Open-weight vision-language embedding models in 9B and 0.6B sizes, returning an int8 vector per document or a document-aware vector per chunk.
pplx-embed-v2-context is a family of open-weight contextual embedding models in 9B and 0.6B sizes for retrieval over text, images, and document pages. It can return an int8 vector for a whole document or vectors for chunks encoded with the surrounding document in view. Perplexity says the two sizes share an embedding space, allowing a 9B model’s 1024-dimensional prefix to be used for indexing and the 0.6B model for queries. The accompanying paper reports evaluations on ConTEB, context-bench, and Q2D-Chunk.
Transformers is used to load and run the model; the model card gives Transformers loading examples and specifies trust_remote_code=True.
