txtai adds LEMUR and mean-centering techniques for late-interaction retrieval. LEMUR learns a fixed-dimensional encoding of multi-vector representations so late-interaction models can use standard vector indexes instead of specialized multi-vector stores, while mean centering corrects anisotropy in token vectors by subtracting the collection mean to expose directional variance. Benchmarks on three BEIR datasets show LEMUR improving 49-62% over size-matched alternatives, though results were limited to exact search with a single model and required careful configuration.
