Hierarchical Repository-Level Code Summarization with Quantized Small Language Model and Graph-Guided Context

Abstract

Despite recent progress in function-level docstring generation, code summarization systems have mostly focused on creating concise descriptions for individual functions using automatically generated docstrings, leaving the difficult task of generating structured repository-level summaries unaddressed. We propose a novel hierarchy-driven architecture that uses a compact and quantized language model with structural context provided by a graph. In this approach, a hierarchy graph is created to capture control flow, data flow, and inter-file relationships, which are then combined with identifier masking and structured encoding to create function and repository-level summaries. Moreover, we propose a novel reference-less evaluation scheme to evaluate the quality of summaries against source code based on semantic, coverage, and structural aspects. Our empirical analysis shows clear advantages over a naive approach in all metrics, with the largest improvements found in flow coverage and entity fidelity, likely due to improved context rather than model size.