Open-source AI Assistant AIlice: Capable of Controlling Multiple Agents to Collaboratively Complete Complex Tasks
-
MyShellAI has just launched the open-source project AIlice, which can control multiple agents to collaboratively complete complex tasks such as document retrieval, downloading, data analysis, and drawing. It can also generate code for tasks and run it in a virtual environment.
The AIlice project aims to create a self-contained AI assistant, similar to JARVIS, using open-source models. AIlice builds a "text computer" with large language models (LLMs) as the core processor. Currently, AIlice has demonstrated proficiency in a range of tasks, including topic research, coding, system administration, literature retrieval, and complex mixed tasks beyond these basic capabilities. AIlice leverages GPT-4 to achieve near-perfect performance in daily tasks and is moving toward practical applications.
Key technical features of AIlice include deep research capabilities on specialized topics, the ability to read and analyze articles and academic works, programming and script execution with advanced automation features (similar to a full-fledged coder and efficient system management tools), voice interaction support, seamless integration with commercial models like GPT-4, and a more intuitive and flexible approach to user interaction. It supports multiple models, features a natural and highly fault-tolerant interactive agent invocation tree architecture, flexible parsing of LLM outputs, self-building and dynamic loading of modules to interact with the environment, and offers unlimited possibilities for extending functionality. Through dialogue with AIlice, users can accomplish various tasks such as topic research, coding, and system management. Currently, AIlice lacks runtime control mechanisms, so when using commercial LLMs, its operations need to be closely monitored. AIlice is built through multi-agent collaboration, with the user being one of them. When additional information is needed, AIlice will request input from the user, and the level of detail provided by the user is crucial for its success.
One of AIlice's future goals is to achieve automatic dataset collection and construction. While the current researcher functionality has some shortcomings, it can already produce some interesting results. AIlice can also build external interaction modules (called ext-modules), granting it unlimited extensibility. With just a simple prompt from the user, AIlice can construct modules and load newly implemented modules via commands.
Project entry: https://github.com/myshell-ai/AIlice