News • Setup • How To Use • Paper • Docs • Paper List • Demo • Citation •

Read this in Chinese.

BMTools is an open-source repository that extends language models using tools and serves as a platform for the community to build and share tools. In this repository, you can (1) easily build a plugin by writing python functions (2) use external ChatGPT-Plugins.

This project is inspired by the open-source project LangChain and optimized for the usage of open-sourced tools like ChatGPT-Plugins, striving to achieve the open-source academic version of ChatGPT-Plugins.

• For new features and further developments, please go to XAgent.

• A demo of using BMTools to manipulate tools for meta analysis.

• [2023/5/28] We release ToolBench, a large-scale tool learning benchmark together with a capable model.

• [2023/5/25] The evaluation data used in the paper is partially released at data-test, we have also created large-scale SFT (100k+) high-quality tool-use training data at data-sft.

• [2023/5/19] Three new Tools are supported: Baidu Map, Google Scholar Search, and Zillow

• [2023/5/18] WebCPM is accepted by ACL 2023, a Chinese version of WebGPT.

• [older] Auto-GPT and BabyAGI are supported in BMTools.

git clone [email protected]:OpenBMB/BMTools.git
cd BMTools
pip install --upgrade pip
pip install -r requirements.txt
python setup.py develop

To support CPM-Bee:

git clone -b main --single-branch https://github.com/OpenBMB/CPM-Bee.git
cp -rf CPM-Bee/src/cpm_live bmtools/models/

Add your api keys to secret_keys.sh, then start the local tools

source secret_keys.sh
python host_local_tools.py

Then set the url of the plugin to http://127.0.0.1:8079/tools/{tool_name}/ (Remember the tailing /).

Just load it with the URL pointed to the .well-known/ai-plugin.json For example, set the url to https://www.klarna.com/, where https://www.klarna.com/.well-known/ai-plugin.json is a valid configuration.

from bmtools.agent.singletool import load_single_tools, STQuestionAnswerer

tool_name, tool_url = 'klarna',  'https://www.klarna.com/'
tool_name, tool_config = load_single_tools(tool_name, tool_url)
print(tool_name, tool_config)
stqa =  STQuestionAnswerer()

agent = stqa.load_tools(tool_name, tool_config)
agent("{Your Question}")

We can use multiple tools at the same time. Basically, the language model will do it recursively. It will treat the whole tool as an API, send questions to it, and the tool calls its sub-APIs to solve the question and send it back to parent tools.

Try this functionality using scripts like:

from bmtools.agent.tools_controller import load_valid_tools, MTQuestionAnswerer
tools_mappings = {
    "klarna": "https://www.klarna.com/",
    "chemical-prop": "http://127.0.0.1:8079/tools/chemical-prop/",
    "wolframalpha": "http://127.0.0.1:8079/tools/wolframalpha/",
}

tools = load_valid_tools(tools_mappings)

qa =  MTQuestionAnswerer(openai_api_key='', all_tools=tools)

agent = qa.build_runner()

agent("How many benzene rings are there in 9H-Carbazole-3-carboxaldehyde? and what is sin(x)*exp(x)'s plot, what is it integrated from 0 to 1? ")

1. Add your plugin to the mappings at beginning of web_demo.py

2. Start the webdemo

python web_demo.py

To develop a tool locally, you need to write a python function to build the tool and register it to the registry.

For example, you can write a tool that can execute python code and return the result. The following is a sample code:

from bmtools.tools import Tool
from pydantic import BaseModel

class ExecutionQuery(BaseModel):
    code: str

class ExecutionResult(BaseModel):
    result: str

def build_python_tool(config) -> Tool:
    tool = Tool(
        "PythonTool",
        "A plugin that can execute python code",
        name_for_model="python", 
        description_for_model="A plugin that can execute python code",
        contact_email="your@email",
    )

    @tool.post("/execute")
    def execute_python_code(query : ExecutionQuery) -> ExecutionResult:
        return ExecutionResult(
            result=eval(query.code)
        )
    
    return tool

Then you need to register the tool to the registry using the following code:

from bmtools.tools import register

@register("python")
def register_python_tool():
    return build_python_tool

Here we register the tool with the name python.

After you have developed a tool, you can contribute it to BMTools by following the steps below:

1. Fork this repository
2. Create a folder in bmtools/tools/{tool_name}
3. Add an api.py to the folder: bmtools/tools/{tool_name}/api.py and a __init__.py to the folder: bmtools/tools/{tool_name}/__init__.py
4. Register the tool in the __init__.py file you created in step 3 using the code in section 3.1
5. Import your tool in the __init__.py file under bmtools/tools
6. Add a test.py to test your tool automatically
7. Add a readme.md in your folder containing a brief introduction, contributor information, or anything you want to let others know.

The functions you wrote will be converted into an interface compatible with the OpenAI plugin. The AI models will read the name, description of the tools, as well as the name and descriptions of the tools' APIs. You can adjust the following aspect to make your API better understood by AI models.

• (1) name_for_model (tell the model what the tool is)
• (2) description_for_model (this will be displayed to the model before the tool is called, and you can include information on how to use the APIs)
• (3) The function name for each API function, as well as the name in @tool.get(). It's best if these two names match, as the name plays an important role in the model's API selection.
• (4) The function's doc string (can suggest to the model whether to use this API or not)
• (5) The function's return value, which can provide the model with error messages to guide its next steps, such as retrying or indicating a preferred next step
• (6) Reduce the errors in your API function.

A simple example to refer to is the Wolfram Alpha API.

If you use BMTools in your research, please cite:

@misc{qin2023tool,
      title={Tool Learning with Foundation Models}, 
      author={Yujia Qin and Shengding Hu and Yankai Lin and Weize Chen and Ning Ding and Ganqu Cui and Zheni Zeng and Yufei Huang and Chaojun Xiao and Chi Han and Yi Ren Fung and Yusheng Su and Huadong Wang and Cheng Qian and Runchu Tian and Kunlun Zhu and Shihao Liang and Xingyu Shen and Bokai Xu and Zhen Zhang and Yining Ye and Bowen Li and Ziwei Tang and Jing Yi and Yuzhang Zhu and Zhenning Dai and Lan Yan and Xin Cong and Yaxi Lu and Weilin Zhao and Yuxiang Huang and Junxi Yan and Xu Han and Xian Sun and Dahai Li and Jason Phang and Cheng Yang and Tongshuang Wu and Heng Ji and Zhiyuan Liu and Maosong Sun},
      year={2023},
      eprint={2304.08354},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}