Independent chunking drops headings, definitions, and pronouns that live outside the window. Contextual embedding models encode the document once (late chunking), then pool tokens inside each span so similar sentences in different files can still be disambiguated. Perplexity's pplx-embed-context-v1 API and the 2026 pplx-embed-v2-context-9b-preview follow this pattern; they still emit one vector per chunk at inference, not an extra reranker.