Retrieval-Augmented Generation Assistant for Anatomical Pathology Laboratories

A team of researchers has published a study proposing a specialized AI assistant to solve a critical problem in healthcare: outdated and fragmented laboratory documentation. In Anatomical Pathology (AP), where up to 70% of medical decisions rely on lab results, technicians often struggle with static PDFs and printed manuals, leading to workflow errors and diagnostic delays. The researchers' solution is a Retrieval-Augmented Generation (RAG) system tailored for AP labs, designed to provide context-grounded, accurate answers to protocol queries by dynamically searching through a curated knowledge base.

The study meticulously evaluated the system using a novel corpus of 99 real AP protocols from a Portuguese institution and 323 constructed question-answer pairs. Through ten experiments, they tested various chunking strategies, retrieval methods, and embedding models, assessing performance with the RAGAS framework. Key findings show that a combination of recursive chunking and hybrid retrieval formed the strongest baseline. Crucially, integrating the domain-specific MedEmbed embedding model significantly boosted performance, achieving scores of 0.74 for answer relevance, 0.70 for faithfulness, and 0.77 for context recall. The analysis also revealed that retrieving just the single top-ranked document chunk (k=1) was optimal for both accuracy and efficiency, reflecting the modular nature of lab protocols. This research provides a blueprint for deploying reliable RAG systems in high-stakes medical environments, demonstrating their potential to turn cumbersome documentation into an interactive knowledge assistant that supports technicians and enhances patient safety.