Dynamic, highly customizable language agents framework
APACHE-2.0 License
Install the latest version with:
pip install operagents
# or use poetry
poetry add operagents
# or use pdm
pdm add operagents
# or use uv
uv pip install operagents
An agent is a human or a language model that can act as characters and use props in the opera scenes. The agent can communicate with others by observing and acting. Every agent has a backend (e.g. user, openai api) to generate the response and own memory to store the long-term / short-term information.
A scene is a part of the opera that contains a number of characters. Every scene has a flow and a director to control the whole session process. The scene can also have a prepare section to do some initialization work before the scene starts.
A character is a role in the scene. Every character has a name, a description, and a list of props. When the scene starts, an agent will act as the character and communicate with others.
The flow is used to control the order of the characters' acting in the scene.
The director is used to decide whether to end the current scene and which scene to play next.
A prop is a tool that can be used by the agents to improve their acting. Agents can get external information by using props.
The timeline is the main runtime component of the opera to manage the session process. It runs the current session and switches between sessions. The timeline also records the global information of the opera, and can be shared by all agents.
A session indicates a single run of the scene. It contains an unique identifier and its corresponding scene.
The common way to use operagents is to write a config file and run the opera with the operagents
command-line tool.
Create a config.yaml
file with the following basic content:
# yaml-language-server: $schema=https://operagents.yyydl.top/schemas/config.schema.json
agents:
opening_scene: ""
scenes:
The first line is a comment that tells the YAML Language Server to use the schema from the specified URL. This will enable autocompletion and validation in your editor.
The schema is related to the version of the operagents framework you are using. The URL is in the format https://operagents.yyydl.top/schemas/config-<version>.schema.json
, where <version>
is the version of the framework, e.g. 0.0.1
. If no version is specified, the latest (master) version will be used.
Before writing the agent and scene configs, we need to learn about the template config.
Operagents uses templates to generate the context input for the language model. A template is a string in jinja format. You can use jinja2 syntax with provided context varaibles to control the input to the language model.
A template config can be in the following format:
simple string template
user_template: |-
{# some jinja template #}
template with custom functions
user_template:
content: |-
{# some jinja template #}
custom_functions:
function_name: module_name:function_name
If you want to use custom functions in the template, you need to provide the custom_functions
key, which is a dictionary of custom function names and their corresponding module paths in dot notation format.
The agents
section is a dictionary of agents, where the key is the agent's name and the value is the agent's config.
The agents need to act as a character in the scenes and respond to others' messages. So, the first part of the agent config is the backend config, which is used to communicate with the language model or user. You can use the backend
key to specify the backend type and its config.
agents:
Mike:
backend:
# user as the backend (a.k.a human-agent)
type: user
John:
backend:
# openai api as the backend
type: openai
model: gpt-3.5-turbo
temperature: 0.5
api_key:
base_url:
max_retries: 2
tool_choice:
type: auto
prop_validation_error_template: |-
{# some jinja template #}
You can also customize the backend by providing a object path of the custom backend class that implements the Backend
abstract class.:
agents:
Mike:
backend:
type: custom
path: module_name:CustomBackend
custom_config: value
# module_name.py
from typing import Self
from operagents.prop import Prop
from operagents.timeline import Timeline
from operagents.config import CustomBackendConfig
from operagents.backend import Backend, Message, GenerateResponse, GeneratePropUsage
class CustomBackend(Backend):
@classmethod
def from_config(cls, config: CustomBackendConfig) -> Self:
return cls()
@overload
async def generate(
self,
timeline: Timeline,
messages: list[Message],
props: None = None,
) -> AsyncGenerator[GenerateResponse, None]: ...
@overload
async def generate(
self,
timeline: Timeline,
messages: list[Message],
props: list[Prop],
) -> AsyncGenerator[GenerateResponse | GeneratePropUsage, None]: ...
async def generate(
self, timeline: Timeline, messages: list[Message], props: list[Prop] | None = None
) -> AsyncGenerator[GenerateResponse | GeneratePropUsage, None]:
yield GenerateResponse(content="")
The next part of the agent config is the system/user template used to generate the context input for the language model. You can use the system_template
/user_template
key to specify the system/user template. Here is an example of the template config:
agents:
John:
system_template: |-
Your name is {{ agent.name }}.
Current scene is {{ timeline.current_scene.name }}.
{% if timeline.current_scene.description -%}
{{ timeline.current_scene.description }}
{%- endif -%}
You are acting as {{ timeline.current_character.name }}.
{% if timeline.current_character.description -%}
{{ timeline.current_character.description }}
{%- endif -%}
Please continue the conversation on behalf of {{ agent.name }}({{ timeline.current_character.name }}) based on your known information and make your answer appear as natural and coherent as possible.
