Alibaba’s H+ Embedding cuts retrieval vectors by 13.7% while matching token-level quality

_reachsumit · x · 2026-08-04

Alibaba’s H+ Embedding introduces context-dependent phrase units between global vectors and token-level retrieval.

The paper argues that terminology-heavy retrieval needs something more granular than a single embedding but cheaper than full token-level late interaction. H+ Embedding predicts variable-length phrase partitions, keeps uncovered tokens as singletons, and applies importance-guided selection with weighted MaxSim. Across 16 scientific, medical, and bilingual tasks, the phrase branch beats the global branch by 6.91 macro nDCG@10, nearly matches token-level retrieval, and uses 13.7% fewer document vectors. The authors position it as a practical middle ground between compression quality and indexing/scoring cost.

Original post →

More from Infra

Infra channel →