<p>NER performance gains from RAG relative to in-context learning. A -shot model refers to one prompted with labeled learning examples selected using the Inquiry-KNN method. The “+ -RAG” condition additionally prepends knowledge snippets retrieved from the RAG corpus. Only metrics showing performance improvement with RAG snippets are reported. A dash (-) indicates no observed benefit. The best-performing configuration for each task is underscored. The 95% confidence intervals constructed by bootstrap are shown in parentheses.</p>

NER performance gains from RAG relative to in-context learning. A -shot model refers to one prompted with labeled learning examples selected using the Inquiry-KNN method. The “+ -RAG” condition additionally prepends knowledge snippets retrieved from the RAG corpus. Only metrics showing performance improvement with RAG snippets are reported. A dash (-) indicates no observed benefit. The best-performing configuration for each task is underscored. The 95% confidence intervals constructed by bootstrap are shown in parentheses.