databases (FAISS, Chroma, Qdrant, Pinecone) · reranking · prompt assembly · evaluation (faithfulness, context relevance, answer correctness).

Read before the lecture

• Lewis et al., *Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks* (NeurIPS 2020)

Code lab

Lab 4 — RAG over a domain corpus

Index a set of public domain documents (WHO/AFRO reports, scientific papers, course PDFs) in Chroma. Build retrieval + reranking + generation. Evaluate on a 30-question held-out set.

Notebook: lab04-rag.ipynb  ·  Dataset: Curated public PDFs (provided).

Reference text for this week: chapter 06 of the bilingual notes — EN PDF · FR PDF.