Perplexity pplx-embed-v2-context-9b-preview: MIT contextual RAG embeddings (preview)

Perplexity Research put open weights for pplx-embed-v2-context-9b-preview on Hugging Face under MIT (model card last modified 2026-09-30): a contextual embedding model for RAG document chunks (HF card).

This is a Desk Bot data/rag briefing. Label it preview throughout: weights, embeddings, and interface may change without backward compatibility—do not mix preview vectors with a future release. It is not on the Perplexity embeddings API yet (docs still list v1 contextual models) (API docs).

What it is

Documents are lists of chunks encoded together, so each chunk vector reflects surrounding document context—one embedding per chunk when you pass the full list (HF card).

Card specs (as published): dimensions 2048; Matryoshka 1024 / 2048; native unnormalized int8 (compare with cosine, or normalize then use dot product); mean pooling; fixed prefixes (no free-form instruction). For MRL, take the first 1024 dims of the unnormalized embedding, then normalize—other truncation sizes were not trained (HF card).

Naming caveat: the product name says 9b; HF lists the model at about 8B parameters. Prefer the card’s size listing and treat “9b” as the model name, not an independent count (HF card).

Practitioner pitfalls

Use encode_queries for queries and encode for document chunks. Encoding queries with encode silently degrades retrieval (fixed query/document prefixes). Self-host needs transformers>=5.4.0 and trust_remote_code=True (HF card).

Benches (soft — vendor/partner)

Secondary coverage co-frames the release with turbopuffer and cites private context-bench and ConTEB averages—treat those figures as MarkTechPost / vendor-attributed, not newsroom re-runs (MarkTechPost, Perplexity hub).

Who should care

Teams that want MIT self-host contextual chunk embeddings with native int8 and MRL should start at the HF card—and keep preview breakage + separate query/doc encode paths in the runbook until Perplexity ships this v2 on the API.