Source: Qdrant tutorial: PDF Retrieval at Scale (ColPali/ColQwen)
Qdrant tutorial that makes ColPali/ColQwen indexing practical by HNSW-indexing row/column mean-pooled multivectors and reranking with the full multivector, which is stored without an HNSW graph.
Contents
- What: vendor tutorial
- Author/publisher: Qdrant
- Date: read 2026-11-14 (no publication date shown)
- Original: PDF Retrieval at Scale
Summary
Shows why naive HNSW indexing of visual late-interaction vectors is slow, and a two-stage fix: mean-pool patch vectors by rows and by columns into small multivectors, index only those with HNSW, prefetch candidates from them, then rerank with the original full multivector.
Key claims
- ColPali produces about 1,030 vectors per page; ColQwen about 700 (up to 768).
- Inserting one page into an HNSW graph costs on the order of 700 x 700 x 100, about 49 million vector comparisons; quantization saves memory but not these comparisons.
- Row and column mean pooling turns ColPali’s 32x32 patch grid into 32 vectors each (special tokens appended back).
- Setup:
originalmultivector with HNSW disabled (m=0) used only for reranking; two pooled multivectors with default HNSW; prefetch 100 from each, rerank, return top 10. - Indexing is an order of magnitude faster with retrieval quality described as comparable (no numbers given).