Perplexity previews context-aware retrieval beyond gold chunks
Perplexity described a teacher-trained 9B contextual embedding preview that represents passages using surrounding document context and evaluates retrieval of answers plus supporting evidence. The company reports leading results on public ConTEB and a new context-bench maintained privately with turbopuffer; existing MIT-licensed v1 contextual models remain downloadable.
Why it matters: Retrieval evaluation that rewards complete evidence rather than one labeled passage better matches source-grounded RAG and agent workflows. The preview also exposes a harder deployment question: whether contextual gains transfer to a private corpus without…