Perplexity's New Visual Retriever Has a Size Problem. We Fixed It. - Qdrant

This post details how Qdrant's TurboQuant quantization was used to address the size problem of Perplexity's pplx-embed-v2-late models, which embed document pages as images. The authors demonstrate a 21x reduction in vector size (from 856 KB to 40 KB per page at 1-bit) while retaining 96.5% of retrieval quality. Rescoring with original vectors boosts quality to 99.9%. The post also explores using smaller query models with larger index models and compares image-based search with OCR-derived text search, showing that image search can be competitive or superior depending on the content.