Re-ranker
ColBERT re-ranker
ColBERT is a technique that creates separate detailed multi-vector representations for both queries and documents. It then uses a soft and contextual approach to locate each token in the query within the document. For each query embedding, it identifies the most similar embedding in the document using a “MaxSim” operator, and calculates the score of the document by summing up all of these MaxSims. The “MaxSim” operator was chosen carefully so that the document embeddings can be indexed for Approximate Nearest Neighbor (ANN) search, making it possible to apply this technique to millions of passages while keeping the response time in the order of tens of milliseconds.