Cascading retrieval with multi-vector representations: balancing efficiency and effectiveness

What this post added

Introduces ConstBERT, a novel multi-vector retrieval model that uses fixed-size document representations to reduce storage and improve query efficiency compared to traditional token-level multi-vector approaches. Details its integration into cascading retrieval pipelines as a second-stage re-ranker, complementing existing dense and sparse retrieval methods before final cross-encoder re-ranking. Highlights the benefits of fixed-size representations for indexing, cache locality, and scalability.