Think-Search-Patch (TSP): A Retrieval-Augmented Reasoning Framework for Repository-Level Code Repair

Table of Contents

• Overview
• Framework Architecture
• Key Features
• Installation
• Project Structure
• Quick Start
• Usage
• License

Overview

Think-Search-Patch (TSP) is a retrieval-augmented reasoning framework designed for repository-level code repair. This framework combines advanced retrieval techniques with reasoning capabilities to provide effective solutions for code repair tasks at the repository level.

Framework Architecture

TSP Framework Overview

Index Construction Pipeline

Key Features

• Retrieval-Augmented Reasoning: Combines retrieval mechanisms with reasoning capabilities
• Repository-Level Understanding: Handles code repair at the repository scale
• Information Masking: Advanced information masking techniques in both training frameworks
• Comprehensive Evaluation: Search reward calculation and LLM-judge evaluation methods
• Modular Architecture: Well-organized components for easy development and deployment

Installation

Prerequisites

• Python 3.9+
• CUDA-compatible GPU (recommended)

Install Required Frameworks

# Install SFT training framework
cd verl && pip install -e .
cd ..

# Install DPO training framework
cd OpenRLHF && pip install -e .
cd ..

# Install index construction framework
cd SWE-bench && pip install -e .
cd ..