AOCI: AI-Oriented Code Indexing — Bridging the Repository-Level Cognitive Gap Through Structured Semantic Compression

Read full text Summarise with VIRAT
Authors: Jinshi Liu, Hanying Zuo, Anran Zhang, Xinzhou Xie
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large language models excel at local code tasks but struggle with repository-level reasoning—we define this phenomenon as the Repository-Level Cognitive Gap (RLCG). This paper proposes AOCI (AI-Oriented Code Indexing), a plain-text protocol that compresses large codebases into structured indices through dual-layer encoding: discrete [ABCDE] tags for architectural positioning and continuous [F-R-A-S] elements for business semantics. On a 156K-line production system, AOCI achieves a 313:1 line compression ratio (~40:1 token compression ratio), with semantic fidelity validated through the high code quality of downstream production systems. In an 8-month longitudinal case study, two non-programmers independently built approximately 510K lines of production code across five systems serving over 27,000 registered users. To validate the practical effectiveness of the AOCI protocol, we conducted rigorous industrial-grade code quality assessments on the generated systems (across two independent technology stacks: Node.js and Go). Results show that code generated under AOCI guidance passed 1,134 automated test cases (100% pass rate), achieved 0 Bug and 0 Vulnerability ratings in SonarQube deep analysis, maintained a full-stack defect density of only 0.0007%, and passed all 12 OWASP security penetration tests. This demonstrates that approximately 600 lines of plain-text index serve as a high-fidelity cognitive contract, sufficient to support non-programmers in continuously building highly secure, highly maintainable large-scale production systems. AOCI Platform (30,018 lines of code, built from zero to deployment in 4 hours) provides automated index management.

Fields of Study & Subjects

Software Engineering Research Software Engineering Techniques and Practices Model-Driven Software Engineering Techniques

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
28 March 2026
Language
EN
Document type
Preprint