Perplexity opens paired visual-document retrieval models
Perplexity released MIT-licensed 9B and 0.6B pplx-embed-v2-late models for retrieving text, images and document pages without first extracting page text. The models share an embedding space, so teams can index with the larger model and query with the smaller one.
Why it matters: The paired design creates a practical quality-versus-latency tradeoff for visually rich PDFs, slides and scans. It can preserve layout information that OCR-only pipelines lose, but increases index size and GPU requirements.