Module 2 · Retrieval that reasons · Lesson 5
Hybrid search and reranking
Why keyword search still matters for precise, specialist text, how BM25 scores documents, how reciprocal rank fusion combines keyword and dense results, and when a cross-encoder or LLM reranker is worth its latency.
Setup for the exercises:Install Claude CodePython with uvGet an OpenRouter keyLocal models with Ollama
By the end of this lesson you can
- Compute BM25 term weights and explain the roles of IDF, term-frequency saturation (k1) and length normalisation (b)
- Fuse BM25 and dense rankings with reciprocal rank fusion, by hand and in code, and explain why rank fusion beats score averaging
- Explain the difference between bi-encoders, cross-encoders, late interaction and LLM rerankers, and estimate reranking cost and latency
- Design a two-stage pipeline for case and reference-library search and decide from your eval whether each stage earns its place