Retrieval-Augmented Classification

Retrieval-augmented classification studies how models can use stored examples, prototypes, label definitions, or domain rules when assigning categories. Common approaches include nearest-neighbor and prototype methods, retrieval-augmented language models, and cascades that retrieve candidates before a classifier verifies them; some also adapt thresholds or search hierarchies. These methods can help with rare classes, large label sets, changing data, and settings where predictions must stay within an approved taxonomy, while reducing the need to retrain models. A central challenge is retrieving evidence that is both relevant and discriminative: similarity alone may not resolve confusable labels, and systems must balance accuracy with context limits, cost, robustness, and reliable abstention.