Perplexity releases pplx-embed-v2-context-9b-preview, tops ConTEB embedding benchmark
On October 1, Perplexity released a preview contextual embedding model, pplx-embed-v2-context-9b-preview, on Hugging Face. It uses a new training approach in which every chunk of a document is encoded with visibility into the entire document, addressing the information loss that chunking causes in RAG pipelines. The model achieved the highest average nDCG@10 among tested models on ConTEB, and the team says it will keep improving performance and bring the model to the API.
Confirmed
- The model was released as a preview on Hugging Face, announced by Perplexity's official account and researcher Antoine Chaffin.
- The training method gives each chunk's encoding access to full-document context, putting the "contextual models" idea into practice to improve RAG chunking.
- On ConTEB, its average nDCG@10 was the highest among tested models, though not first on every task; comparisons were also based on turbopuffer evaluations.
- It surpasses voyage-context-4 on chunk retrieval while requiring only one-eighth the storage.
- Chaffin said the model is already strong and the release mainly validates the method; the team is improving performance and the model will come to the API.
Why it matters
- Context-aware embeddings directly target the pain point of chunking severing semantics in RAG pipelines; if adopted widely, it could change default retrieval system design.
- Leading accuracy at one-eighth the storage of a competitor has real cost implications for large-scale vector databases.
2026-10-01 ~ 2026-10-01 · 5 related posts
Primary sources
- Perplexity's pplx-embed-v2-context-9b-preview sets SOTA on ConTEB benchmark — perplexity_ai ·
- Perplexity embed model beats voyage-context-4 on chunk retrieval with 8x less storage — perplexity_ai ·
- Perplexity releases pplx-embed-v2-context-9b-preview embedding model, API version coming — antoine_chaffin ·
- [source] Perplexity's pplx-embed-v2-context-9b-preview sets SOTA on ConTEB benchmark — perplexity_ai · 2026-10-01
- [source] Perplexity embed model beats voyage-context-4 on chunk retrieval with 8x less storage — perplexity_ai · 2026-10-01
- [source] Perplexity releases pplx-embed-v2-context-9b-preview embedding model, API version coming — antoine_chaffin · 2026-10-01
- Perplexity releases contextual embedding model with new SOTA training approach — antoine_chaffin · 2026-10-01
1 near-duplicate retellings: perplexity_ai