Details
- Perplexity introduced pplx-embed-v2-context-9b-preview, a 9-billion-parameter contextual embedding model for retrieval systems.
- The model encodes each document chunk with the full document in view, then pools chunk vectors for retrieval.
- Its training uses relevance signals distilled from a query-aware context compression model rather than relying on one gold chunk per query.
- The system scores document tokens against a query and converts those scores into chunk-level targets for retrieving both answer and supporting passages.
- Perplexity says the model leads on Answer and Evidence retrieval at every cutoff in turbopuffer’s privately held context-bench, which contains 2,099 queries and 38,894 documents.
- On ConTEB, the preview records the highest average nDCG@10 among tested models, although it does not lead on every task.
- Perplexity reports that it beats voyage-context-4 on chunk retrieval while using 1 KB per vector, compared with 8 KB for voyage-context-4’s 2,048-dimensional float32 vectors.
- The model is available as a preview through Hugging Face, while Perplexity’s earlier 4-billion-parameter context model remains slightly ahead on some document-retrieval cutoffs.
Impact
The release targets a growing retrieval bottleneck: chunking long documents can discard the context needed to identify relevant evidence. Perplexity’s lower storage requirement could reduce vector-database costs for large deployments, while its benchmark results position it against Voyage AI’s contextual embedding products. Because context-bench is privately held and the model is still a preview, independent validation and production pricing will determine whether the reported quality and storage advantages translate into broader adoption.
