什么是 Claude Sonnet 4.5？

Databricks' Instructed-Retriever-1 cuts search latency by 3x and TTFT to ~2s via parallel test-time scaling without quality loss. The unified model handles query generation and reranking in parallel using multi-pivot groupwise reranking, achieving Pareto-optimal recall-precision tradeoffs for enterprise RAG systems.

November 2025 was a critical inflection point for LLM development, with model performance changing hands five times among three major vendors in six months, coding agents achieving qualitative leaps to daily usability, and emerging tools like Warelay beginning to appear.