Context
Perplexity announced publicly available pplx-embed-v2-late models in 0.6B and 9B sizes. The late-interaction retrieval family supports text and image inputs, including rendered PDF pages, and shares a multi-vector embedding space across sizes.
What changed
Perplexity published its multi-vector multimodal embedding release and benchmark methodology, enabling a smaller query encoder to search document representations indexed with the larger model. The announcement does not establish that these models power its consumer search.
Why it matters
The work shows a concrete retrieval design that preserves token-level and visual evidence instead of compressing each document into one vector. For AI discoverability, source retrieval fidelity and final citation selection must be evaluated separately.
Visibility OS interpretation
Visibility OS interpretation: retrieval systems may surface different documents when queries match multiple facets or visual PDF evidence. The paper and announcement do not establish a production citation ranking change at Perplexity or any other AI answer engine.
What remains uncertain
Are these checkpoints deployed in Perplexity consumer search? How well do the controlled benchmark improvements reproduce independently and predict live-web source citation or answer absorption?