Reranking

Reranking takes the results of an initial, fast retrieval pass and re-scores them with a slower, more precise model, so a broader first pass can afford to over-fetch and still hand the model only the strongest results. The common pattern is to retrieve generously with vector search, then rerank down to a small, focused set.

Vector similarity alone is a rough filter: it finds passages that are topically related, not necessarily the specific one that actually answers the question. In a regulated answer, the model needs to see the passage that actually applies, not merely one that mentions similar terms.

Retrieving a broad set of passages that all mention a termination clause and reranking so the one from the contract that is actually in question ends up first, ahead of similar clauses from unrelated agreements.

In Dynamiq, a Cohere cross-encoder reranker, an LLM-based ranker, or a time-weighted ranker that favors recent documents can sit between any retriever and the model, on plain vector search results or on GraphRAG's returned facts, with over-fetching on the retriever and trimming on the reranker as the standard pattern.