Reranking

First-stage retrieval, keyword and embedding search, is built for speed: it scans millions of documents and returns perhaps a hundred candidates, with imperfect precision. The reranker then examines each candidate against the query closely, typically with a cross-encoder model that reads query and passage together, and produces a refined relevance ordering.

Only the top handful of reranked passages enter the model's context window. In practice, reranking is the cut that separates pages that influence an answer from pages that merely existed in the candidate pool.

Cross-encoders evaluate genuine semantic fit: does this passage actually answer this question? They punish superficial keyword matches, off-topic tangents, and passages where the relevant information is diluted across unrelated text. They reward direct answers, completeness within the passage, and precise alignment with the query's specific intent, including its qualifiers and constraints.

Authority and freshness signals are often blended in at this stage too, so a clear passage on a trusted, recently updated page beats an equally clear one on a stale or unknown domain.

Write passages a strict judge would score highly: the question's answer stated explicitly, qualifiers addressed ("for small teams," "in 2026"), no burying the point under context-setting. Each section should survive being read alone against a specific query, that is literally the test it will face.

You see reranking outcomes only in their end product: citations. Comparing which of your pages get cited versus which never appear, across many tracked prompts in Geonimo, shows where your content clears the rerank and where it dies in the candidate pool, guiding content gap fixes.