Announcement
FLINT proposes fixed-budget vectors for long-query search
The authors report index sizes reduced by 5.3–9.3 times versus token-level models such as ColBERTv2.
TLDR
In a paper posted to arXiv, FLINT’s authors describe pooling both queries and documents into a fixed number of vectors instead of keeping a vector for every token. They report comparable ranking effectiveness to the closest baselines on BEIR, stronger results on several benchmarks with longer queries, and 1.5–2.8 times faster end-to-end retrieval than ConstBERT on long-query corpora.
Combined views
1.4K
3 Sources, first seen ago
