Optimizing ColPali for Retrieval at Scale, 13x Faster Results - Qdrant

What this post added

Introduced a two-stage retrieval process for ColPali to address scaling challenges with visually rich PDFs. This involved pooling (mean and max) to reduce the number of vectors per page from 1030 to 38 for initial retrieval, followed by reranking with original high-resolution embeddings. Achieved a 13x speed improvement with mean pooling maintaining high NDCG@20 (0.952) and Recall@20 (0.917).