报告开始:2026年11月21日 15:15(Asia/Kolkata)
报告时间:15min
所在会场:[S4] Artificial Intelligence and Intelligent Computing [S4-11] Artificial Intelligence and Intelligent Computing
The reason behind the rapid decay of software documentation is its manual maintenance and disconnection from version-control lifecycle. While Single-pass Large Language Model (LLM) generators can improve the situation to certain extent, they still heavily depend on repository grounding. Without it, the LLM generators hallucinate a structure they do not have access to and publish flawed drafts of documentation without quality control. This paper presents DocuBear, a system that utilizes a hybrid retrieval-augmented generation (RAG) engine and a multi-agents pipeline with seven stages orchestrated by LangGraph to maintain the documentation updated and in line with the relevant changes in the code. The retrieval process consists of three components: dense vector search, BM25 lexical matching, and dependency graph traversal, combined with Reciprocal Rank Fusion and cross-encoder rerankings over Abstract Syntax Tree (AST) code chunks; the nature of the indexing is incremental.A formal problem model is provided, description on the end-to-end architecture and the coordination algorithm used, a comparison on the structural properties of the design with that of baseline single-pass and static-extraction methods is presented. We conclude with an account of findings such as the important limitations of the present model and ideas for future studies.
11月20日
2026
11月22日
2026
初稿截稿日期
提前注册日期
注册截止日期
发表评论