backend

🐝 OpenBee Backend

OpenBee is an offline AI Voice Assistant backend powered by LiveKit Agents, local inference plugins, and modern GPU/CPU acceleration.

🚀 Setup Instructions

1. Prerequisites

• uv installed.
• Python 3.12+

2. Install Dependencies

uv sync

3. Setup Environment

Create a .env file in the backend/ directory (copy from .env.example) and add your LiveKit credentials:

LIVEKIT_URL=wss://your-livekit-url
LIVEKIT_API_KEY=your-api-key
LIVEKIT_API_SECRET=your-api-secret

---

📥 Download Models

Before running the agent for the first time, you must download the local model files for Whisper (STT), Gemma (LLM), and Kokoro (TTS).

Run the provided script:

uv run python utils/download_all_models.py

This will place the weights into the models/ directory.

💡 Changing the Local LLM

If you want to use a different quantized version of Gemma 3 (e.g., a smaller 4-bit version for speed), follow these steps:

1. Browse Models: Go to unsloth/gemma-3-1b-it-GGUF.
2. Update Download Script: In utils/download_all_models.py, change the filename in the hf_hub_download call to your desired .gguf file.
3. Update Agent Config: In core/agent.py, update the model_path in the LocalLlamaLLM initialization to match the new filename:

   llm=LocalLlamaLLM(model_path="models/your-new-model.gguf")

---

🏃‍♂️ Running the System

Start the API Server

The API server handles connection details for the frontend:

uv run server.py

---

📂 Project Structure

• core/: Core agent logic, server definitions, and configurations.
• plugins/: Custom STT, TTS, and LLM implementations for local inference.
• utils/: Helper scripts for token generation and model management.
• models/: Local storage for model weights and binaries.