RECAST Beats RAG by 16 Points

🔥 What happened Google researchers dropped RECAST: a router model that doesn't just retrieve evidence — it computes it. A 9B RouterLM plans sequential filter, aggregation, and code operations instead of relying on fixed similarity search. 💡 Why it matters Classic RAG fails when the answer isn't in any single source but must be derived. RECAST hits 75.6% success, beating the strongest baseline by 15.9 points — with only 9B parameters. ⚡ Our take RAG is yesterday's news. Anyone still betting on vector search in 2026 will lose to routers that compute instead of guess.