Perplexity open-sources new Embedding model: reads the whole document at once, and chunks can still carry context.

动察 Beating AI News Flash: Perplexity has open-sourced a new context embedding model, pplx-embed-v2-context-9b-preview, primarily for RAG retrieval. Ordinary embeddings typically split long documents into small chunks and understand them separately, while the new model encodes these chunks together, allowing each chunk to reference the context of the entire document when generating its vector. The model is now available on Hugging Face under the MIT license.

For example, if a passage only says "it grew by 20%," it is difficult to know who "it" refers to when taken in isolation. The new model can understand this sentence by combining it with other parts of the document. During training, it also specifically learns to simultaneously retrieve the answer and other passages that support the answer, so that downstream AI does not only get the answer, but also the evidence needed to verify it.

In the context-bench blind test designed by turbopuffer, the new model achieved an Answer@10 of 45.5%, 14.4 percentage points higher than Voyage Context 4; its Evidence Recall@10 was 40.6%, 5 percentage points higher. The test includes 38,894 long documents and 2,099 queries, and the questions are not public, which can reduce the risk of the test set being contaminated by training data.

The model is currently still a Preview version. Perplexity cautions that future weights, interfaces, and generated embeddings may all change, so vectors generated now may not necessarily be directly compatible with future versions.