When Is Retrieval Enough? A Survey of Stopping Decisions in RAG

Retrieval-Augmented Generation (RAG) is increasingly moving from fixed-depth pipelines toward systems that decide whether additional evidence acquisition is warranted. Yet methods that emit the same STOP/CONTINUE action often judge different quantities, including evidence coverage, answer reliability or stability, unresolved information gaps, and continuation value.

This focused review analyzes those decision semantics, the supervision used to obtain them, the continuation actions available after an insufficient state, and the terminal distinction between stopping retrieval and releasing an answer. We organize evaluation around four failures: premature stopping, unnecessary continuation, unsupported answering, and unnecessary abstention. The literature shows that stopping quality is policy-relative: whether another retrieval would help depends on the retriever, action space, corpus, budget, and answerer.

We conclude with directions for marginal-utility supervision, counterfactual trajectory evaluation, whole-loop cost accounting, and transfer to richer multimodal, graph structured, and tool-augmented evidence settings.