Reproduce “AOCI: Symbolic-Semantic Indexing for Practical Repository-Scale Code Understanding with LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinshi Liu, Hanying Zuo, Congyin Cao, Anran Zhang, Yixuan Liu, Xinzhou Xie
AOCI introduces a symbolic-semantic indexing protocol and platform to enable large language models to understand repository-scale codebases by providing a structured "file-level intent layer." This approach achieved 97.67% accuracy in file localization and prevented all defects in real-world industrial tasks while significantly reducing token consumption by 4x to 130x compared to agent-based tools.
Key insights
- LLMs demonstrate improved code understanding and reduced defects when provided with a structured, systematically indexed architectural blueprint rather than raw code or dynamic exploration.
- Explicitly encoding a "file-level intent layer," including architectural coordinates and semantic summaries of files' roles and relations, is crucial for LLMs to build a comprehensive system view.
- A stable, precomputed global context layer prevents common classes of defects, such as multi-path query omissions and schema incompatibilities, and drastically reduces token consumption compared to agent-based dynamic context acquisition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.