Perplexity released pplx-embed-v2-late, a family of late-interaction embedding models. The models use multi-vector representations instead of single vectors. They support text-to-text, text-to-image, and text-to-visual-document retrieval. A shared embedding space lets the 0.6B model query indexes built with the 9B model. Both models are available on Hugging Face. They achieve state-of-the-art results on vision and text retrieval benchmarks. The 0.6B model matches larger models on ViDoRe V3. Training used 186M query-document pairs across 46 languages. The models were trained without benchmark contamination. Evaluation shows strong performance on domain-specific, web-scale, and agentic retrieval tasks.
What changed
Previous embedding models used single-vector representations for text-only retrieval.
What it unlocks
Token-level matching between queries and documents for multimodal retrieval without OCR or parsed text.
- 186M query-document pairs
- 594 datasets in 46 languages
- 81.3% nDCG@10 (9B model)
- 78.0% nDCG@10 (0.6B model)
What you need to act on it
- access to Hugging Face
- use of Perplexity's embedding models
Sources