Abstract
The advent of Large Language Models opened new perspectives concerning their usage within the digital health domain. However, their intrinsic probabilistic and unpredictable behavior needs the design of trustworthy strategies aiming to avoid the creation of hallucinations that, especially within the digital health domain, may lead to severe harm. Such an issue has been addressed with the adoption of Retrieval-Augmented Generation solutions, where the text generation task is supported by controlled knowledge injected into the prompts. Even if the hallucination issue is mitigated, the generation of certified information (such as trustworthy content granted by the system’s owner) requires more sophisticated strategies. In this work, we propose an approach where the classic Retrieval-Augmented Generation pipeline is enhanced with a further initial step where the Large Language Model is asked to generate a preliminary text used to query the repository of certified information for presenting the appropriate content to the final user.