Perplexity kills the gold-chunk dogma

🔥 What happened Perplexity Research and turbopuffer dropped pplx-embed-v2-context-9b-preview, a self-hostable contextual embedding model for RAG. Instead of training on one gold chunk per query, it distills soft targets from a teacher that scores every token. Weights are on Hugging Face under MIT. 💡 Why it matters The old recipe labels one chunk positive and dumps every other chunk — including the evidence sentences — into the negatives. This model beats voyage-context-4 by 14.4 points on answer recall at K=10. And 1024-dim int8 (1 KB) edges out 2048-dim float32 (8 KB) — 8x less storage. ⚡ Our take This is the first real strike against chunking orthodoxy. Anyone still training on one-hot gold labels in 2025 is burning money and recall.