Context
A Perplexity and turbopuffer research collaboration introduced pplx-embed-v2-context, contextual embedding models in 9B and 0.6B sizes, and benchmarks that distinguish answer-bearing chunks from supporting verification evidence.
What changed
The preprint describes distillation of graded chunk relevance from a query-aware context compression model, plus Context-bench and Q2D-Chunk evaluations. Its authors report strong results on their retrieval benchmarks and visual retrieval tasks.
Why it matters
AI search depends on finding not only answer snippets but supporting source evidence. Contextual chunk retrieval may change which document passages are available for grounding, though live citation impact is not demonstrated.
Visibility OS interpretation
Visibility OS interpretation: document-contextual retrieval and supporting-evidence coverage should be measured separately from whether a source is ultimately cited in generated answers.
What remains uncertain
Will independent studies reproduce the reported gains? Are these models deployed in consumer Perplexity retrieval? Do they affect citation frequency, fidelity or publisher referral traffic?