Perplexity open-sources two models for searching text, images and document pages
Redacción IvokaOct 11, 6:00 PM 3 min
The pplx-embed-v2-late models (0.6B and 9B) are available on Hugging Face under the MIT license and are meant for building search over your own files.
On October 7, Perplexity open-sourced two embedding models called pplx-embed-v2-late, at 0.6B and 9B parameters. They are what lets a search tool find the right document, image or page inside your files even when you don't use the same words.
What they do
- They retrieve text, images and document pages (including PDFs) in a shared space across both sizes.
- They can be combined: a corpus indexed with the 9B model can be queried with the lighter 0.6B model without rebuilding the index.
- According to reports, the 0.6B model scores 62.3% on the ViDoRe v3 benchmark, rising to 63.5% when the index is built with the 9B.
- Both are publicly available on Hugging Face under the MIT license. API access has been announced as forthcoming; for now the release is for those who self-host.
Who it's for
Mostly developers: it's the piece behind assistants that "answer from your documents" (catalogs, manuals, contracts). If you don't code, you don't need to install anything.
Performance figures are as reported by Perplexity itself, according to the outlets covering the release.
Sources
