Modelscope Hub ｜ Paper ｜ Demo
中文 ｜ English
Modelscope-Agent is a customizable and scalable Agent framework. A single agent has abilities such as role-playing, LLM calling, tool usage, planning, and memory. It mainly has the following characteristics:
- Simple Agent Implementation Process: Simply specify the role instruction, LLM name, and tool name list to implement an Agent application. The framework automatically arranges workflows for tool usage, planning, and memory.
- Rich models and tools: The framework is equipped with rich LLM interfaces, such as Dashscope and Modelscope model interfaces, OpenAI model interfaces, etc. Built in rich tools, such as code interpreter, weather query, text to image, web browsing, etc., make it easy to customize exclusive agents.
- Unified interface and high scalability: The framework has clear tools and LLM registration mechanism, making it convenient for users to expand more diverse Agent applications.
- Low coupling: Developers can easily use built-in tools, LLM, memory, and other components without the need to bind higher-level agents.
- 🔥🔥🔥Aug 8, 2024: A new graph based code generation tool CodexGraph is released by Modelscope-Agent, it has been proved effective and versatile on various code related tasks, please check example.
- 🔥🔥Aug 1, 2024: A high efficient and reliable Data Science Assistant is running on Modelscope-Agent, please find detail in example.
- 🔥July 17, 2024: Parallel tool calling on Modelscope-Agent-Server, please find detail in doc.
- 🔥June 17, 2024: Upgrading RAG flow based on LLama-index, allow user to hybrid search knowledge by different strategies and modalities, please find detail in doc.
- 🔥June 6, 2024: With Modelscope-Agent-Server, Qwen2 could be used by OpenAI SDK with tool calling ability, please find detail in doc.
- 🔥June 4, 2024: Modelscope-Agent supported Mobile-Agent-V2arxiv，based on Android Adb Env, please check in the application.
- 🔥May 17, 2024: Modelscope-Agent supported multi-roles room chat in the gradio.
- May 14, 2024: Modelscope-Agent supported image input in
RolePlayagents with latest OpenAI modelGPT-4o. Developers can experience this feature by specifying theimage_urlparameter. - May 10, 2024: Modelscope-Agent launched a user-friendly
Assistant API, and also provided aTools APIthat executes utilities in isolated, secure containers, please find the document - Apr 12, 2024: The Ray version of multi-agent solution is on modelscope-agent, please find the document
- Mar 15, 2024: Modelscope-Agent and the AgentFabric (opensource version for GPTs) is running on the production environment of modelscope studio.
- Feb 10, 2024: In Chinese New year, we upgrade the modelscope agent to version v0.3 to facilitate developers to customize various types of agents more conveniently through coding and make it easier to make multi-agent demos. For more details, you can refer to #267 and #293 .
- Nov 26, 2023: AgentFabric now supports collaborative use in ModelScope's Creation Space, allowing for the sharing of custom applications in the Creation Space. The update also includes the latest GTE text embedding integration.
- Nov 17, 2023: AgentFabric released, which is an interactive framework to facilitate creation of agents tailored to various real-world applications.
- Oct 30, 2023: Facechain Agent released a local version of the Facechain Agent that can be run locally. For detailed usage instructions, please refer to Facechain Agent.
- Oct 25, 2023: Story Agent released a local version of the Story Agent for generating storybook illustrations. It can be run locally. For detailed usage instructions, please refer to Story Agent.
- Sep 20, 2023: ModelScope GPT offers a local version through gradio that can be run locally. You can navigate to the demo/msgpt/ directory and execute
bash run_msgpt.sh. - Sep 4, 2023: Three demos, demo_qwen, demo_retrieval_agent and demo_register_tool, have been added, along with detailed tutorials provided.
- Sep 2, 2023: The preprint paper associated with this project was published.
- Aug 22, 2023: Support accessing various AI model APIs using ModelScope tokens.
