RAG to customize prompts according to the input text

[wbcbugfree]

Currently, few-shot learning prompts contain only fixed examples and thus may not benefit from those examples when processing text that is significantly different from the examples in the prompts. Therefore, introducing Retrieval-Augmented Generation (RAG) enables the customization of prompts according to the input text, i.e., the selection of ground-truth text-RDF pairs that are semantically closest to the input text. Obviously, RAG relies on a manually annotated and/or censored dataset containing ground-truth text-RDF pairs.

[robknapen]

Perhaps something else to consider is agentic RAG that can self-correct based on evaluation results.