Part 1 of 5 · 1 min read
Overview
In this module you will
- Explain how BM25 (keyword) search scores passages and when it beats vector search.
- Explain when vector search beats keyword search, and why neither is enough alone.
- Combine two ranked lists with reciprocal rank fusion (RRF) and with a weighted score.
- Add a reranker as a second stage and say what it can and cannot fix.
- Measure all of this by kind of question, not with one average.
Agents ask retrieval questions of every shape: an exact error code from a log, a product name, a vague description of a symptom. A vector-only system can blur a rare identifier such as E-4021 into its neighbours. A keyword-only system cannot connect "money back" to "refund". A hybrid system with a reranker is the practical default because it covers both, and because the model downstream only sees a handful of passages, so the order of the top few matters a lot.
About the examples
The corpus is an invented 18-passage help centre and the 18 questions were written by me. The "dense" retriever is a toy: related words share coordinates through a hand-made table, and tokens containing digits are deliberately down-weighted to stand in for the way real models often handle rare identifiers poorly. The reranker is a small rule program that scores a (query, passage) pair together. They show the mechanism and the direction of each effect; the percentages are not benchmarks, and a real embedding model or cross-encoder will differ. No real model or service was called.