Perplexity's New Late-Interaction Models Meet Qdrant's 21x Vector Compression

Perplexity’s pplx-embed-v2-late models enable document retrieval by embedding PDF pages as images, preserving visual context such as tables, charts, and layouts without requiring OCR, but their late-interaction architecture produces roughly 1,700 vectors per page and requires substantial storage. Tests using Qdrant’s 1-bit TurboQuant reduced vector storage from 856 KB to 40 KB per page, a 21-fold reduction, while retaining 96.5% of float32 retrieval quality, or 99.9% when top candidates were rescored using original vectors. Across four ViDoRe v3 datasets totaling 11,609 pages, image-based page retrieval slightly outperformed Markdown-converted pages, though Markdown used less compressed storage and may be preferable when textual content is sufficient. The evaluation also found that pages indexed with the larger 9B-parameter model could be queried with the smaller 0.6B model with only a modest quality loss, supporting an approach that concentrates compute during indexing rather than querying. The authors recommend testing compressed image-page retrieval and optional rescoring against unquantized baselines, while noting that OCR remains necessary when applications need extractable text or structured fields.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.