Answer With Citations, or Say You Can't

Answer a question from a set of documents, cite the documents you used, and say plainly when they do not cover it.

Implement answer_with_citations(llm, question, docs) returning {"answer": str, "citations": [int, ...], "supported": bool}.

Number the documents from 0 and ask the model for JSON: the answer, the indices it relied on, and whether the documents support the answer. Then enforce what the model cannot be trusted to:

• Citations must be valid indices. The model will cite document 9 of 4. Drop anything out of range.
• Supported requires a citation. If the model says supported: true and cites nothing, set it false and make the answer say the documents do not cover it.
• When it is not supported, say so. "I can't answer that from these documents" is a correct output.

The catch: asked for the CEO's home address, a model given documents about refunds and parking will often invent one. With nothing to cite, the citation rule stops it being marked supported.

Tests use a real model: a refund question must cite the refund document and contain the number; an off-topic question must come back unsupported with no citations.