Late interaction seems like it could be the sweet spot between full cross-attention and bi-encoders. Curious how it scales with really large document collections in production, and if we really want that in any case vs. optimizing compaction methods 🙂

→ View original post on X — @whats_ai