Speeding Up Multi-Vector Retrieval with CITADEL

Ever wondered why it takes so long to find the exact information you need? Multi-vector retrieval methods, which mix the best of both worlds from sparse and dense search techniques, are incredibly effective but slow and space-hungry. Enter CITADEL, a new way of thinking about token routing that makes multi-vector retrieval much faster without sacrificing accuracy. CITADEL works by cleverly directing different token vectors to specific "keys" so that a query token only interacts with document token vectors sent to the same key.

This clever routing system drastically cuts down on the computing power needed, making it nearly 40 times faster than previous top performers like ColBERT-v2. Whether you're searching within specific domains like MS MARCO or broader categories like BEIR, CITADEL delivers the same or even slightly better results. The code and data for CITADEL are readily available for anyone to explore and build upon.

Actions