Perplexity releases pplx-embed-v2 embedding models trained on distilled chunk relevance scores
Article timeline
The articles in this story. Times are ET.
Perplexity DevelopersOfficialPerplexity's text-only 9B embedding model beats leading retrievers on visual documentsAIPerplexity says its 9B embedding model, trained only on text, outperforms the leading dense retrievers on visual documents. Using two 1024-dimensional embeddings per page, it beats Nemotron-3.5-Embed-8B on ViDoRe-V3 with a 4x smaller index.
- Perplexity DevelopersOfficialPerplexity's embedding models rank top on ViDoRe-V3 markdown retrieval
AIPerplexity says its 0.6B embedding model is the strongest sub-1B model on ViDoRe-V3 markdown, whether chunked or unchunked. Its contextual 9B model with 128-token chunks ranks second only to voyage-context-4. Both models also support a non-contextual mode that creates one embedding per document without chunking.
- Perplexity DevelopersOfficialPerplexity trains embedding models on distilled chunk-level relevance scores
AIPerplexity says its contextual embedding models are trained differently from other contextual embeddings models. Instead of one gold chunk per query, chunk-level relevance scores are distilled from a query-aware context compression model in its answer pipeline, and supporting chunks are no longer treated as negatives.
- Perplexity DevelopersOfficialPerplexity releases open-weight pplx-embed-v2-context embedding models
AIPerplexity releases pplx-embed-v2-context, open-weight vision-language embedding models in 9B and 0.6B sizes. Both return one int8 vector per document, or one per chunk that sees the whole document.
