Test the result, not just the happy path
Build a small evaluation set that reflects your actual task.
Start with representative questions
Include straightforward questions, questions requiring more than one source, and examples with incomplete context.
Check factual grounding
Compare each response with the original material. Confirm that the answer does not turn an assumption into a fact.
Check behavior at the boundaries
Ask for information outside the sources, an unsupported action, and a conflicting instruction. Review how the agent handles each case.
Make a controlled change
Adjust one instruction or retrieval setting, then rerun the same questions. Keep a record of the changes that improve the result.