When RAG retrieval disappoints, engineers are told to 'add a reranker'. But explanations often stay at the architecture level—calling it a cross-encoder, attention mechanism, or relevance-trained model—without revealing what the model actually learned. This gap matters: teams cannot justify reranker use or spot when cheaper keyword lookup would suffice.

The honest answer is that a reranker is not smarter than the embeddings step below it. It uses the same statistical token association from training data, just conditioned on the query-passage pair instead of independent texts. Once understood, the decision to use a reranker shifts from 'because the tutorial said so' to 'only when this specific tradeoff is worth paying for'.

The article walks through the mechanism: a reranker is a transformer (cross-encoder) trained on relevance labels, producing a score based on token co-occurrence, not true understanding. It highlights three architectural consequences for enterprise RAG systems. A companion notebook lets users test the reranker on PDFs, showing which keywords drive rankings.

This article is Part I of the Enterprise Document Intelligence series, alongside the embeddings triptych (2A/2B/2C), and builds on the Prompt, Context, Loop framework.