RelRepair: : Retrieving Relevant Information to Enhance Automated Program Repair

RelRepair is a novel Retrieval-Augmented Generation (RAG) framework that improves automated program repair by retrieving relevant information and then steering an LLM to generate higher-quality patches.

Prerequisites

 
  Install Defects4J from https://github.com/rjust/defects4j 
  export PATH=$PATH:"path2defects4j"/framework/bin 
 

 
sudo apt-get install openjdk-8-jdk -y
export JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-amd64
export PATH=$JAVA_HOME/bin:$PATH
export CLASSPATH=.:$JAVA_HOME/lib/dt.jar:$JAVA_HOME/lib/tools.jar
 

 
numpy==1.24.3
pandas==2.0.3
torch==2.0.1
torchvision==0.15.2
transformers==4.29.2
openai==1.30.1
sentence_transformers==2.6.1
scikit-learn==1.4.2
 

Plausible Patches Generation and Validation

To run different components of RelRepair, follow the instructions below.

▶️ Run the baseline generator (base_gen)

 
  python3 base_gen.py -d ./dataset/defects4j-sf.json -bug Math-2
 

▶️ Run SigRepair and SnipRepair

Run retrieval first (SigRAG / SnipRAG)

 
  python3 sig_rag.py
  python3 snip_rag.py
 

Then run the generators (SigGen / SnipGen), using similar arguments as base_gen.py:

 
  python3 sig_gen.py -d ./dataset/defects4j-sf.json -bug Math-2
  python3 snip_gen.py -d ./dataset/defects4j-sf.json -bug Math-2