A new retrieval system outperforms the voyage-context-4 by 14.4 points in answer recall@10 during a blind evaluation on the turbopuffer private context-bench. Perplexity developed the tool, named pplx-embed-v2-context-9b-preview, to encode document chunks with the entire document in view, and the company is open sourcing the model.

The 9B parameter model sets a new state of the art on ConTEB and turbopuffer benchmarks by distilling relevance from a context compression model that scores every token against a query. At 1,024 dimensions in int8, the system uses 1KB per vector and still exceeds the performance of voyage-context-4, which requires 8KB per vector.