Moss

Moss is my graduation project based on HuggingGPT. It is a system based on large language model agents designed to tackle complex artificial intelligence tasks, powerd by LangChain.

🏗️ Road Work Ahead

Moss presents a collaborative system comprising LLM agents serving as controllers and multiple expert tools acting as collaborative executors. The workflow encompasses:

1. Task Planning: Employing LLMs to analyze user requests to comprehend their intentions and break them down into feasible solvable tasks.
2. Tool Selection: In order to address the planned tasks, LLM chooses appropriate expert tools considering their descriptions.
3. Task Execution: LLM invokes and executes each selected tool, subsequently delivering the results.
4. Response Generation: Ultimately, LLM integrates the predictions made by all tools and generates appropriate responses in friendly natural language.

Installation

• Clone this repository

$ git clone https://github.com/MoyusiteruIori/moss.git
$ cd moss

• Create configuration file from template

$ cp .env.example .env

Replace OPENAI_API_KEY, HF_TOKEN and SD_TOKEN in .env file with your own keys/tokens. If you don't know what these are, see openai-api, huggingface and stability.ai.

Install with Docker (recommended)

Note: Only Linux systems are guaranteed good support as the host machine for this Docker image. Ubuntu 22.04 is recommended.

Option 1: Use docker-compose (recommended):

$ docker compose up -d
# To utilize your local GPUs for certain local inferences:
# INFERENCE_MODE=HYBRID docker compose up -d

Option 2: Build a Docker image locally and manually start:

$ docker build -t moss .
# To utilize your local GPUs for certain local inferences:
# docker build --build-arg INFERENCE_MODE=HYBRID -t moss .
$ docker run -p 7860:7860 --name moss moss

Install without Docker

Linux and macOS are supported.

• Install dependencies

conda create -n moss python=3.11
conda activate moss
pip install -r requirements.txt

Optional: To utilize your local GPUs for certain local inferences, you'll need some extra packages (Linux only) :

pip install accelerate==0.29.3 diffusers==0.27.2 controlnet-aux==0.0.8 transformers==4.40.1

• Run gradio demo

python gradio_demo.py

You can then access localhost:7860 to use moss.

• Run command line version

python cli.py

System Demo

Demo video for my final presentation!

License

MIT

Credits

• HuggingGPT