Ablite Web (Secure AI)

🇹🇷 Türkçe okumak için tıklayın (Read in Turkish) - README_TR.md

Ablite Web is a Highly Encrypted Artificial Intelligence (AI) assistant that allows you to safely and quickly chat with HuggingFace models locally (both MLX and GGUF formats). It is specifically designed for conducting sensitive and confidential conversations with Uncensored / Abliterated models.

While it pushes the limits of Apple Silicon (M-series chips) for MacBook users via MLX, it also offers full Cross-Platform Support (Windows & Linux) via the GGUF format!

💡 Performance Note: This system was tested on a MacBook with an M4 processor and 24 GB of RAM, and runs flawlessly and smoothly even with massive parameter models such as mlx-community/Qwen3.8-27B-4BIT. Please note that when using RAG with Qwen 3.8 27B and appending the last 3 messages as context, it pushes the limits of 24 GB RAM. Devices with larger RAM will have no issues.

---

⚡ Lightweight CLI Mode (No Web/RAG)

If the combination of the RAG system and the web client is consuming too much of your system's resources or RAM (especially when using very large models like Qwen 3.8 27B on devices with 24GB or less RAM), you can use the standalone script to chat with the model directly from your terminal.

# Chat with the default model
python run_without_rag_and_web.py -m "Your prompt here"

# Enable thinking mode (if supported by the model)
python run_without_rag_and_web.py -m "Your prompt here" -t

# Use a different model repo
python run_without_rag_and_web.py -m "Your prompt here" -r "mlx-community/Llama-3.2-3B-Instruct-4bit"

🔥 Key Features

• 🔒 Maximum Privacy (Data at Rest Security): Your entire chat history saved in the database (SQLite) is stored using the AES-GCM encryption standard. When your device is turned off, stolen, or physically falls into the hands of unwanted individuals, it is mathematically impossible to access the content of your messages without your password.
  • This is all anyone will see if they manage to breach your chats.db:

  9e9c3e9a-7c9c-469b-980b-dfb01e38a2e5|user|nONzL7w1h3R1t6O...|2026-08-22
  9e9c3e9a-7c9c-469b-980b-dfb01e38a2e5|assistant|Q2-iaKhsI8nxteckbgTr7h...|2026-08-22
  

• 🧠 Smart Decrypt-on-the-fly RAG (Retrieval-Augmented Generation): To prevent the AI from losing its "memory" during long conversations, all your past messages are instantly converted into vectors (embeddings) in the background. Even these vectors are written to the database encrypted. When you ask a new question, the system decrypts the data only momentarily in the RAM, finds the relevant context, passes it to the AI, and immediately deletes it from memory.
• 💻 Cross-Platform (Mac, Windows, Linux): Thanks to the GGUF format (llama.cpp), the system runs not only on Macs but also on Windows and Linux machines. While M-series Macs utilize the MLX framework to push the hardware limits, other devices utilize the CPU or GPU (CUDA/Vulkan) via GGUF.
• 🌐 Modern Web Interface: Instead of clunky terminal screens, it offers a transparent (Glassmorphism), smoothly animated, dark-themed chat interface. Designed with Vanilla JS & CSS; no Node.js required!
• 📦 Built-in Model Manager: Through the web interface, you can select your downloaded HuggingFace models, enter a new HuggingFace Repo ID to download any model (MLX and GGUF formatted), and monitor download percentages (Progress Bar) in real-time.
• 🚀 One-Click Installation: No hassle of setting up a virtual environment (venv) and downloading packages manually! The install.sh script handles everything automatically.

---

🛠️ Installation

The project is configured to install completely automatically. Simply paste the following commands into your terminal:

# Clone the repository
git clone https://github.com/furkanhaydari/abliterated.git
cd abliterated

# Start the automatic installation wizard (Installs the virtual environment and packages)
./install.sh

🖥️ Windows & Linux GPU Setup (CUDA)

For MacBook (Apple Silicon) users, GPU (Metal) support is enabled automatically and requires no extra setup. However, if you are on Windows or Linux and want to utilize Nvidia GPU (CUDA) acceleration, you must reinstall the GGUF engine with CUDA support after the standard installation:

For Windows (PowerShell):

# Activate the environment
.\env\Scripts\activate
# Install with CUDA support
$env:CMAKE_ARGS="-DGGML_CUDA=on"
pip install --force-reinstall --no-cache-dir llama-cpp-python

For Linux:

# Activate the environment
source env/bin/activate
# Install with CUDA support
CMAKE_ARGS="-DGGML_CUDA=on" pip install --force-reinstall --no-cache-dir llama-cpp-python

---

🚀 Getting Started (Usage)

Once the installation is complete, starting the server is very simple:

# Activate the virtual environment (if not already active):
source env/bin/activate

# Start the Server:
ablite

Alternatively, after manually doing pip install -e . without using install.sh, you can call the ablite command from anywhere.

When the server starts, go to http://localhost:8000 from your browser.

1. Set a password on the main screen to unlock the application.
2. Click the model manager icon in the top right to manage your downloaded models or download a new one.
3. Enjoy a confidential AI experience with Abliterated models!

---

💡 Security Note (Important!)

The password you enter on the lock screen is converted into an ultra-secure encryption key in the background using PBKDF2 algorithms, valid only for that session. This key is NEVER written to disk; it is kept strictly in RAM. When you close your browser or the server, your password vanishes from RAM.

⚠️ WARNING: If you forget or lose your password, everything in the database remains cryptographically locked. There is no recovery mechanism such as "I forgot my password". If the password is forgotten, everything related to your old conversations vanishes permanently and you have to start from scratch!

⚠️ BROWSER USAGE WARNING: If you are using this application via a web browser, it is highly recommended to use a New Incognito/Private Window with ALL extensions disabled. Also, ensure any built-in browser data collection or translation features are turned off to maintain maximum privacy and prevent extensions from leaking your sensitive, decrypted chat data.

📝 License

Released as open source under the MIT License. You may use, modify, and distribute it as you wish.