Perplexity releases pplx-embed-v2 context embedding model

On October 1, 2026, MarkTechPost reported that Perplexity Research and turbopuffer released pplx-embed-v2-context-9b-preview, a contextual embedding model built for RAG pipelines. According to the report, each chunk is embedded with the full document in view, rather than being handled in isolation.

The report frames the real change as one of training signal. The model is trained to retrieve the answer together with the context needed to verify it, instead of retrieving a single “gold passage.” That training objective is described as distinct from earlier approaches that optimize for one passage at a time.

On deployability, the report answers yes, though the summary stops there and does not list deployment methods, supported platforms, or access channels. As of that report, the verifiable details are limited to the following: the release date of October 1, 2026; the two parties behind it, Perplexity Research and turbopuffer; the model name pplx-embed-v2-context-9b-preview; its positioning as a contextual embedding model for RAG; its handling of chunks with the full document in view; a training goal that covers both the answer and the context required to verify it; and confirmation that it can be deployed. The report does not mention pricing, context length, benchmark results, or the scope of availability, and no material currently covers those points.