A library RAG assistant must reduce hallucination risk on catalog-policy answers. What grounding discipline should they enforce?

A library RAG assistant should ground answers only on retrieved catalog-policy context from SQL to reduce hallucination risk on policy questions. Grounding discipline means the architecture retrieves catalog-policy chunks and the prompt constrains the model to those municipal sources rather than inventing circulation or holds rules. Allowing answers from unconstrained model invention with no retrieved catalog context fails by concept because that is the hallucination-prone pattern RAG exists to avoid for public-facing library guidance. Grounding answers on unrelated weather feeds instead of catalog policies fails because off-domain context cannot authorize catalog-policy claims and can steer the assistant into irrelevant advice. Disabling retrieval and banning any policy text from reaching the model fails by removing the grounding source entirely so every answer becomes unanchored. Exam caveat: grounding reduces but does not eliminate hallucination—prompts should still require refusal when retrieved context is insufficient, and UI citations should point back to chunk IDs from SQL. Operationally, require retrieval before generation, insert catalog-policy chunks into the prompt, refuse or escalate when similarity scores are weak, and periodically red-team the assistant with questions outside the retrieved policy set.