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

Large language models (LLMs) often struggle with complex software engineering tasks in large industrial repositories because they rely on retrieving isolated code fragments. This paper investigates how to better recover the broader, multi-relational context—such as configuration files, dependency injection, and cross-module constraints—required to solve these tasks effectively under strict computational budgets.

The authors introduce DeepDiscovery, a two-stage framework that treats repository understanding as a task-level context recovery problem rather than a search problem.

1. Location Stage: The system uses environment-aware analysis and rule-guided scoring to identify high-confidence task anchors (entry points) within the repository.
2. Inference Stage: Starting from these anchors, the system performs multi-relational expansion, navigating explicit dependencies (e.g., imports, calls) and implicit links (e.g., event bindings, configuration-to-code mappings) to construct a coherent implementation path.

DeepDiscovery employs a metadata-first approach, loading full-text content only when necessary to stay within budget constraints. It avoids the need for expensive, frequently stale offline preprocessing by performing these operations at task-time.

DeepDiscovery consistently outperforms traditional retrieval-augmented generation (RAG) methods across multiple benchmarks. In controlled evaluations on SWE-bench Verified, a system equipped with DeepDiscovery achieved a 78.6% solve rate, an 8.2 percentage point improvement over the baseline. In production-scale industrial environments, the method improved Full Recall Rate for task-relevant files, with gains ranging from 1.6% to 9.2% on large subprojects. These results demonstrate that moving beyond fragment-level retrieval to structured, path-based context recovery significantly enhances the ability of AI coding agents to handle complex, real-world software engineering tasks.