Perplexity open-sources pplx-embed-v2-late embedding models for text, images and pages
We're releasing pplx-embed-v2-late, two late-interaction embedding models that retrieve text, images, and pages with a shared embedding space for cross-model querying. Both models achieve frontier performance and are publicly available on Hugging Face.
Context
Developers get free multimodal search models, including a small one that can run queries on device.
Embedding models turn documents into vectors so search systems can find them. Aravind Srinivas said the release has 9B and 0.6B sizes in one shared space, so data can be indexed with the large model and queried on device with the small one. He added that it can search PDF pages without OCR.
Deep dive
Perplexity released pplx-embed-v2-late on October 7, two open embedding models for search across text, images and rendered document pages. Embedding models turn content into vectors that a search system can compare. These use a late-interaction design, which keeps a small vector for each token rather than one per document and scores matches token by token.
The two sizes, 9 billion and 0.6 billion parameters, share one embedding space, so a developer can build an index with the large model and query it with the small one, including on a device. Aravind Srinivas said the models can search PDF pages without OCR, and cited scores of 92.4% on MADQA and 64% on BrowseComp+.
Both models are on Hugging Face. Google released its own open multimodal embedding model, EmbeddingGemma 2, around the same time, adding to the open options for on-device multimodal search.