Perplexity releases new open-source embedding model: can process entire documents as a whole and retain context even when chunked.

Beating AI News: Perplexity has open-sourced a new context embedding model, pplx-embed-v2-context-9b-preview, built primarily for Retrieval-Augmented Generation (RAG). Traditional embedding models typically split long documents into small chunks and process each individually, but this new model encodes all chunks together, allowing each chunk’s vector to reference the full document context. The model is now available on Hugging Face under the MIT license. For example, a sentence stating “It grew by 20%” is ambiguous on its own, as the subject of “it” is unclear; the new model can interpret this sentence by integrating context from other sections of the document. During training, it is specifically designed to retrieve both answers and supporting paragraphs, enabling downstream AI systems to receive not just answers, but also the evidence needed to verify them. In a blind test conducted on turbopuffer’s context-bench, the new model achieved an Answer@10 score of 45.5%, 14.4 percentage points higher than Voyage Context 4. Its Evidence Recall@10 reached 40.6%, a 5 percentage point lead. The test includes 38,894 long documents and 2,099 queries, with undisclosed questions to reduce the risk of test set contamination by training data. The model remains in preview. Perplexity warns that future weights, APIs, and generated embeddings may change, so vectors produced now may not be directly compatible with future versions.