RAG-Based Dynamic Prompting for Biomedical NER

Prompt Structure

The full dataset-specific prompts are provided in supplementary materials (Tables 8–12) of the paper. The base prompt begins with:

You are a medical AI ...
[Dynamically retrieved annotated examples inserted here via RAG]
Input: ['token1', 'token2', ...]
Output: ['token1-TAG', 'token2-TAG', ...]

Input/Output Format

The paper uses two equivalent I/O formats. Example from the paper:

Token-level format:

Input: ['I', 'was', 'a', 'codeine', 'addict.']
Output: ['I-O', 'was-O', 'a-O', 'codeine-B-Clinical_Impacts', 'addict.I-Clinical_Impacts']

Sentence-level format:

Input: ['I was a codeine addict.']
Output: ['I-O', 'was-O', 'a-O', 'codeine-B-Clinical_Impacts', 'addict.I-Clinical_Impacts']

Usage Notes

This prompt is from the paper “Retrieval augmented generation based dynamic prompting for few-shot biomedical named entity recognition using large language models” (Ge et al., 2025).

• Approach: Uses RAG to dynamically select the most relevant annotated examples as in-context learning demonstrations for each input, rather than using a fixed set of examples.
• Retrieval methods evaluated: TF-IDF, SBERT, ColBERT, and DPR for selecting contextually relevant examples.
• Key innovation: The prompt is not static — for each input text, similar annotated examples are retrieved and inserted as few-shot demonstrations.
• Full prompts: Dataset-specific prompts are available in supplementary materials (Tables 8–12) of the paper.