Prompting with retrieved context

Structure beats concatenation

Passages pasted into one block become indistinguishable. Wrapping each in a delimiter with its source id lets the model attribute claims and lets you check that it did.

State the grounding rule explicitly: answer only from the provided context, and say when the context does not contain the answer. Without that instruction the model falls back on training knowledge and you cannot tell which sentence came from where.

Position matters. Models attend most reliably to the beginning and end of a long context, so put the strongest passages at the edges rather than burying them in the middle.