Key Info
Perplexity built a new training method for contextual embedding models that encode each chunk with the whole document in view, and released pplx-embed-v2-context-9b-preview as a Hugging Face preview.
Highlights
- Sets a new state of the art on ConTEB and turbopuffer's privately held context-bench
- Highest average nDCG@10 on ConTEB among tested models, though not on every task
- Beats voyage-context-4 on chunk retrieval with 8x less storage per vector: 1 KB (1024 dims, int8) vs 8 KB (2048, float32)
- Available to view on Hugging Face
