Qdrant Shrinks Perplexity's Visual Retriever Vectors 21x With TurboQuant, Keeps 99.9% Quality

qdrant_engine · x · 2026-10-14

Qdrant benchmarked Perplexity's newly released visual retrieval embedding model pplx-embed-v2-late and found a practical problem: vectors weigh 856 KB per page, making storage costly.

Using Qdrant's TurboQuant quantization across 4 ViDoRe v3 datasets, they compressed vectors to 40 KB per page — a 21x reduction — while retaining 99.9% of baseline retrieval quality.

The trade-off: rescoring is required to preserve accuracy, meaning the original (uncompressed) vectors must be kept available. For PDF/document visual retrieval, where tables and charts need their context retrieved, this is a workable storage optimization.

Original post →

Qdrant Shrinks Perplexity's Visual Retriever Vectors 21x With TurboQuant, Keeps 99.9% Quality

More from Models