Qdrant says it shrank Perplexity’s visual-search vectors 21-fold

Qdrant says its TurboQuant method can shrink the vectors used by Perplexity’s visual document-retrieval model by 21 times while retaining 96.5% of retrieval quality. If the result holds up in practice, visual search could keep page layouts intact without paying quite so much in storage.

Watch Desk analysis

What happened

In a post published on 13 October, Qdrant describes tests on Perplexity’s pplx-embed-v2-late models. It says the method uses 1-bit quantisation, with optional rescoring, to reduce the size of page representations. The approach searches document pages as images, without converting them through OCR or text extraction.

Qdrant reports that the compressed vectors retained 96.5% of the original retrieval quality. That is the company’s result from its evaluation, not an independent comparison across other systems or workloads.

Why it matters

Keeping pages as images can preserve layout and visual context that plain text extraction may lose. The reported reduction could make storing and searching those representations less costly, especially for large collections of documents. A 21-fold compression claim is eye-catching; the more useful number for prospective users is how much quality survives alongside it.

Our read

This is a concrete infrastructure result, not just a clever-sounding compression technique: Qdrant gives a specific size reduction and a retrieval-quality figure, and explains how it avoids OCR. Those figures are promising, but buyers and developers will want to see how they translate to their own documents and retrieval tasks. Vector compression is less glamorous than a new model, but storage bills are famously unmoved by glamour.

What to watch

• How the method performs on document collections beyond Qdrant’s evaluation.
• Whether the 96.5% quality figure holds across different retrieval tasks.
• How optional rescoring affects storage, speed and retrieval quality.

Discussion spark: Would you accept a 3.5% drop in retrieval quality to cut visual-search storage by 21 times, or does that trade-off depend too much on the documents being searched?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.