Late-Interaction Retrieval With Perplexity's New Embeddings

On October 7, 2026 Perplexity released pplx-embed-v2-late, an open-weight retriever in 0.6B and 9B sizes. It is not a normal embedding model. Instead of squeezing a document into one vector, it keeps one 128-dimension vector per token and scores with MaxSim, the ColBERT recipe. It also reads page images, so a screenshot of a chart and a paragraph of text live in the same index.

Why care now? Single-vector search is where most RAG pipelines quietly lose answers: a long passage gets averaged into one point, and a short, specific query like an error code or a plan name drowns in the average. Late interaction keeps the detail. The new models are MIT-licensed on the 0.6B card, run through plain Sentence Transformers, and the two sizes share one embedding space, so you can index with the big model and query with the small one.