Perplexitymodelpplx-embed-v2-context
Two open-weight Perplexity models that turn documents and images into searchable number strings while keeping each passage's surrounding context, for developers building their own search.
Details
- Comes in two sizes, 9B and 0.6B. Both return one compact vector per document, or one per chunk, where each chunk's vector is built after seeing the whole document.
- Reads text and images. A document page supplied as an image gets split into horizontal bands, with one vector per band.
- Trained differently from other contextual models. Instead of one correct chunk per query, relevance scores come from a compression model in Perplexity's own answer pipeline, so chunks that support an answer no longer count as wrong.
- Both also run in non-contextual mode, with one vector per document and no chunking. Documents embedded with one size can be queried with the other at 1024 dimensions.
- Weights are on Hugging Face. The 0.6B model card lists no hosted API, so expect to run it yourself.
Best for
Search across long documents, where a single passage only makes sense next to the rest of the document.
When to use it
Your team's custom search or assistant keeps pulling the wrong passage from long documents and you have engineers who can run an open model.
Who benefits
Engineers building search into internal AI tools. Marketers would only feel this through the tools built on top of it.
Join the newsletter to get ahead with AI.
I share strategies, new modules, free guides, and tools to get you better at using AI in your work.
Leading a team? See team programs →