Reranks

Reranks refer to the process of reordering a subset of retrieved items using a secondary, often more expensive model after an initial retrieval stage. The aim is to improve relevance and user satisfaction by applying stronger signals or more sophisticated modeling to a small candidate set, while keeping overall latency low.

In a typical pipeline, a lightweight retriever generates a short list of candidates from a large collection. A reranker then analyzes these candidates with richer features, such as contextual signals, deeper textual representations, or user intent, producing a final ranked list. Reranking can be performed with traditional machine learning methods, such as learning-to-rank algorithms, or with neural models, including cross-encoder or bi-encoder architectures. Listwise or pairwise objectives are common, and rerankers may incorporate user feedback, session history, or diversity considerations.

Common techniques involve models that compute more detailed relevance scores for each candidate or consider interactions among candidates. Rerankers may use sophisticated text embeddings, query-document interactions, or multimodal features. They often operate on a much smaller candidate set (for example, top 100) to balance quality and latency. Training data typically comes from labeled relevance judgments, click data, or weak supervision, and may require calibration to avoid overfitting to historical biases.

Evaluation of reranking systems uses metrics such as NDCG, precision@k, or MRR, often assessed offline and through online A/B testing. Challenges include computational cost, latency constraints, data drift, and potential biases that can be amplified by reranking. Reranks are widely used in web search, e-commerce, question answering, and recommendation systems to enhance matching quality while maintaining responsive performance.