How do we evaluate RAG accuracy without labelled data?

We're building a document QA system over internal financial reports. We don't have labelled question-answer pairs and building a ground-truth dataset would take months. How are teams evaluating retrieval quality and answer correctness in practice without ground truth?