v0.0.10

• Add the Tool-Integrated Reasoning (TIR) agent for Qwen2.5-Math.
• Add support for caching the LLM's responses. LLM caching can be enabled by providing cache_dir in the LLM's config, i.e.,llm = get_chat_model(cfg={..., "cache_dir": "/tmp/directory_to_save_llm_cache"}).
• Add self-play agents to simulate a dialogue between a user and an AI agent, mainly for data synthesis and testing purposes.

v0.0.9

• Bug fix: Incomplete PyPI packaging caused by the absence of __init__.py.

v0.0.8

• Multimodal function calling for Qwen2-VL. Please see the function calling example and the agent example for example usage.

v0.0.7

• Make dependencies for the code interpreter and RAG optional.
  • pip install qwen-agent now only installs minimal dependencies for function calling.
  • pip install qwen-agent[rag,code_interpreter,gui] installs optional dependencies for RAG, Code Interpreter, and GUI.

v0.0.6

• Added experimental support for parallel tool calling and tool choice.

v0.0.5

• Added long-context agents to understand 1-million-token documents. Read the blog for technical details.

v0.0.4

• Added experimental support for Gradio-based web GUI components for easy prototyping.
• Added vector retrieval, hybrid retrieval, and keyword stemming to improve RAG performance.

v0.0.3

• Improved the accuracy of the fast RAG strategy.
• Reorganized the implementations of RAG-related tools.

v0.0.2

• Improved documentation.
• Support local image files for DashScope Qwen-VL.

v0.0.1

• Qwen-Agent v0.0.1 released and uploaded to pypi.org.