Perplexity breaks the single-vector bottleneck

🔥 What happened Perplexity dropped pplx-embed-v2-late: late-interaction models that keep a vector per token instead of crushing everything into one. They're the first to combine multi-vector retrieval, text-plus-image search, and a shared embedding space across 0.6B and 9B sizes. 💡 Why it matters The 0.6B model matches models with 5x the active parameters on ViDoRe V3. It searches rendered PDF pages directly — no OCR, no parsing. And the small model can query an index built by the big one, pairing cheap queries with high-quality docs. ⚡ Our take One vector per document was always a lossy hack. Perplexity just killed it — and humiliated bigger rivals with a 0.6B model.