PR Generator

A modular AI-powered system for intelligently analyzing Git changes and generating pull request (PR) recommendations using different strategies — including LLM-guided reasoning, CrewAI agents, and TF-IDF similarity.

---

🚀 Features

• Automatically detects unstaged Git changes in a repository
• Groups related files into logical PRs with:
  • 📝 Titles
  • 📄 Descriptions
  • 💡 Reasoning
• Supports multiple PR generation strategies:
  • CrewAI Agentic Analysis
  • LLM-Guided Grouping
  • TF-IDF Similarity Clustering
• Integrates with multiple LLM providers (OpenAI, Ollama)
• Uses standardized, strongly typed Pydantic 2.0 models
• Outputs structured JSON results for each PR suggestion
• Modular architecture for easy extension and experimentation

---

🧠 Approaches

1. crewai_approach/ (Primary Focus)

Uses CrewAI to orchestrate a team of agents that analyze and group code changes into PRs. Agents perform tasks like code analysis, pattern recognition, validation, and content generation.

2. llm_guided_approach/ (Legacy/Experimental)

Uses custom LLM pipelines and generators to group files and generate PR content.

3. tf_idf/ (Legacy/Experimental)

Uses traditional NLP techniques (TF-IDF + cosine similarity) to find related file changes.

---

🗂️ Project Structure (Simplified)

pr-generator/
├── crewai_approach/            # CrewAI-powered PR generation
│   ├── tools/                  # Modular tools for file analysis and grouping
│   ├── models/                 # Pydantic 2.0 data models
│   ├── config/                 # Agent/task configuration in YAML
│   ├── test_*.py               # Test files for toolchain and logic
│   ├── crew.py                 # Crew definition and setup
│   └── run_crew_pr.py          # Script entry point
│
├── llm_guided_approach/        # Legacy approach using LLM-only logic
│   ├── generators/             # Grouping and PR generation logic
│   ├── llm_service.py
│   └── scripts/generate_prs.py
│
├── tf_idf/                     # Traditional NLP-based grouping
│   └── tf_idf.py
│
├── shared/                     # Shared models and utilities
│   ├── config/
│   ├── models/
│   ├── tools/
│   └── utils/
│
├── embeddings/                 # Embedding-based similarity analysis
├── output/                     # Output from runs (structured groupings)
├── docs/                       # Design notes and architectural documentation
├── generate_graph.py           # PR graph visualizer
├── compare.py                  # Compare outputs from different strategies
└── README.md                   # This file

---

📦 Package Migration

The project has been migrated from setup.py to pyproject.toml for modern Python packaging:

• Why the change? PEP 621 introduced pyproject.toml as the new standard for Python package metadata, offering a more structured and comprehensive way to define package configuration.
• Benefits:
  • Cleaner dependency management with optional dependency groups
  • Better integration with modern Python tooling
  • Standardized configuration for development tools (black, isort, pytest)
  • Improved package discovery and data file handling

The migration maintains backward compatibility while adding new features like optional dependency groups for MCP servers and monitoring tools.

---

🧩 Setup

Python Version

• Python 3.8+

Create Virtual Environment

python -m venv venv
source venv/bin/activate

Installation Options

Option 1: Install Package (Recommended)

The project now uses pyproject.toml for package configuration:

# Install the package in development mode
pip install -e .

# Install with optional dependencies
pip install -e ".[mcp-server,monitoring,dev]"

# Install GitHub dependency (required)
pip install git+https://github.com/manavgup/llm-clients.git

Option 2: Install Dependencies Directly

For quick testing without installing the package:

pip install -r requirements.txt

For MCP server development:

pip install -r mcp_servers/requirements.txt

Verify Installation

After installing the package, you can verify that it's correctly installed by running the test script:

python test_install.py

This script will attempt to import key modules from the package and report any issues.

Environment Variables

For OpenAI usage:

export OPENAI_API_KEY=your_openai_key

For Ollama usage:

# Default URL is http://localhost:11434
export OLLAMA_API_URL=http://localhost:11434

---

🧪 Example Usage (CrewAI)

# Run CrewAI-based PR generator on a Git repo
python crewai_approach/run_crew_pr.py /path/to/your/repo --max-files 30 --provider openai

Optional flags:

• --provider openai|ollama – Choose LLM backend
• --max-files – Max files per PR group
• --model – LLM model to use (e.g., gpt-4, llama3)
• --output – Path to save JSON output
• --verbose – Enable detailed logging

---

🧠 Agent Architecture (CrewAI)

• Repository Analyzer Tool: Scans for changed files, computes metrics.
• Pattern Analyzer Tool: Extracts filename/pattern similarities.
• Batch Splitter Tool: Splits files into manageable chunks for analysis.
• Grouping Strategy Selector Tool: Determines which grouping strategy to use.
• Group Refiner/Validator Tools: Improves and validates group quality.
• PR Crew: Orchestrates agents and passes structured data between tools.

---

📦 Output Format

Each generated PR suggestion includes:

{
  "title": "feat(core): improve auth",
  "description": "Enhanced auth flow...",
  "reasoning": "These files all deal with authentication middleware.",
  "files": [
    "backend/core/auth_middleware.py",
    "backend/core/token_utils.py"
  ]
}

---

✅ Tests

To run tests (example: for CrewAI tools):

pytest crewai_approach/

---

📌 Roadmap

• Add LangGraph-style composable agents (under langgraph_approach/)
• Web-based visualization for PR suggestions
• Fine-tuned LLM models for code grouping
• Human-in-the-loop feedback loop

---

🧑‍💻 Contributing

If you're collaborating specifically on crewai_approach/, please refer to its internal requirements.txt and test files. Contributions are welcome via PRs.

---

🪪 License

MIT License