Perplexity previews contextual embedding model for evidence-aware retrieval
Perplexity Research and turbopuffer have released pplx-embed-v2-context-9b-preview, a contextual embedding model designed for retrieval-augmented generation pipelines. It embeds each document chunk while considering the full document, aiming to retrieve an answer together with the material needed to verify it. The model is available as a self-hosted preview, with weights on Hugging Face under the MIT license. It is not currently offered through the Perplexity API, and its weights and interface may change without backward compatibility. Loading requires transformers>=5.4.0 and trust_remote_code=True.
The training approach replaces the usual single “gold” passage with softer relevance targets. A query-aware context compression model scores tokens across the document; the scores are pooled within chunks and converted into a distribution. The embedding model learns this distribution through forward KL divergence, while a document-level InfoNCE objective scores a document by its strongest chunk. Random chunking strategies are used during training, allowing boundaries to change without new annotations.
The model builds on an in-house 9B ColBERT retrieval model and produces 2048-dimensional embeddings, with Matryoshka training supporting 1024 dimensions. Quantization-aware training enables native int8 embeddings. Training covered roughly 430 datasets and more than 50 languages, excluding ConTEB data.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.