Code Sushi 🍣

Code Sushi is a tool to help developers take large repos and cut them down into bite-sized pieces that can be fed into an LLM for answering questions in a few minutes.

Demo

🌟 Features

• Privacy-First: All code processing happens locally in the CLI tool. Your code only leaves your machine for inference.
• Multi-Language Support: Can work with any programming language. (currently only Python, Typescript, and PHP is supported)
• Modular Components: Easily swap out LLM, vector DB, and embedding providers with your preferred choices.

🧑‍🤝‍🧑 Who is this for?

• You work in a company that has very strict data privacy policies, operates in a regulated industry, or you are just paranoid about your code.

If not, Cursor is an excellent choice with much more features.

• You are dealing with a very large codebase (+50k lines of code) that cannot fit into the context window.

If not, I recommend Gitingest as an alternative for smaller repos.

• You just want question-answering capabilities for your codebases.

If you need more like code completion, etc... there are no open-source IDEs that can do that yet. Continue is a good open-source VSCode extension.

📦 Installation

You just need to have Python 3.12 or higher.

pip install code-sushi

🚀 Quick Start

Here is how you can use Code Sushi locally.

sushi init   # Will create a .sushiignore file in your project root
sushi run    # Starts the process of cutting down your repo and vectorizing the chunks

# After the process is done, you can ask questions to the LLM
sushi ask "How long are payment links valid for unverified users?"

Before the process starts, the tool will show a summary of the files detected and ask you to confirm the process.

📖 Usage

Basic Commands

sushi init          # Creates a .sushiignore and sushi-config.json file in your project root
sushi slice         # Will start the process of cutting down your repo into smaller pieces and summarize them
sushi clean         # Will clean up the stored vectors of the summaries for the project
sushi run           # Will run the entire process to slice and vectorize your repo in one command
sushi chat          # Will start a chat with the LLM
sushi ask           # Will ask a single question to the LLM

You can run the commands with the -h flag to get more information about the parameters each command accepts.

Configuration

The sushi init command automatically creates a sushi-config.json file in your project root:

{
    "max_agents": 10,
    "ai_provider": "openai",
    "vector_db_provider": "pinecone",
    "embedding_max_chunk_size": 128,
    "vector_db_max_concurrent_requests": 25
}

This is also where we read the API keys for the third party services from. You should make sure this file is not checked into your repo.

🔍 Inner Workings

Code Sushi uses a combination of AI services for each part of its workflow to understand and classify your code effectively:

• We use Together.ai to access Llama 3.3 70B for understanding the code and answering questions.
• We use Voyage.ai for creating embeddings and reranking search results.
• We use Pinecone to store the vector embeddings and search against them for RAG.

🔄 Modular Architecture

Each component is designed to be easily swappable. If you desire to do so, you can implement your own LLM provider, vector database, and more.

My only ask: I would kindly ask you to contribute your changes back to the project so that others can benefit from your work instead of a fork. 😇

📝 License

This project is licensed under the MIT License – see the LICENSE file for details.

🙏 Acknowledgements

• Thank you to Romain for the initial inspiration with his gitingest project.

👥 Contributors

Want to be part of building Code Sushi? Check out our Contributing Guidelines to get started! We always welcome contributions to improve the tool for benefit of the community.

---

Made with ❤️ and many 🍣. - The Code Sushi team!