AI Dev Assistant (Local, Self-Hosted)

This project provides a fully local, containerized AI development assistant that can:

• Index Flutter and React codebases
• Answer natural language queries about your code
• Generate or refactor code based on your task prompts
• Operate completely offline — no external API usage

---

🚀 Tech Stack

• LlamaIndex v0.12.32
• Ollama (serving llama3)
• FastAPI (backend API)
• Docker + Docker Compose
• CLI interface (via Python + Rich)

---

⚙️ System Requirements

Make sure the following are installed on your local machine:

• Docker
• Docker Compose

Recommended:

• At least 8 GB RAM for LLaMA 3 (use 70B only if you have a powerful machine)
• Directory ./repos with:
  • flutter-app/ codebase
  • react-app/ codebase

---

🧰 Project Structure

.
├── docker-compose.yml
├── ollama/
│   └── Dockerfile
├── llama-index/
│   ├── Dockerfile
│   ├── requirements.txt
│   └── app/
│       ├── config.py
│       ├── ingest.py
│       └── main.py
├── cli/
│   ├── Dockerfile
│   └── cli.py
└── repos/
    ├── flutter-app/
    └── react-app/

---

🛠️ Setup Instructions

1. Place your codebases in the ./repos/ folder:

   mkdir -p repos/flutter-app repos/react-app
   

2. Build and start the assistant:

   docker compose up --build
   

   This will:

   • Start Ollama
   • Pull llama3 model
   • Start the indexing and API services

3. Use the assistant via CLI:

   docker compose run --rm cli
   

   Example prompts:

   > What does the AuthService class do?
   > Where is the login widget defined?
   > Refactor the main entry point
   

---

📌 Notes

• The assistant will build a vector index on first run from your code.
• The embedding model is sentence-transformers/paraphrase-MiniLM-L6-v2 (local).
• Ollama will automatically pull llama3 on first boot if not found.
• Everything runs offline — no OpenAI or external APIs.

---

🧪 Troubleshooting

• Model not found: Make sure Ollama is allowed to pull llama3 (requires internet on first run).
• Slow first response: First-time indexing and embedding can take 1–2 minutes depending on repo size.
• No answers?: Make sure your code contains enough comments or function names for semantic matching.

---

🙋 Contact

If you want to extend this with:

• Git PR integration
• VSCode plugin
• Web UI interface

Let’s connect!