ColBERT models fine-tuned for efficient retrieval. Optimized from pre-trained architectures for token-level similarity.