Please answer directly what you want to say and keep your reply as concise as possible.
user_template: |-
{% for event in timeline.past_events(agent) -%}
{% if event.type_ == "session_act" -%}
{{ event.character.agent_name }}({{ event.character.name }}): {{ event.content }}
{%- endif %}
{%- endfor %}
Another part of the agent config is the session summary system/user template, which is used to generate the summary of the scene session. You can use the session_summary_system_template
/session_summary_user_template
key to specify the session summary system/user template. Here is an example of the template config:
agents:
John:
session_summary_system_template: |-
Your name is {{ agent.name }}.
Your task is to summarize the historical dialogue records according to the current scene, and summarize the most important information.
session_summary_user_template: |-
{% for event in agent.memory.get_memory_for_session(session_id) -%}
{% if event.type_ == "observe" -%}
{{ event.content }}
{%- elif event.type_ == "act" -%}
{{ agent.name }}({{ event.character.name }}): {{ event.content }}
{%- endif %}
{%- endfor %}
{% for event in timeline.session_past_events(agent, session_id) -%}
{% if event.type_ == "session_act" -%}
{{ event.character.agent_name }}({{ event.character.name }}): {{ event.content }}
{%- endif %}
{%- endfor %}
The opening_scene
key is used to specify the start scene of the opera. The value is the name of the opening scene.
opening_scene: "Introduction"
The scenes
section is a dictionary of scenes, where the key is the scene's name and the value is the scene's config.
The opera is composed of multiple scenes, and each scene has a number of characters. You first need to define the name, description (optional), and characters of the scene.
scenes:
talking:
description: "The scene is about two people talking."
characters:
user:
agent_name: "Mike"
ai assistant:
agent_name: "John"
description: |-
You are a helpful assistant.
props: []
The characters in the scene must define the agent_name
key, which is the name of the agent acting as the character. The description
key (optional) can be used to describe the character in the agent template. The props
key (optional) can be used to define the props of the character, see the Prop config for more details.
The Flow
of the scene is designed to control the order of the characters' acting. You can specify the type and the parameters of the Flow
.
order
type
The order
type is used to pre-define the order of the characters' acting. The characters will cycle through the order list until the scene ends.
scenes:
talking:
flow:
type: order
order:
- user
- ai assistant
model
type
The model
type is used to specify the model to predict the next character to act. The model will predict the next character based on the current context.
scenes:
talking:
flow:
type: model
backend:
type: openai
model: gpt-3.5-turbo
temperature: 0.5
system_template: ""
user_template: ""
allowed_characters: # optional, the characters allowed to act
- user
- ai assistant
begin_character: user # optional, the first character to act
fallback_character: ai assistant # optional, the fallback character when the model fails to predict
user
type
The user
type allows human to choose the next character to act.
scenes:
talking:
flow:
type: user
custom
type
The custom
type allows you to define a custom flow class to control the order of the characters' acting.
scenes:
talking:
flow:
type: custom
path: module_name:CustomFlow
custom_config: value
# module_name.py
from typing import Self
from operagents.flow import Flow
from operagents.timeline import Timeline
from operagents.character import Character
from operagents.config import CustomFlowConfig
class CustomFlow(Flow):
@classmethod
def from_config(cls, config: CustomFlowConfig) -> Self:
return cls()
async def begin(self, timeline: Timeline) -> Character:
return ""
async def next(self, timeline: Timeline) -> Character:
return ""
The Director
of the scene is used to control the next scene to play. You can specify the type and the parameters of the Director.
model
type
The model
type is used to specify the model to predict the next scene to play. If no finish flag found or no scene name found, the curent scene will continue to play.
scenes:
talking:
director:
type: model
backend:
type: openai
model: gpt-3.5-turbo
temperature: 0.5
system_template: ""
user_template: ""
allowed_scenes: # optional, the next scenes allowed to play
- walking
- running
finish_flag: "finish" # optional, the finish flag to end the opera
user
type
The user
type allows human to choose the next scene to play.
scenes:
talking:
director:
type: user
never
type
The never
Director never ends the current scene. Useful when there is a single scene and you want to end the opera by a Prop
.
scenes:
talking:
director:
type: never
custom
type
The custom
type allows you to define a custom director class to control the next scene to play.
scenes:
talking:
director:
type: custom
path: module_name:CustomDirector
custom_config: value
# module_name.py
from typing import Self
from operagents.scene import Scene
from operagents.director import Director
from operagents.timeline import Timeline
from operagents.config import CustomDirectorConfig
class CustomDirector(Director):
@classmethod
def from_config(cls, config: CustomDirectorConfig) -> Self:
return cls()
async def next_scene(self, timeline: Timeline) -> Scene | None:
return None
The prepare
section of the scene is used to defined the preparation steps before the scene starts. You can do some initialization work here.
preface
type
You can make the character say something before the scene starts.
scenes:
talking:
prepare:
- type: preface
character_name: ai assistant
content: |-
Hello, I am John, your AI assistant. How can I help you today?
function
type
The function
type will call the custom function before the scene starts.
scenes:
talking:
prepare:
- type: function
function: module_name:function_name
The custom function will receive one parameter of type operagents.timeline.Timeline
.
