Recoleta
GitHub
Source note

BLAgent: Agentic RAG for File-Level Bug Localization

Authors
Md Afif Al Mamun, Gias Uddin
Source
arXiv
Published
Collection
Software Intelligence
Bug LocalizationAgentic RAGCode IntelligenceAutomated Program RepairSoftware Agents
Open arXivMarkdown source

Summary

BLAgent is an agentic RAG pipeline for file-level bug localization in code repositories. It ranks likely faulty files from a bug report, then improves downstream automated repair by giving the repair system better file context.

Problem

  • File-level bug localization is a bottleneck for debugging, triage, root-cause analysis, and automated program repair because later steps fail when the wrong files are selected.
  • SWE-bench repositories average over 11,000 functions and 168,000 statements, so direct statement-level search is too expensive for LLM-based repair systems.
  • Prior RAG methods often use static retrieval over weak code chunks, while graph-based agent methods can be costly or require extra model training.

Approach

  • BLAgent indexes source files with AST-aware chunks, so chunks align with functions, classes, or other syntactic units instead of arbitrary text spans.
  • It prepends each chunk with the relative file path, which helps match bug reports that mention modules, tracebacks, or package-qualified names.
  • It rewrites each bug report into two retrieval queries: a structural query for identifiers and modules, and a behavioral query for observed versus expected behavior.
  • It retrieves candidate files from both queries, keeps up to 15 candidates, and ranks files by the best matching chunk in each file.
  • It reranks candidates in two phases: a ReAct agent scores file skeletons, then a one-shot LLM compares pruned file contexts for the top 5 files using the most relevant retrieved chunks.

Results

  • On SWE-bench Lite, BLAgent reaches 86.7% Top-1 accuracy and 0.900 MRR with a closed-source model.
  • With an open-source model, it reaches 78.6% Top-1 accuracy and 0.851 MRR on SWE-bench Lite.
  • The paper claims BLAgent beats LocAgent while using the same model at over 18× lower API cost.
  • When integrated into Agentless, an open-source APR system, BLAgent improves issue resolution by over 20%.
  • The paper reports that removing file-level localization from a multi-granularity localization pipeline caused a 94% drop in Top-5 accuracy and a 96% drop in statement-level MAP, which motivates the file-level focus.

Link