Recoleta
GitHub
Source note

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

Authors
Jiawei He, Weisong Sun, Mengyu Shi, Jie Jia, Tong Bian, Xikai Yang, et al.
Source
arXiv
Published
Collection
Software Intelligence
Code IntelligenceRepository UnderstandingSoftware AgentsContext RetrievalAutomated Software Engineering
Open arXivMarkdown source

Summary

DeepDiscovery recovers task-level context in large codebases by finding reliable task anchors, then expanding along code, configuration, test, and organization links. The paper claims better relevant-file recovery in industrial repositories and a higher solve rate on SWE-bench Verified.

Problem

  • LLM coding systems often retrieve local code snippets, while repository tasks need connected files across interfaces, business logic, configuration, tests, and dependencies.
  • Large industrial repositories change often, so pre-built vector indexes or static graphs can go stale and add maintenance cost.
  • Missing bridge files, such as registration, dependency-injection, or test-to-implementation files, can reduce automated coding success.

Approach

  • DeepDiscovery uses a two-stage Location-Inference workflow: Location finds a small set of high-confidence task anchors, and Inference expands from those anchors to recover a broader implementation path.
  • The Location stage scores candidate files using semantic and lexical match, compressed structural summaries, rule-template matches, and task-conditioned artifact-role priors.
  • The Inference stage expands over a multi-relation repository graph covering explicit dependencies, implicit links such as configuration-to-code bindings, and organizational links such as folder and module proximity.
  • A budget-aware priority score chooses expansion actions by estimated gain divided by cost, with default settings of 8 anchors, a 0.15 expansion stopping threshold, and a 0.62 full-text promotion threshold.
  • The final context is metadata-first: most selected entities keep compact metadata, while likely modified or high-value files get full text.

Results

  • On 27 medium-scale tasks with 135 manually annotated gold relevant files, DeepDiscovery reports the best file-recovery quality among five baselines: DeepWiki, CodeWiki, RAG, GraphRAG, and AST+GraphRAG.
  • The industrial repository setting covers 2.67 million lines of code and more than 25,000 files.
  • In organization-internal industrial evaluations, DeepDiscovery improves Full Recall Rate by 2.5 to 7.4 percentage points on 27 medium-scale tasks across multiple AI coding systems.
  • On 40 large-subproject industrial tasks, it improves Full Recall Rate by 1.6 to 9.2 percentage points across multiple AI coding systems.
  • On SWE-bench Verified, the system with DeepDiscovery reaches a 78.6% Solve Rate, 8.2 percentage points above the corresponding baseline.

Link