# cAST: Enhancing Code Retrieval-Augmented Generation with Structural Chunking via Abstract Syntax Tree

**Authors:** Yilin Zhang, Xinran Zhao, Zora Zhiruo Wang, Chenyang Yang, Jiayi Wei, Tongshuang Wu  
**arXiv:** [2506.15655](https://arxiv.org/abs/2506.15655)  
**Date:** 2025-06-18  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2506.15655v1/

## Key points

- Retrieval-Augmented Generation (RAG) enhances large-scale code generation tasks by grounding predictions in external code corpora
- Traditional line-based chunking methods disrupt semantic structures, leading to a degradation in the quality of generated code
- Chunking via Abstract Syntax Trees (CAST) breaks down large AST nodes into smaller, more manageable chunks and consolidates sibling nodes while adhering to predefined size limits
- CAST method generates self-contained and semantically coherent units across various programming languages and tasks, enhancing performance on diverse code generation tasks
- CAST has been shown to boost Recall@5 by 4.3 points on RepoEval retrieval and Pass@1 by 2.67 points on SWE-bench generation
- Structure-aware chunking is crucial for scaling up retrieval-enhanced code intelligence
- Design goals for CAST include maintaining syntactic integrity, maximizing information density within each chunk, ensuring language invariance, and enabling seamless integration within existing RAG pipelines
- AST Parsing supports syntax-aware chunking, accurately identifying and extracting meaningful chunks from source code while preserving its structural formatting

## Summary

, , , , 

In the realm of code generation, Retrieval-Augmented Generation (RAG) has emerged as a crucial technique for enhancing large-scale code generation tasks by grounding predictions in external code corpora to improve accuracy. However, a key yet often overlooked aspect of RAG pipelines is the process of chunking, which involves dividing documents into retrievable units. Traditional line-based chunking methods frequently disrupt semantic structures by splitting functions or merging unrelated code segments, leading to a degradation in the quality of generated code. To address this issue, we introduce chunking via Abstract Syntax Trees (CAST), a structure-aware approach that breaks down large AST nodes into smaller, more manageable chunks and consolidates sibling nodes while adhering to predefined size limits. This method generates self-contained and semantically coherent units across various programming languages and tasks, ultimately enhancing performance on diverse code generation tasks. For instance, our approach has been shown to boost Recall@5 by 4.3 points on RepoEval retrieval and Pass@1 by 2.67 points on SWE-bench generation. Our work underscores the significance of structure-aware chunking in scaling up retrieval-enhanced code intelligence. By focusing on the initial stage of the RAG pipeline - chunking - we aim to parse source code into meaningful units such as functions or classes while preserving the overall structure of the code. These units are then grouped into coherent chunks that serve as retrievable contexts for subsequent retrieval processes and language model prompts. The design goals for CAST revolve around four key principles: maintaining syntactic integrity by aligning chunk boundaries with complete syntactic units whenever possible; maximizing information density within each chunk up to a specified size limit; ensuring language invariance so that the algorithm can be applied uniformly across different programming languages and tasks; and enabling seamless integration within existing RAG pipelines through plug-and-play compatibility. AST Parsing plays a crucial role in supporting syntax-aware chunking, allowing us to accurately identify and extract meaningful chunks from source code while preserving its structural formatting. Our experiments demonstrate promising results, showcasing how leveraging AST-based chunking can significantly improve retrieval-enhanced code intelligence by enhancing semantic coherence and overall performance across various coding tasks.