Perplexity's Embedding Model Beats Voyage 8x Smaller

Perplexity Research and the vector database company turbopuffer have put out pplx-embed-v2-context-9b-preview, a contextual embedding model built for retrieval-augmented generation pipelines. It's live on Hugging Face under the MIT license, loadable with transformers 5.4.0 or later and trust_remote_code=True. It hasn't landed on the Perplexity API yet, and the model card warns that both the weights and the interface could change without keeping backward compatibility, so this is a preview in the literal sense.

The design answers a specific failure mode in how RAG systems chop documents into chunks. A chunk can depend on something stated pages earlier, a defined term, a heading, an entity introduced in a different section. Contextual embedding models have tried to fix this with late chunking, encoding the whole document first and pooling per chunk afterward.

But the training behind most of these models still picks a single "gold" chunk per query and treats everything else as a negative, even the supporting sentences a user would need to check the answer. Perplexity's team names three more cracks in that setup: binary labels are too coarse, LLM annotation costs scale badly, and labels get locked to one chunking scheme.