Reranking: Definition and Examples

Reranking is a post-processing step used in information retrieval and Retrieval-Augmented Generation (RAG) systems. Its principle is simple: after a fast initial retrieval phase that returns a set of candidates, a more sophisticated model reassesses and reorders these results to place the most relevant ones at the top.

In a typical RAG pipeline, the initial search often relies on fast but approximate methods, such as cosine similarity between embeddings or keyword-based search (BM25). These approaches are effective at reducing a corpus of millions of documents to a few dozen candidates, but they sometimes lack nuance. The reranker then acts as a second filter, using a cross-encoder model that jointly analyzes the query and each candidate document to produce a much more accurate relevance score.

The most common reranking models are transformers specifically trained for this task, such as Cohere Rerank, cross-encoder models from the SBERT family, or Jina AI re-rankers. Unlike the bi-encoders used during the retrieval phase, these models take the query-document pair as input and can capture fine semantic interactions between the two texts.

Reranking has become an essential component of modern RAG architectures because it significantly improves the quality of responses generated by LLMs, ensuring that the context provided to the model contains the most relevant information. The computational cost remains reasonable since the reranker only processes the top N results from the initial search, typically between 20 and 100 documents.