How I implemented a unified search architecture and embedded LLM intelligence at key decision points to improve intent parsing and RAG orchestration for my Copilot Studio agent