You are building a RAG-based customer support assistant on Databricks. Compliance requires that every answer include inline citations pointing to the specific source document IDs used to generate the response. The retriever already returns each chunk's document ID in its metadata, and these IDs are injected into the prompt context. The model currently produces fluent answers but frequently omits citations or invents document IDs that were not in the context. Without changing the retriever or fine-tuning the model, what is the most effective way to make the assistant reliably cite the correct source IDs?