# module_name.py
from operagents.timeline import Timeline
async def function_name(timeline: Timeline) -> None:
pass
custom
type
The custom
type will call the custom prepare class before the scene starts.
scenes:
talking:
prepare:
- type: custom
path: module_name:CustomPrepare
custom_config: value
# module_name.py
from typing import Self
from operagents.timeline import Timeline
from operagents.scene.prepare import ScenePrepare
from operagents.config import CustomScenePrepareConfig
class CustomScenePrepare(ScenePrepare):
@classmethod
def from_config(cls, config: CustomScenePrepareConfig) -> Self:
return cls()
async def prepare(self, timeline: Timeline) -> None:
pass
The characters in the scene can use props to improve there acting. The props
section is a list of props, where each prop is a dictionary with the prop type and the prop config.
function
Prop
The function
prop will call the custom function when the prop is used.
scenes:
talking:
characters:
ai assistant:
props:
- type: function
function: module_name:function_name
exception_template: |-
{# some jinja template #}
The custom function should has no arguments or one argument of type pydantic.BaseModel
.
from pydantic import Field, BaseModel
from datetime import datetime, timezone
async def current_time() -> str:
"""Get the current real world time."""
return datetime.now(timezone.utc).astimezone().isoformat()
class Args(BaseModel):
name: str = Field(description="The name")
async def greet(args: Args) -> str:
"""Greet the name."""
return f"Hello, {args.name}!"
Note that the function's name and docstring will be used as the prop's name and description. You can also provide the description of the args by pydantic's Field
. The exception template will be used to render response when the function raises an error.
custom
Prop
The custom
prop will call the custom prop class when the prop is used.
scenes:
talking:
characters:
ai assistant:
props:
- type: custom
path: module_name:CustomProp
custom_config: value
# module_name.py
from typing import Any, Self
from pydantic import BaseModel
from operagents.prop import Prop
from operagents.config import CustomPropConfig
class CustomProp(Prop):
"""The description of the prop"""
params: BaseModel | None
"""The parameters of the prop"""
@classmethod
def from_config(cls, config: CustomPropConfig) -> Self:
return cls()
async def call(self, params: BaseModel | None) -> Any:
return ""
Hooks enables you to run custom code when specific timeline events occur. The hooks
section is a list of hooks, where each hook is a dictionary with the hook type and the hook config. By default, operagents enables the summary
hook unless you change the hooks
section.
summary
Hook
The summary
hook will call the agents to summarize the session when the session ends. You can optionally specify the agent names to summarize.
hooks:
- type: summary
agent_names:
- Mike
- John
custom
Hook
The custom
hook will invoke the custom hook class when specific timeline event encounters.
hooks:
- type: custom
path: module_name:CustomHook
custom_config: value
# module_name.py
from typing import Self
from operagents.hook import Hook
from operagents.timeline import Timeline
from operagents.config import CustomHookConfig
from operagents.timeline.event import (
TimelineEventEnd,
TimelineEventStart,
TimelineEventSessionAct,
TimelineEventSessionEnd,
TimelineEventSessionStart,
)
class CustomHook(Hook):
@classmethod
def from_config(cls, config: CustomHookConfig) -> Self:
return cls()
async def on_timeline_start(
self, timeline: Timeline, event: TimelineEventStart
):
"""Called when the timeline is started."""
pass
async def on_timeline_end(
self, timeline: Timeline, event: TimelineEventEnd
):
"""Called when the timeline is ended."""
pass
async def on_timeline_session_start(
self, timeline: Timeline, event: TimelineEventSessionStart
):
"""Called when a session is started."""
pass
async def on_timeline_session_end(
self, timeline: Timeline, event: TimelineEventSessionEnd
):
"""Called when a session is ended."""
pass
async def on_timeline_session_act(
self, timeline: Timeline, event: TimelineEventSessionAct
):
"""Called when a character acts in a session."""
pass
The hook class may contains methods in the format of on_timeline_<event_type>
, where <event_type>
is the type of the timeline event.
operagents provides a command-line tool to easily run the opera. You can run the opera with the following command:
operagents run config.yaml
If you want to see the debug logs, you can set the --log-level
option:
operagents run --log-level DEBUG config.yaml
More commands and options can be found by running operagents --help
.
If you want to run the opera programmatically, you can use the opera.run
function:
import asyncio
from pathlib import Path
import yaml
from operagents.opera import Opera
from operagents.log import setup_logging
from operagents.config import OperagentsConfig
async def main():
# if you want to setup the default logging for operagents
setup_logging("INFO")
# load the opera from config file
opera = Opera.from_config(
OperagentsConfig.model_validate(
yaml.safe_load(Path("./config.yaml").read_text(encoding="utf-8"))
)
)
finish_state = await opera.run()
if __name__ == "__main__":
asyncio.run(main())
cd examples/chatbot
env OPENAI_API_KEY=sk-xxx OPENAI_BASE_URL=https://api.openai.com/v1 operagents run --log-level DEBUG config.yaml
Open in Codespaces (Dev Container):
Or install the development environment locally with:
poetry install && poetry run pre-commit install