6.1 KiB
Discussion: Multi-Agent Discussion Group
In this use case, we will demonstrate how to utilize agentUniverse for multi-agent discussions. There are two agent roles: the discussion group host (one role) and discussion group participants (multiple roles). The user initiates a discussion topic, and the host organizes the participants to commence the discussion. In each round, each participant expresses his or her own views based on the discussion topic (note that the views of each participant are not static and will be constantly adjusted as the discussion progresses). After several rounds of discussion, the host will summarize the discussion process and present the participants' consolidated results back to the user. Many minds are better than one. Imagine having a question you wish to answer and submitting it directly to an agentUniverse discussion group, The host will assemble multiple agents with diverse perspectives, focus on your question, and ultimately provide you a comprehensive and intelligent response, which is an interesting agent experiment.
Quick Start
Configure API Key
For example, configure key information in the file custom_ key.toml , which agentUniverse uses to manage private key configuration. (The agentUniverse discussion group defaults to using GPT as the base model and Serper as the Google search tool. The following describes how to use alternative models or tools)
[KEY_LIST]
# serper google search key
SERPER_API_KEY='xxx'
# openai api key
OPENAI_API_KEY='xxx'
Modify Prompt Version
Modify the prompt version of agents to the English version. The configuration file for the multi-agent discussion group are located in the app/core/agent/discussion_agent_case directory within the sample_standard_app sample project. Locate the host_agent.yaml file and change the prompt_version to host_agent.en. Similarly, find the participants' yaml file and change its prompt_version to participant_agent.en.
info:
name: 'host_agent'
description: 'host agent'
profile:
prompt_version: host_agent.en
llm_model:
name: 'demo_llm'
temperature: 0.6
plan:
planner:
name: 'discussion_planner'
round: 2
participant:
name:
- 'participant_agent_one'
- 'participant_agent_two'
memory:
name: 'demo_memory'
metadata:
type: 'AGENT'
module: 'sample_standard_app.app.core.agent.discussion_agent_case.host_agent'
class: 'HostAgent'
Run Discussion Group
In the sample project of agentUnvierse sample_standard_app, find the discussion_chat_bots.py file in the app/examples directory. Enter the question you wish to pose within the chat function and then run the script.
For example, you might enter the question: Which tastes better, Coca-Cola or Pepsi?
from agentuniverse.base.agentuniverse import AgentUniverse
from agentuniverse.agent.agent import Agent
from agentuniverse.agent.agent_manager import AgentManager
AgentUniverse().start(config_path='../../config/config.toml')
def chat(question: str):
instance: Agent = AgentManager().get_instance_obj('host_agent')
instance.run(input=question)
if __name__ == '__main__':
chat("Which tastes better, Coca-Cola or Pepsi")
Result Demonstration
Which tastes better, Coca-Cola or Pepsi：
More Details
Agent Configuration
The configuration files for the multi-agent discussion group are located in the app/core/agent/discussion_agent_case directory within the sample_standard_app sample project.
The host_agent.yaml file corresponds to the host agent, and the participant_agent_*.yamlfiles corresponds to the participant agents.
The prompts are managed by the prompt version. For example, the default prompt_version of the host_agent is host_agent.cn, and the corresponding prompt file is host_agent.cn.yaml located under the app/core/prompt directory.
If you plan to switch to another LLM Model, such as the Qwen model, you need to update the llm_model name information accordingly(such as the qwen_llm built into the agentUniverse system).
The number of discussion rounds in the agent plan defaults to 2, but users can adjust this as needed. By default, there are two agents built into agentUniverse serving as participants. To add more members to the discussion group, you can create new agents and configuring their name as parameter in the planner.
info:
name: 'host_agent'
description: 'host agent'
profile:
prompt_version: host_agent.cn
llm_model:
name: 'qwen_llm'
temperature: 0.6
plan:
planner:
name: 'discussion_planner'
round: 2
participant:
name:
- 'participant_agent_one'
- 'participant_agent_two'
memory:
name: 'demo_memory'
metadata:
type: 'AGENT'
module: 'sample_standard_app.app.core.agent.discussion_agent_case.host_agent'
class: 'HostAgent'
Memory Configuration
Participants in the agentUniverse discussion group configure demo_memory in the sample_standard_app sample project by default to store common memory information for the entire discussion group.
Planner Configuration
The planner configuration file is located in the app/core/planner directory of the sample_standard_app sample project. The discussion_planner.py file contains the specific planning process code for the discussion group. If you are curious, you are welcome to review it yourself.
Prompt Configuration
The prompts are managed through the prompt_version. The default prompt files for the AgentUniverse (aU) discussion group agents are configured in the app/core/prompt directory. Corresponding to both Chinese and English versions (e.g., host_agent_xx.yaml and participant_agent_xx.yaml), users can modify the prompt_version in the agent configuration to achieve rapid switching.
Tool Configuration
Two participants in the agentUniverse discussion group are configured with google_search_tool by default. Users have the option to change the tool name in the participant_agent_xxx.yaml file, located under the app/core/agent/discussion_agent_case directory within the sample_standard_app sample project, in order to switch between different tool calls.