reranked

Reranking refers to the process of reordering a set of items that have already been retrieved by an initial retrieval stage. The goal is to improve relevance or other objectives by applying a more sophisticated model or a different signal to the candidate list without performing a full search over the entire collection.

Often used in two-stage retrieval pipelines: a fast first-stage retriever (such as BM25 or hashed dense representations) returns a candidate list, which is then re-ranked by a more expensive model—commonly a neural network such as a cross-encoder or a more powerful re-ranking model. Re-ranking can incorporate richer features such as query-document interactions, context, entity information, or user signals. Training typically uses learning-to-rank approaches with pairwise or listwise losses, and can be tuned to optimize metrics like NDCG, MRR, or MAP.

Applications include web search, enterprise search, e-commerce product search, and question-answering systems where precise ordering of results is critical. In practice, latency constraints are a central concern; reranking adds computational cost, so systems balance speed and accuracy, sometimes performing reranking only on the top-k candidates.

Challenges include maintaining fairness and reducing bias, handling domain shifts, and providing robust performance with limited annotated data. Ongoing research explores more efficient architectures, end-to-end training strategies, and hybrid models that combine dense and sparse signals for reranking.