You want a model to answer questions about a company policy document. What is the most reliable way to make sure it works from the real policy?

The principle — A model cannot consult anything you have not given it. Private or recent material has to travel into the prompt.

Why the key is correct — Google's advice is to include the information the model needs rather than assume it has it, and its own worked example does exactly this: a generic troubleshooting answer becomes a correct one once the relevant text is pasted in and the model is told to answer using only that text. The scoping instruction is what stops it drifting back to generic knowledge.

Why the others are wrong — Naming a document does not retrieve it. Describing it substitutes your summary for the source. Background about your role is not the material the question is about.

Remember this — If it matters, paste it. Then bound the answer to it.

Sources — Google, Prompt design strategies.