Chapter 4

Retrieval Engineering

BM25 · Hybrid · Rerank · Eval

BM25

Sparse / lexical

SKUs, IDs, exact terms · rank_bm25 code

Dense

Semantic

BGE / E5 / OpenAI · FAISS index

Hybrid

RRF merge

BM25 + dense parallel · $\\frac{1}{k+rank}$

Rerank

Two-stage

Top-50 recall → cross-encoder → Top-5

Pipeline

Production stack

hybrid → rerank → dedupe → LLM

Eval

Golden set + RAGAS

Recall@k · MRR · faithfulness · context recall

Lab

Ablation study

4 configs × 20 questions → pick winner

Retrieval done

Scroll lab: course-full-retrieval.html#retrieval-practice

Next: ⑤ LangChain & LangGraph

Intro
1 / 1 Hub Space F