Segment-Level Grounding: Retrieval Triggered by an Uncertainty Signal

Retrieval-augmented generation (RAG) reduces hallucinations by grounding generation in external evidence, but standard RAG retrieves for every query regardless of necessity. Self-RAG addresses this by training the model to retrieve adaptively through reflection tokens. We propose a complementary approach: segment the generated response, detect uncertainty at the segment level, and retrieve only for the marked segments. Confirmed evidence is injected into the marked segment, after which confidence is recomputed. The approach works with black-box outputs and requires no fine-tuning. This is a position paper: we define segmentation levels, uncertainty-detector features, the injection format, axioms for confidence recomputation, and an evaluation protocol on HaluEval and TruthfulQA with a baseline that isolates the contribution of the trigger.