Recoleta
GitHub
Source note

ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation

Authors
QiHong Chen, Aaron Imani, Iftekhar Ahmed
Source
arXiv
Published
Collection
Software Intelligence
Code GenerationRepository Level GenerationProcedural RetrievalCode IntelligenceAgentic Workflow
Open arXivMarkdown source

Summary

ProjAgent improves repository-level code generation by retrieving functions with similar computational procedures, even when their names and domains differ. It combines procedural, lexical, and semantic retrieval with static-analysis-based code repair.

Problem

  • Repository-level generation must handle cross-file dependencies, project APIs, types, and coding conventions.
  • BM25, dense retrieval, and related methods can miss useful functions whose implementation steps resemble the target but use different vocabulary or serve another domain.
  • Missing or misleading context can cause hallucinated APIs, invalid code, and lower generation accuracy.

Approach

  • Decompose repository functions and the target task into logical implementation steps, such as input validation or state transformation.
  • Represent each step with an LLM reasoning-subspace projection derived from hidden states, then compare steps with cosine similarity.
  • Reduce representation anisotropy by subtracting the mean and first principal component before similarity search.
  • Use an agentic workflow to identify and validate procedurally similar functions, combine them with lexical and semantic context, and generate the target function with Qwen2.5-Coder-14B-Instruct.
  • Run compiler and static-analysis feedback over generated code and iteratively repair detected errors.

Results

  • ProjAgent reaches 41.14% Pass@1 on the REPOCOD repository-level code-generation benchmark.
  • The paper reports that this result outperforms existing retrieval-based baselines, but the provided excerpt does not include the baseline scores or the size of the improvement.
  • The method retrieves cross-domain procedural matches such as validation steps in BlackBody.evaluate and FLRW.m_nu, despite low lexical overlap and no direct call dependency.
  • PCA debiasing is reported to improve the discriminative value of procedural cosine similarity; the excerpt provides no separate ablation number.
  • The provided text contains no further quantitative results for individual retrieval components or static-analysis repair.

Link