Related Experiment Video
Updated: Jun 20, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
LabSage: Structural-Semantic Decoupling for Enhanced Retrieval-Augmented Generation in Clinical Laboratories.
Hang Zhang1, Yuelyu Ji2, Chenyu Li3
1Intelligent Systems Program, University of Pittsburgh, Pittsburgh, PA, USA.
Summary
LabSage improves clinical laboratory AI by decoupling retrieval and reasoning, enhancing accuracy and compliance in regulated environments. This approach addresses context fragmentation in standard methods.
Area of Science:
- Clinical Laboratory Science
- Artificial Intelligence in Medicine
- Regulatory Compliance
Background:
- Clinical laboratories require strict adherence to Standard Operating Procedures (SOPs) within a regulated environment.
- Standard Retrieval-Augmented Generation (RAG) methods struggle with context fragmentation, disrupting procedural dependencies in this domain.
Purpose of the Study:
- To introduce LabSage, a novel domain-adapted framework designed to overcome the limitations of standard RAG in clinical laboratory settings.
- To improve the accuracy and completeness of information retrieval for laboratory procedures.
Main Methods:
- Developed LabSage, a framework employing structural-semantic decoupling with a hierarchical architecture.
- Indexed compact search units for precision and dynamically retrieved expanded context units for completeness.
- Utilized Qwen-2.5-7B as the backbone for evaluation.
Main Results:
- LabSage achieved an Answer Accuracy of 0.780 and Context Recall of 0.909.
- Outperformed standard RAG by 8.3% in accuracy and 5.7% in recall.
- Demonstrated mitigation of safety-critical omissions and ensured compliance.
Conclusions:
- Decoupling vector search from reasoning context is crucial for regulated medical domains.
- LabSage represents a significant architectural advancement for AI in clinical laboratories.
- The framework enhances AI reliability and adherence to regulatory standards.