Single agent vs multi-agent: check the context before you split

After hybrid search and a reranker, my RAG system, GroundedDocs, still got [N] of 40 eval questions wrong or only partly right. The next step looked obvious: a planner agent, a retrieval agent, a veri