KwaiAgents is a comprehensive system for building and evaluating generalized information-seeking agents powered by Large Language Models (LLMs). It offers pre-trained LLMs with agent capabilities, a large dataset for fine-tuning, and a robust benchmark for evaluating agent performance. This project is ideal for researchers and developers looking to build sophisticated AI agents that can plan, use tools, and reflect.

How It Works

KwaiAgents comprises several components: KAgentSys-Lite, a simplified agent framework derived from projects like BabyAGI and Auto-GPT; KAgentLMs, LLMs fine-tuned for agent tasks like planning and tool-use using a "Meta-agent tuning" approach; KAgentInstruct, a dataset of over 200k agent-related instructions; and KAgentBench, a benchmark with over 3,000 human-edited evaluation cases across various agent capabilities. The system supports both OpenAI-compatible API endpoints (via vLLM and FastChat) and local CPU inference (via llama.cpp).

Quick Start & Requirements

• Installation: Clone the repository and install dependencies: git clone git@github.com:KwaiKEG/KwaiAgents.git, cd KwaiAgents, conda create -n kagent python=3.10, conda activate kagent, pip install -r requirements.txt, python setup.py develop.
• Prerequisites: Python 3.10+, Miniconda, GPU with CUDA for vLLM serving, or CPU for llama.cpp. chromedriver is required for the browse_website tool.
• Serving: Deploy models via vLLM/FastChat or llama.cpp. Detailed instructions are provided for both GPU and CPU serving.
• Usage: Interact via CLI: kagentsys --query="Who is Andy Lau's wife?" --llm_name="gpt-3.5-turbo". Local LLM usage requires specifying host/port.
• Links: Dataset, Benchmark, Paper.

Highlighted Details

• Fine-tuned models (e.g., Qwen-MAT, Baichuan2-MAT) show significantly improved performance on agent benchmarks compared to base models.
• KAgentBench provides automated evaluation metrics for planning, tool-use, reflection, concluding, and profiling.
• Supports both cloud-based LLMs (OpenAI API) and self-hosted local LLMs.
• Includes a custom tool example and instructions for integrating new tools.

Maintenance & Community

The project is open-sourced by KwaiKEG from Kuaishou Technology. Recent updates include new model releases (Qwen1.5-14B-MAT) and refreshed benchmark results. The project has gained significant media attention.

Licensing & Compatibility

The repository does not explicitly state a license in the README. The models are available on Hugging Face under unspecified licenses. Compatibility for commercial use or closed-source linking is not detailed.

Limitations & Caveats

KAgentSys-Lite has limitations compared to the full KAgentSys, including a reduced toolset, no memory mechanisms, and slightly lower performance. The browse_website tool requires manual chromedriver setup. Network issues might affect the duckduckgo_search tool, potentially requiring proxy configuration.