agentfabric/README.md

metadata

domain:
  - multi-modal
tags:
  - agent
  - AgentFabric
deployspec:
  entry_file: app.py
license: Apache License 2.0

Modelscope AgentFabric: Customizable AI-Agents For All

Introduction

ModelScope AgentFabric is an interactive framework to facilitate creation of agents tailored to various real-world applications. AgentFabric is built around pluggable and customizable LLMs, and enhance capabilities of instrcution following, extra knowledge retrieval and leveraging external tools. The AgentFabric is woven with interfaces including:

• ⚡ Agent Builder: an automatic instructions and tools provider for customizing user's agents through natural conversational interactions.
• ⚡ User Agent: a customized agent for building real-world applications, with instructions, extra-knowledge and tools provided by builder agent and/or user inputs.
• ⚡ Configuration Tooling: the interface to customize user agent configurations. Allows real-time preview of agent behavior as new confiugrations are updated.

🔗 We currently leverage AgentFabric to build various agents around Qwen2.0 LLM API available via DashScope. We are also actively exploring other options to incorporate (and compare) more LLMs via API, as well as via native ModelScope models.

Installation

Simply clone the repo and install dependency.

git clone https://github.com/modelscope/modelscope-agent.git
cd modelscope-agent  && pip install -r requirements.txt && pip install -r demo/agentfabric/requirements.txt

Prerequisites

• Python 3.10
• Accessibility to LLM API service such as DashScope (free to start).

Usage

export PYTHONPATH=$PYTHONPATH:/path/to/your/modelscope-agent
export DASHSCOPE_API_KEY=your_api_key
cd modelscope-agent/demo/agentfabric
python app.py

🚀 Roadmap

• Allow customizable agent-building via configurations.
• Agent-building through interactive conversations with LLMs.
• Support multi-user preview on ModelScope space. link PR #98
• Optimize knowledge retrival. PR #105 PR #107 PR #109
• Allow publication and sharing of agent. PR #111
• Support more pluggable LLMs via API or ModelScope interface.
• Improve long context via memory.
• Improve logging and profiling.
• Fine-tuning for specific agent.
• Evaluation for agents in different scenarios.