Rag and Llms Based Automation of Cosmic Functional Size Measurement

Functional Size Measurement (FSM) methods, such as COSMIC, provide standardized techniques for software effort estimation and efficient project management. However, manual measurement of Natural Language (NL) requirements remains time-consuming, prone to human error, and subject to human interpretation biases. Existing hybrid approach that combines Large Language Models (LLMs) with a rule-based system has demonstrated promising results in automating COSMIC measurement. Yet, its performance varies across domains, with high accuracy in business applications and lower accuracy in real-time systems due to domain-specific terminology and contextual differences. This study enhances the existing LLM-based COSMIC measurement framework by integrating a Retrieval Augmented Generation (RAG) system, which dynamically injects domain-specific COSMIC knowledge into the reasoning process. A rule-based validation layer ensures adherence to COSMIC movement rules and rectifies incorrect classifications. Experimental results show that the RAG enhanced framework achieves more balanced accuracy across both business and real-time domains, reducing domain disparity and improving overall consistency. These results validate that the combination of retrieval and reasoning mechanisms significantly improves the cross-domain robustness and measurement reliability of automated COSMIC sizing.