pplx-embed-v2-late

For developers building search: two Perplexity models that find matching text and images, including PDF pages, without converting them to text first.

Most search models squeeze each document into one summary vector. These keep a small vector for every word or image patch, so different parts of a question can match different parts of a page
Searches rendered PDF pages straight from the image, skipping the step of pulling text out first
Two sizes, 0.6B and 9B parameters, that share the same embedding space, so you can index with one and search with the other
Publicly available on Hugging Face and works with the sentence-transformers and transformers libraries
Joins pplx-embed-v2-context, a contextual embedding model Perplexity released the week before