🔮 Predic: Your Offline & Private AI Coding Assistant

The Private, Offline AI Coding Companion for VS Code.

Predic is a powerful AI coding assistant that runs 100% locally on your machine. It provides chat assistance, code explanation, and intelligent inline completions ("ghost text") without sending a single line of your code to the cloud.

---

✨ Why Predic?

• 🔒 100% Privacy: Your code never leaves your device. Ideal for enterprise or sensitive projects.
• ⚡ Offline Capable: Works perfectly without an internet connection (after initial model download).
• 🧠 Bring Your Own Model: Compatible with any GGUF model supported by KoboldCpp (Qwen, DeepSeek, Llama 3, Phi-3).
• 💰 Free: No subscription fees. You own the compute.

🚀 Features

💬 Context-Aware Chat

Chat with your codebase naturally.

• Smart Context: Type @filename to instantly add a file's content to the AI's context window.
• Streaming Responses: Real-time, typewriter-style generation.
• Pro UI: Clean, native VS Code interface with syntax highlighting, copy/edit/insert actions, and file chips.

👻 Inline Code Completion

Get "Ghost Text" suggestions as you type in the editor.

• FIM Support: Uses "Fill-In-the-Middle" technology to understand code before and after your cursor.
• Low Latency: Optimized for speed on consumer hardware.

🎛️ Model Dashboard

A dedicated control center for your AI.

• Model Manager: Download curated high-performance coding models (like Qwen 2.5-Coder) directly from the UI.
• One-Click Switch: Instantly swap between different models for different tasks (e.g., a small model for speed, a large model for complex logic).
• Official Models: Try ReaComplete, our fine-tuned model optimized for React/TypeScript.

🛠️ Productivity Tools

• Explain Code: Right-click any selection > Predic: Explain Selection.
• Fix Errors: Right-click > Predic: Fix Error.

📦 Requirements

Since Predic runs locally, your hardware matters:

• OS: Windows, macOS, or Linux.
• RAM: Minimum 8GB (for small models), 16GB+ recommended.
• GPU (Optional): NVIDIA GPU with 6GB+ VRAM is highly recommended for faster response times.

🔧 Setup Guide

1. Install the Extension from the VS Code Marketplace.
2. Download the Engine:
   • Predic relies on KoboldCpp to run models.
   • Download the latest koboldcpp executable from here.
3. Open the Dashboard:
   • Click the Predic Icon in the Activity Bar.
   • Click the Settings/Gear Icon in the Chat header.
   • Point Predic to your koboldcpp.exe (or binary).
4. Download a Model:
   • In the Dashboard, click Download on one of the "Recommended Models" (e.g., Qwen 2.5 Coder 1.5B).
5. Start Coding:
   • Click Load on your downloaded model.
   • Go back to the Chat tab and say "Hello!".

⚙️ Extension Settings

You can configure Predic in VS Code Settings (Ctrl+,):

Setting	Description
predic.koboldCppPath	Full path to your koboldcpp executable.
predic.modelDir	Folder where your .gguf models are stored.
predic.gpuLayers	Number of model layers to offload to your GPU (-1 for auto).
predic.contextSize	Maximum context window in tokens (Default: 8192).
predic.port	Local server port (Default: 5001).

🤝 Contributing

Predic is open source! We welcome issues, feature requests, and pull requests. Check out our GitHub Repository.

Note: If you want to add features or fix bugs, please check out the Developer Guide.

📄 License

MIT License

---

🙏 Acknowledgements

• A huge thank you to the team behind koboldcpp for making local, in-browser, and in-app AI accessible to everyone.