A number in a document may lose its meaning when separated from its row, column or chart label. Searching page images can preserve that layout, but representing each page with many vectors creates a different infrastructure problem: how much memory and storage the retrieval system needs.
Qdrant’s October 13 experiment evaluates Perplexity’s late-interaction embedding models on four ViDoRe v3 document sets. It reports reducing a page’s compressed vector representation from 856 KB in float32 to 40 KB with one-bit TurboQuant while retaining 96.5% of baseline ranking quality. Rescoring raises reported retention to 99.9%. The company explicitly says those compressed sizes exclude stored original vectors, and that million-page totals are projections rather than a benchmark at that scale. Read the primary source.
Include the second read
Rescoring uses the original representation to check shortlisted candidates. Keeping compressed vectors in memory and originals on disk can change the memory requirement, but it does not make the originals disappear. Storage capacity, cache behavior and extra reads all belong in an application’s cost model.
The practical opportunity is to test visual document retrieval without first converting every page into plain text. That does not eliminate text extraction when an application needs exact fields or a structured record. A controlled comparison should measure retrieval quality and latency on the same questions, with rescoring clearly enabled or disabled.
Compression could make layout-preserving retrieval more accessible for large document collections. The evidence needed next is performance on changing collections, realistic concurrency and the full storage hierarchy. The important choice is the quality and operating-cost combination, not the smallest compressed-vector number in isolation.
References
Questions, news tips or corrections? Contact [email protected].
