nnRAG – Incorporating Human-Feedback for Neural Network Driven Similarity Search: A Preliminary Feasibility Study

Medical data presents unique challenges for large language models (LLMs), especially because any inaccurate or incomplete information may pose a risk of serious harm to the patient. While Retrieval-Augmented Generation (RAG) architectures offer promising solutions for improving the context and accuracy of LLM-generated responses, it is crucial that the retrievals are highly precise. Due to the potential consequences of errors, Human-in-the-Loop (HITL) processes are essential to ensure reliability and relevance in medical data retrieval. In this work, we propose a first-of-its-kind approach for retrieval workflows that incorporates Human Feedback for a Neural Network-Driven Similarity Search within the RAG architecture (nnRAG). This novel approach stores and integrates user feedback-based reward scores directly into the retrieval process to make it adaptive and refine the accuracy of context selection. We conducted a preliminary study to verify the functionality of this approach, initially evaluating it on a non-medical dataset and subsequently testing its feasibility with a medical dataset. The study demonstrated the ability of the nnRAG system to dynamically update its selection from non- or less-relevant sections to highly precise sections, post-training, on both non-medical and medical datasets. The system was able to perform this precise retrieval, even when the top-k was kept to 1, after a single iteration of training. Concluding, this approach offers a unique solution by introducing dynamism into the retrieval workflows, leveraging user feedback and potentially ensuring improved relevance and utility in the context of diverse and evolving user needs.