Abdelrahman Abdallah PRO

@Nomad-link-id Thanks, this is a fair point. The claim in the post is scoped to replacing the dense + late-interaction pair (one model and one index instead of two), not the lexical arm. UIBK has no sparse/lexical head, and BM25 is clearly stronger on LIMIT (97.8 vs 91.4), so the post doesn’t show that it replaces a dense + BM25 hybrid..

On long inputs: the token head reads the first 64 query tokens and the dense head reads up to 256, by design. We haven’t yet split results by where the decisive term sits in a long query, and we’ll check that on ArguAna/Touché-style queries. For long documents we do have evidence: on LongEmbed, UIBK-Embedding-v1 ranks #1 with a mean of 91.53, ahead of BM25 (90.87) and GTE-ModernColBERT (90.58), and it stays at 97 at 32K tokens, where GTE-ModernColBERT drops to 87. That is long-document retrieval rather than long queries, so it doesn’t replace the slice you suggest, but it shows the model doesn’t degrade with input length.