What: On June 18, 2026 , researchers posted CacheWeaver — a prompt-layer method built on prefix-cache-aware evidence reordering . It changes only the order retrieved RAG chunks appear in the prompt, so the serving engine can reuse more of its KV prefix cache . Why: In retrieval-augmented serving, time-to-first-token is dominated by prefilling the evidence ; reusing a cached prefix skips that work, and CacheWeaver squeezes out the reuse a naive system leaves on the table — with no ...