Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction

Retrieval-augmented generation (RAG) powers many AI applications, but a fundamental challenge remains: how to know if the retrieved context actually supports the generated answer. A new paper from researchers Florian Geissler, Francesco Carella, Laura Fieback, and Jakob Spiegelberg introduces a two-stage framework that adds statistical rigor and factual confidence measures to RAG pipelines. The first stage applies conformal prediction to select only those retrieved chunks that are likely from the correct source. This filtering alone improves answer quality by up to 6% on certain datasets. However, the authors note that standard statistical guarantees do not always hold—they depend on the exchangeability assumption of the retriever setup. To address this, they provide diagnostic metrics to check whether a given setup is suitable for conformal prediction.

The second stage quantifies the confidence that the final generated answer is consistent with the retrieved context. Using an attention-based factuality classifier, the approach can detect inconsistent answers with a success rate of up to 77%. This two-stage pipeline does not just flag potential errors—it enables a new class of certified RAG systems that come with measurable reliability guarantees. For industry applications in legal, medical, or financial domains where factual accuracy is critical, this work provides a practical path toward dependable AI that knows when it doesn't know.