Task-Grounded and EL2N-Informed RAG Switch for Domain-Specific Information Extraction
Abstract
AbstractRetrieval-augmented generation (RAG) is increasingly popular in information extraction (IE), especially in biological and chemical domains where external knowledge is abundant. Existing RAG-enhanced IE systems, however, typically retrieve for every input, even though retrieval may be unnecessary, introduce distracting evidence, and increase inference cost. To address this issue, we propose RAG Switch, a retrieval decision framework for IE that combines difficulty-aware signals with task-grounded supervision. Specifically, we approximate model-specific training difficulty, quantified by Error L2-Norm (EL2N), from inference-time uncertainty signals, and use it to guide retrieval decisions. We then define retrieval necessity directly from instance-level extraction outcomes without retrieval, yielding task-grounded supervision for learning retrieval decision. Experiments on multiple biological and chemical IE benchmarks across diverse IE backbones show that RAG Switch consistently outperforms representative retrieval decision baselines and achieves a strong efficiency–performance trade-off in end-to-end RAG-enhanced IE pipelines.