Why it matters

Retrieval architectures shape RAG. Understanding shapes design.

Advertisement

The architecture

Dense: fast bi-encoder.

Sparse: BM25 + learned SPLADE.

Cross-encoder: reranking.

Late-interaction: ColBERT.

Retrieval architecture spectrumDense bi-encoderfast semanticSparse BM25 / SPLADElexical + learnedReranking + late interactioncross-encoder / ColBERTCombine for best results: dense + sparse retrieve, cross-encoder rerank
Retrieval arch.
Advertisement

How it works end to end

Bi-encoder for scale.

Cross-encoder for precision.

Sparse for lexical.

Hybrid for best of both.