A Reranker Algorithm Showdown for Vector Search

Vector search effectively delivers semantic similarity for retrieval augmented generation but struggles with short keyword searches or out-of-domain terms. Supplementing vector retrieval with keyword search like BM25 and combining the results using a reranker is becoming the standard approach to achieve optimal performance. Rerankers are machine learning models that reorder search results to improve relevance by examining queries paired with each candidate result in detail, which can be computationally expensive but produces more accurate results than simple retrieval methods alone. In a test of six rerankers on the ViDoRe benchmark dataset, all ML-based rerankers tested delivered meaningful improvements over pure vector or keyword search, with Voyage rerank-2 setting the relevance bar. However, tradeoffs exist: superior accuracy is offered by Voyage rerank-2, faster processing by Cohere, and solid middle-ground performance by Jina or Voyage's lite model. Even the open-source BGE reranker adds significant value for teams choosing to self-host.