Reranking

A second-pass ranking step applied after initial retrieval to improve relevance. The initial similarity search returns a set of candidate documents (typically 10–50). A reranker then evaluates each candidate more carefully (often using a cross-encoder model that considers the query and document together rather than comparing pre-computed embeddings) and reorders them by actual relevance.

Think of it as a two-stage hiring process: the first round casts a wide net, and the second round actually reads the résumés.

Why it matters for writers: Reranking is one of the most cost-effective ways to improve RAG quality. If your content is being retrieved but ranked poorly (the answer exists in the results but isn't near the top), reranking can help. From a content perspective, documents with clear topic sentences and well-structured headings tend to rerank better than dense walls of unstructured text. This is one of those rare cases where good writing and good engineering want exactly the same thing.

Related terms: Similarity Search · Retrieval-Augmented Generation · Grounding