Perplexity Open-Sources Contextual Embedding Model

Perplexity is targeting RAG’s most persistent weakness: isolated chunks that retrieve the answer but lose the context needed to interpret it.

• –A context-compression teacher turns token-level relevance into softer chunk-level training signals.
• –The model produces one vector per chunk without adding inference-time reranking or storage overhead.
• –It reports 45.5% Answer@10 and 40.6% Evidence Recall@10 on turbopuffer’s context-bench, ahead of Voyage Context 4.
• –1024-dimensional int8 vectors reduce storage to roughly 1 KB each, making the approach attractive for large-scale vector databases.
• –Turbopuffer’s object-storage economics provide a fitting infrastructure layer for this push toward cheaper, document-aware retrieval.