UIBK-Embedding: Effective Text Retrieval by Unifying Dense and Late-Interaction Retrieval

Overview & Synopsis

Dense bi-encoders and late-interaction retrievers are usually treated as alternatives, yet they fail on different queries: one vector captures what a text is about, token-level matching finds its exact entities and phrases. Using both today means two models, two indexes and a score-fusion step. We present UIBK-Embedding-v1, a 179M-parameter English encoder that produces both representations in one pass and scores them with one unmodified MaxSim. The dense embedding is written into the list of token vectors as three extra vectors, and a one-dimensional tag keeps the two kinds of vectors from matching each other, so that plain MaxSim equals the token-level MaxSim plus a weighted dense cosine. The model therefore runs in a standard late-interaction index without changes; we test it with PLAID. A shared 18-layer trunk with a separate 4-layer top per head, and a cross-encoder distilled into the fused score, turn the two heads into one retriever. On BEIR, UIBK-Embedding-v1 reaches 59.79 nDCG@10. This is 2.57 points above LateOn, the strongest model of its size, and 2.10 points above it when both models read documents of the same length; UIBK-Embedding-v1 is ahead on all 15 datasets. On a decontaminated version of BEIR it stays ahead on average (61.55 against 60.36) and on 8 of 12 datasets. Among the 156 models with complete BEIR results in the MTEB results repository, only models with at least 1.5B parameters score higher. Code, model weights, and a technical blog post are available at https://github.com/DataScienceUIBK/uibk-embedding , https://huggingface.co/DataScience-UIBK/uibk-embedding-v1 , and https://huggingface.co/blog/abdoelsayed/uibk-embedding-v1 .