AI Briefing
KOSign in

Qdrant tests Perplexity’s PDF embeddings: 21× smaller compressed vectors, with originals still stored

·2026.10.17 11:52

Key point

1-bit TurboQuant compression reduces vector size from 856 KB to 40 KB per page while retaining 96.5% of float32 retrieval scores.

Details

Qdrant evaluated Perplexity’s new visual document embeddings on 11,609 pages across four ViDoRe v3 datasets. These models treat pages as images, preserving tables and layout structures during search.

Compression and Accuracy

The test utilized 1-bit TurboQuant to compress vectors, achieving a 21x reduction in size:

  • Original size: 856 KB per page
  • Compressed size: 40 KB per page
  • Retrieval score retention: 96.5% of the float32 baseline

When rescoring against the original uncompressed vectors, the retrieval accuracy improved to 99.9%.

Storage Implications

A critical limitation is that the reported 40 KB compressed size excludes the original vectors, which Qdrant continues to store. The figures regarding million-page scalability are projections, not results from a million-page benchmark. This evaluation is conducted by Qdrant itself.

embeddingsperplexitycompressionvector-databaseqdrantdocument-search
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

TLDR AIPerplexity Releases pplx-embed-v2-late: Multi-Vector Embeddings for Text and Visual DocumentsRedditEmbedFlow Released to Eliminate Downtime During Embedding Model SwapsHacker NewsEmbedding Space Conversion Without Paired Data Raises Vector DB Security ThreatsLangChain BlogQdrant x LangChain: Achieving Final PerformanceAWS BlogAmazon DynamoDB Supports Real-Time Vector Search at Any ScaleHacker News14x Faster Embeddings: How the ONNX Path Was Rebuilt in Manticore