- Aug 7, 2023: The initial version of the modelscope-agent repository was released.
clone repo and install dependency：
git clone https://github.com/modelscope/modelscope-agent.git
cd modelscope-agent && pip install -r requirements.txtThe ModelScope Notebook offers a free-tier that allows ModelScope user to run the FaceChain application with minimum setup, refer to ModelScope Notebook
# Step1: 我的notebook -> PAI-DSW -> GPU环境
# Step2: Download the [demo file](https://github.com/modelscope/modelscope-agent/blob/master/demo/demo_qwen_agent.ipynb) and upload it to the GPU.
# Step3: Execute the demo notebook in order.The agent incorporates an LLM along with task-specific tools, and uses the LLM to determine which tool or tools to invoke in order to complete the user's tasks.
To start, all you need to do is initialize an RolePlay object with corresponding tasks
- This sample code uses the qwen-max model, drawing tools and weather forecast tools.
- Using the qwen-max model requires replacing YOUR_DASHSCOPE_API_KEY in the example with your API-KEY for the code to run properly. YOUR_DASHSCOPE_API_KEY can be obtained here. The drawing tool also calls DASHSCOPE API (wanx), so no additional configuration is required.
- When using the weather forecast tool, you need to replace YOUR_AMAP_TOKEN in the example with your AMAP weather API-KEY so that the code can run normally. YOUR_AMAP_TOKEN is available here.
# 配置环境变量；如果您已经提前将api-key提前配置到您的运行环境中，可以省略这个步骤
import os
os.environ['DASHSCOPE_API_KEY']=YOUR_DASHSCOPE_API_KEY
os.environ['AMAP_TOKEN']=YOUR_AMAP_TOKEN
# 选用RolePlay 配置agent
from modelscope_agent.agents.role_play import RolePlay # NOQA
role_template = '你扮演一个天气预报助手，你需要查询相应地区的天气，并调用给你的画图工具绘制一张城市的图。'
llm_config = {'model': 'qwen-max', 'model_server': 'dashscope'}
# input tool name
function_list = ['amap_weather', 'image_gen']
bot = RolePlay(
function_list=function_list, llm=llm_config, instruction=role_template)
response = bot.run('朝阳区天气怎样？')
text = ''
for chunk in response:
text += chunkResult
- Terminal runs
# 第一次调用llm的输出
Action: amap_weather
Action Input: {"location": "朝阳区"}
# 第二次调用llm的输出
目前，朝阳区的天气状况为阴天，气温为1度。
Action: image_gen
Action Input: {"text": "朝阳区城市风光", "resolution": "1024*1024"}
# 第三次调用llm的输出
目前，朝阳区的天气状况为阴天，气温为1度。同时，我已为你生成了一张朝阳区的城市风光图，如下所示：
An Agent object consists of the following components:
LLM: A large language model that is responsible to process your inputs and decide calling tools.function_list: A list consists of available tools for agents.
Currently, configuration of Agent may contain following arguments:
llm: The llm config of this agent- When Dict: set the config of llm as {'model': '', 'api_key': '', 'model_server': ''}
- When BaseChatModel: llm is sent by another agent
function_list: A list of tools- When str: tool names
- When Dict: tool cfg
storage_path: If not specified otherwise, all data will be stored here in KV pairs by memoryinstruction: the system instruction of this agentname: the name of agentdescription: the description of agent, which is used for multi_agentkwargs: other potential parameters
Agent, as a base class, cannot be directly initialized and called. Agent subclasses need to inherit it. They must implement function _run, which mainly includes three parts: generation of messages/propmt, calling of llm(s), and tool calling based on the results of llm. We provide an implement of these components in RolePlay for users, and you can also custom your components according to your requirement.
from modelscope_agent import Agent
class YourCustomAgent(Agent):
def _run(self, user_request, **kwargs):
# Custom your workflowLLM is core module of agent, which ensures the quality of interaction results.
Currently, configuration of `` may contain following arguments:
model: The specific model name will be passed directly to the model service provider.model_server: provider of model services.
