Reranking

Definition

After the fast but coarse similarity search in the vector index, the top-N candidates are passed to a reranker — typically a cross-encoder or a smaller LLM that scores each question/passage pair individually. The result is a re-ordered list that prioritizes the chunks most aligned with the question. Reranking improves answer quality at the cost of latency and compute, which is why it is typically applied only to a small candidate set.

How Swiss Knowledge Hub uses this term

Swiss Knowledge Hub combines vector retrieval with reranking so that even for ambiguous questions (for example legal phrasing or references to standards), the genuinely relevant passages are prioritized.

Related terms

• RAG (Retrieval-Augmented Generation)
• Embedding
• Semantic search
• Vector database (vector store)

Sources

1. Pinecone — Rerankers explained — https://www.pinecone.io/learn/series/rag/rerankers/