Qdrant tests Perplexity's visual PDF embeddings: 21x smaller vectors, 99.9% after rescoring
jrxna · reddit · 2026-10-16
Qdrant independently benchmarked Perplexity's new visual document embeddings on 11,609 pages across four ViDoRe v3 datasets. The models search pages as images, so tables and layout survive intact.
Key results
- 1-bit TurboQuant shrinks compressed vectors from 856 KB to 40 KB per page (21x smaller) while keeping 96.5% of the float32 retrieval score
- Rescoring against originals brings that to 99.9%
The catch
- The 40 KB excludes original vectors, which Qdrant still stores — the 99.9% figure costs you full-size storage plus rescoring disk reads
- Million-page figures are projections, not benchmarks
For document search teams, the open question is where the tradeoff stops being worth it: rescoring disk reads, indexing cost, or errors on real queries.
More from Infra
- LLM scaling framed as industrial revolution 2.0: turning fuel and power into cognitive labor — menhguin · 2026-10-16
- NVIDIA signals $2B in new European AI deals at GTC Berlin, totaling ~$10B in a year — CurieuxExplorer · 2026-10-16
- Making Your Data Ready for Agentic AI: trust first, then semantics, then writes — bibryam · 2026-10-16
- China's AI boom drives up to 660% profit surges for domestic chip firms; GPU maker MetaX turns profitable — pstAsiatech · 2026-10-16
- When Does an Agent Actually Need an API Gateway? Reddit Weighs Direct Calls vs. Unified Endpoints — TurbulentStiffness · 2026-10-16
- Taiwan to build $4B AI manufacturing complex in Texas to back Nvidia and TSMC — Polymarket · 2026-10-16