Perplexity and turbopuffer release pplx-embed-v2-context-9b-preview embedding model

The model improves retrieval‑augmented generation by training a token‑level teacher that scores every token for a query‑document pair, producing soft targets instead of a single “gold” chunk. It encodes an entire document in one pass, then pools per chunk, allowing retrieval of answers together with the evidence needed to verify them. Benchmarks on the private context‑bench (2,099 queries, 38,894 documents, 2,458,072 chunks) show 45.5% answer recall and 40.6% evidence recall at K = 10, a 14.4‑point and 5.0‑point lead over the competing voyage‑context‑4 model respectively. The model stores 1024‑dim int8 vectors (≈1 KB each) and supports 2048‑dim float32 vectors.