Release Highlights

Numerical statistics

• Huge number of commits: increased from ~1300 to ~2000
• Significantly improved test coverage: from ~30% to 80%+

Ease of use

• Example simplification: Now only 7 lines of code are needed to implement a general election debate. Simplified debate and knowledge base related examples
• Pure asynchronous: LLM API calls are changed to pure asynchronous, removing all synchronization code

Framework update

• Integrated code to support Huggingface demo and AgentStore
• Added serialization and deserialization of important objects
• Supported breakpoint recovery
• All types modified to base on pydantic v2. In order to keep it simple enough, all private variables are changed to public
• Moved most logic of ActionOutput to ActionNode

LLM provider update

• OpenAI package upgraded to v1.6.0
• Supported or optimized LLMs: Gemini, ZhipuAI, Ollama, and other Open LLMs etc.

Other update

• Removed OpenInterpreter and StableDiffusion part of the code
• Added ToT strategy
• Installation process optimized: now you can use pip install -e .[dev] / pip install -e .[test] to install the package corresponding to the scenario

What's Changed

Backwards Incompatible Changes

• IMPORTANT: In accordance with OpenAI's official specifications, the OPENAI_API_BASE has been renamed to OPENAI_BASE_URL in config.yaml
• The underlying abstraction has been renamed from BaseGPTAPI to BaseLLM, making it more generic and capable of supporting API / UI / local interfaces.
• Role._rc -> Role.rc
• Removed BaseLLM.ask and BaseLLM.ask_code, maintained the asynchronous BaseLLM.aask and BaseLLM.aask_code only

Features

• Github related: Add pre-commit code format check and corrected the format #587 #594 by @voidking
• Optimized example, used purely asynchronous api call, integrated code to support Huggingface and AgentStore, upgraded OpenAI package, built up test suite #589 #671 by @geekan @better629 @seehi @iorisa @shenchucheng @Stitch-z @Justin-ZL @voidking @garylin2099
• Based important objects to pydantic BaseModel, added serialization and deserialization #590 by @better629
• Supported Google Gemini as underlying LLM #595 by @better629
• Github related: Provided issue and pr template #598 by @better629
• Supported Ollama as underlying LLM #603 by @better629
• Enabled MetaGPT to be used as a dependency for web applications, such as https://huggingface.co/spaces/deepwisdom/MetaGPT. #621 by @shenchucheng
• Experimental implementation of ToT strategy #643 by @stellaHSR

Bugfixes and improvements

• Solve Message initialization issue #596 by @orange-crow #599 by @better629
• Fixed QaEngineer missing test_round #600 by @better629
• Upgraded langchain to be compatible with openai new version #605 by @seehi
• Update examples #606 by @better629
• Fixed bug in sk_agent #623 by @femto
• Set openai proxy for class ZhiPuAPTAPI #630 by @springasa
• Fix filename problem in research.py #634 @shenchucheng

Documents

• Fixed broken docs links #575 by @shenchucheng
• Fixed typos #602 by @eltociear