BM25 is 50 years old and still essential in 2026. Here's why every serious RAG system uses it alongside dense vectors.
Hybrid retrieval combines dense and sparse methods, fused into one ranked list. It's the production default for a reason.
HyDE has the LLM hallucinate an answer first, then retrieves documents similar to that hallucination. Counterintuitive and often effective.
One query produces one retrieval. Multiple query variations produce many retrievals that can be fused for better coverage.
Users ask poorly-phrased questions. An LLM can rewrite them into forms that retrieve better. Here's how and when to do this.
Reranking is the single highest-leverage retrieval improvement. A cross-encoder on top of initial retrieval typically adds 10-30% to quality.
The core retrieval primitive in RAG. Simple in concept, with a few sharp edges worth knowing.