Production RAG Hits Three Failure Modes

This reads like the real gap between demo-grade RAG and production QA: the hard problems are not answer generation, they’re retrieval control, abstention, and temporal scoping.

• –Scatter usually needs structure, not just larger `k`: hybrid retrieval, query routing, and graph or facet-based expansion tend to outperform raw vector top-k on broad comparison tasks
• –Negative knowledge needs an explicit reject path: if the retriever can’t support answerability, the system should abstain before the LLM ever gets a chance to improvise
• –Temporal questions are often two retrieval problems, not one: split pre/post retrieval, then force the generator to compare evidence across the boundary
• –GraphRAG-style indexing may help on the scatter case, but it is not a blanket fix for uncertainty or chronology
• –The production pattern that seems most robust is an answerability gate plus multiple targeted retrieval passes, not a single prompt with better instructions