Do Not Treat Code as Natural Language: Implications for Repository-Level Code Generation and Beyond

Repository-Level Code Generation • Structure-Aware Indexing • Dependency-Aware Retrieval

Quick Start • Documentation • Paper

Introduction

Large language models for code (CodeLLMs) have demonstrated remarkable success in standalone code completion and generation, yet their effectiveness diminishes in repository-level settings where cross-file dependencies and structural context are essential. Existing Retrieval-Augmented Generation (RAG) approaches often borrow strategies from NLP, relying on chunking-based indexing and similarity-based retrieval that overlook structural relationships and miss functionally relevant dependencies.

We present Hydra, a repository-level code generation framework that treats code as structured code rather than natural language. Our approach introduces: (i) structure-aware indexing that preserves code structure and dependencies, (ii) a lightweight dependency-aware retriever (DAR) that identifies true dependencies, and (iii) hybrid retrieval combining dependency-aware and similarity-based methods.

Extensive experiments on DevEval and RepoExec benchmarks show that Hydra achieves state-of-the-art performance, surpassing the strongest baseline by over 5% in Pass@1 and enabling smaller models to match larger ones.

Quick Start

Prerequisites

Create a new conda environment and install dependencies:

# Create conda environment
conda create -n hydra python=3.10.0
conda activate hydra

# Install required packages
pip install -r requirements.txt

Setup and Installation

Important: You must complete the following setup steps before running any experiments.

1. Extract benchmark data:

   cd data
   unzip temp.zip
   
   # Extract RepoExec benchmark
   cd ../benchmark/RepoExec
   unzip test-apps.zip
   
   # Extract DevEval benchmark
   cd ../DevEval
   tar -xzf data.tar.gz
   wget https://huggingface.co/datasets/LJ0815/DevEval/resolve/main/Source_Code.tar.gz
   tar -xvzf Source_Code.tar.gz
   

2. Prepare structured context (required for experiments):

   # For RepoExec benchmark
   bash src/context_formulation/structured_indexer/run.sh --dataset RepoExec
   
   # For DevEval benchmark  
   bash src/context_formulation/structured_indexer/run.sh --dataset DevEval
   

Main Results

Comparison of Hydra with prior retrieval-based approaches and no-context baselines. Results are reported in Pass@1/3/5.

GPT-4.1 mini

QwenCoder-1.5B-Instruct

QwenCoder-7B-Instruct

Documentation

Important Before reproducing experiments, you must first train the DAR (Dependency-Aware Retriever) model or using our pretrained model at huggingface DAR model.

For detailed instructions and comprehensive guides, please refer to:

• Training.md - DAR (Dependency-Aware Retriever) training guide including:

  • Dataset construction methodology
  • Model architecture and training procedures

• Reproduce.md - Complete experimental reproduction guide including:

  • Benchmark setup and data preparation
  • Research questions reproduction (RQ1-RQ4)
  • Code generation pipeline
  • Evaluation and metrics calculation

Citation

If you found this repository to be useful, please cite:

@misc{leanh2026treatcodenaturallanguage,
      title={Do Not Treat Code as Natural Language: Implications for Repository-Level Code Generation and Beyond}, 
      author={Minh Le-Anh and Huyen Nguyen and Khanh An Tran and Nam Le Hai and Linh Ngo Van and Nghi D. Q. Bui and Bach Le},
      year={2026},
      eprint={2602.11671},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2602.11671}, 
}