When RAG Users Ask Vague Questions: Clarify Once, Learn the Default

When a RAG system receives vague or ambiguous user questions, silently guessing a missing field is risky. This post presents a production pattern: emit a single structured clarification request to the user, then store the answer as a learned default so future identical questions are answered silently. Two Pydantic schemas handle the contract — ClarificationRequest for asking and ClarificationDefault for learning. A confidence-gated loop (ask below 0.6, apply silently above 0.85, occasionally refresh in between) drives the decision. A worked broker insurance example walks through three cases: first-time ask, silent default application, and failure-triggered stratification. The pattern integrates with the broader enterprise RAG pipeline via the query log and ontology tables, with full auditability.