Reranking Is Now Optional on Multimodal Collections

  • rerank=false is honoured everywhere: multimodal collections previously forced reranking on and ignored the flag. They now rerank by default but respect an explicit rerank=false. The response carries a warnings entry noting that cross-modal ranking quality is reduced, since results from different modalities are then ordered by their own pipeline scores. Reranking costs roughly 200 ms, so this is a real lever for latency-sensitive workloads such as voice assistants. Text-only collections are unchanged. See Multimodal Search.
  • New rerank.reason value: v3 query responses report default_for_multimodal when a multimodal collection reranks without the caller asking. required_for_multimodal remains valid in the schema for pinned clients but is no longer returned, and a disabled rerank now reports disabled on every collection type.
© 2026 Captain