From Fragments to Paths: Task-Level Context Recovery for Large Industrial Codebases

DeepDiscovery is a new task-level repository-understanding method designed for large industrial codebases, utilizing a two-stage Location-Inference framework to enhance context recovery for complex software engineering tasks. It demonstrates superior performance in file recovery and downstream software engineering tasks, achieving a 78.6% Solve Rate in controlled evaluations and improving Full Recall Rate by up to 9.2 percentage points on large subprojects compared to existing baselines. This advancement is significant for practitioners as it enables more effective coding agents in navigating and understanding extensive codebases without the need for offline preprocessing.