BaseChatModel, as a base class of llm, cannot be directly initialized and called. The subclasses need to inherit it. They must implement function _chat_stream and _chat_no_stream, which correspond to streaming output and non-streaming output respectively.
Optionally implement chat_with_functions and chat_with_raw_prompt for function calling and text completion.
Currently we provide the implementation of three model service providers: dashscope (for qwen series models), zhipu (for glm series models) and openai (for all openai api format models). You can directly use the models supported by the above service providers, or you can customize your llm.
For more information please refer to docs/modules/llm.md
We provide several multi-domain tools that can be configured and used in the agent.
You can also customize your tools with set the tool's name, description, and parameters based on a predefined pattern by inheriting the base tool. Depending on your needs, call() can be implemented. An example of a custom tool is provided in demo_register_new_tool
You can pass the tool name or configuration you want to use to the agent.
# by tool name
function_list = ['amap_weather', 'image_gen']
bot = RolePlay(function_list=function_list, ...)
# by tool configuration
from langchain.tools import ShellTool
function_list = [{'terminal':ShellTool()}]
bot = RolePlay(function_list=function_list, ...)
# by mixture
function_list = ['amap_weather', {'terminal':ShellTool()}]
bot = RolePlay(function_list=function_list, ...)image_gen: Wanx Image Generation. DASHSCOPE_API_KEY needs to be configured in the environment variable.code_interpreter: Code Interpreterweb_browser: Web Browsingamap_weather: AMAP Weather. AMAP_TOKEN needs to be configured in the environment variable.wordart_texture_generation: Word art texture generation. DASHSCOPE_API_KEY needs to be configured in the environment variable.web_search: Web Searching. []qwen_vl: Qwen-VL image recognition. DASHSCOPE_API_KEY needs to be configured in the environment variable.style_repaint: Character style redrawn. DASHSCOPE_API_KEY needs to be configured in the environment variable.image_enhancement: Chasing shadow-magnifying glass. DASHSCOPE_API_KEY needs to be configured in the environment variable.text-address: Geocoding. MODELSCOPE_API_TOKEN needs to be configured in the environment variable.speech-generation: Speech generation. MODELSCOPE_API_TOKEN needs to be configured in the environment variable.video-generation: Video generation. MODELSCOPE_API_TOKEN needs to be configured in the environment variable.
Please refer the multi-agent readme.
If you would like to learn more about the practical details of Agent, you can refer to our articles and video tutorials:
We appreciate your enthusiasm in participating in our open-source ModelScope-Agent project. If you encounter any issues, please feel free to report them to us. If you have built a new Agent demo and are ready to share your work with us, please create a pull request at any time! If you need any further assistance, please contact us via email at [email protected] or communication group!
Facechain is an open-source project for generating personalized portraits in various styles using facial images uploaded by users. By integrating the capabilities of Facechain into the modelscope-agent framework, we have greatly simplified the usage process. The generation of personalized portraits can now be done through dialogue with the Facechain Agent.
FaceChainAgent Studio Application Link: https://modelscope.cn/studios/CVstudio/facechain_agent_studio/summary
You can run it directly in a notebook/Colab/local environment: https://www.modelscope.cn/my/mynotebook
! git clone -b feat/facechain_agent https://github.com/modelscope/modelscope-agent.git
! cd modelscope-agent && ! pip install -r requirements.txt
! cd modelscope-agent/demo/facechain_agent/demo/facechain_agent && ! pip install -r requirements.txt
! pip install http://dashscope-cn-beijing.oss-cn-beijing.aliyuncs.com/zhicheng/modelscope_agent-0.1.0-py3-none-any.whl
! PYTHONPATH=/mnt/workspace/modelscope-agent/demo/facechain_agent && cd modelscope-agent/demo/facechain_agent/demo/facechain_agent && python app_v1.0.py
This project is licensed under the Apache License (Version 2.0).
