ML//RAG//ColBERT

ColBERT is a late-interaction retrieval model, a descendant of BERT: it keeps per-token embeddings instead of compressing each text into a single vector. It bridges cheap bi-encoders and expensive cross-encoder rerankers.

ColBERT is a late-interaction retrieval model, a descendant of BERT: it keeps per-token embeddings instead of compressing each text into a single vector. It bridges cheap bi-encoders and expensive cross-encoder rerankers.

Its MaxSim score matches each query token against its best document token and sums the results, which is more expressive than cosine similarity. ColPali and ColQwen extend it to vision, retrieving documents by their visual layout as well as their text.