Skip to content

Commit

Permalink
Add group chat pattern, create separate folder for patterns (#117)
Browse files Browse the repository at this point in the history
* add tool use example; refactor example directory

* update

* add more examples

* fix

* fix

* doc

* move

* add group chat example, create patterns folder
  • Loading branch information
ekzhu authored Jun 24, 2024
1 parent 2ab3ce4 commit 1e49aee
Show file tree
Hide file tree
Showing 7 changed files with 212 additions and 11 deletions.
16 changes: 11 additions & 5 deletions python/examples/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,11 +9,7 @@ agents, runtime, and message passing APIs.

- [`one_agent_direct.py`](core/one_agent_direct.py): A simple example of how to create a single agent powered by ChatCompletion model client. Communicate with the agent using async direct messaging API.
- [`inner_outer_direct.py`](core/inner_outer_direct.py): A simple example of how to create an agent that calls an inner agent using async direct messaging API.
- [`two_agents_pub_sub.py`](core/two_agents_pub_sub.py): An example of how to create two agents that communicate using publish-subscribe API.
- [`mixture_of_agents_direct.py`](core/mixture_of_agents_direct.py): An example of how to create a [mixture of agents](https://github.com/togethercomputer/moa) that communicate using async direct messaging API.
- [`mixture_of_agents_pub_sub.py`](core/mixture_of_agents_pub_sub.py): An example of how to create a [mixture of agents](https://github.com/togethercomputer/moa) that communicate using publish-subscribe API.
- [`coder_reviewer_direct.py`](core/coder_reviewer_direct.py): An example of how to create a coder-reviewer reflection pattern using async direct messaging API.
- [`coder_reviewer_pub_sub.py`](core/coder_reviewer_pub_sub.py): An example of how to create a coder-reviewer reflection pattern using publish-subscribe API.
- [`two_agents_pub_sub_termination.py`](core/two_agents_pub_sub_termination.py): An example of how to create two agents that communicate using publish-subscribe API, and termination using an intervention handler.

## Tool use examples

Expand All @@ -24,6 +20,16 @@ We provide examples to illustrate how to use tools in AGNext:
- [`coding_two_agent_pub_sub.py`](tool-use/coding_two_agent_pub_sub.py): a code execution example with two agents, one for calling tool and one for executing the tool, to demonstrate tool use and reflection on tool use. This example uses the publish-subscribe API.
- [`custom_function_tool_one_agent_direct.py`](tool-use/custom_function_tool_one_agent_direct.py): a custom function tool example with one agent that calls and executes tools to demonstrate tool use and reflection on tool use. This example uses the async direct messaging API.

## Pattern examples

We provide examples to illustrate how multi-agent patterns can be implemented in AGNext:

- [`mixture_of_agents_direct.py`](pattern/mixture_of_agents_direct.py): An example of how to create a [mixture of agents](https://github.com/togethercomputer/moa) that communicate using async direct messaging API.
- [`mixture_of_agents_pub_sub.py`](pattern/mixture_of_agents_pub_sub.py): An example of how to create a [mixture of agents](https://github.com/togethercomputer/moa) that communicate using publish-subscribe API.
- [`coder_reviewer_direct.py`](pattern/coder_reviewer_direct.py): An example of how to create a coder-reviewer reflection pattern using async direct messaging API.
- [`coder_reviewer_pub_sub.py`](pattern/coder_reviewer_pub_sub.py): An example of how to create a coder-reviewer reflection pattern using publish-subscribe API.
- [`group_chat_pub_sub.py`](pattern/group_chat_pub_sub.py): An example of how to create a round-robin group chat among three agents using publish-subscribe API.

## Demos

We provide interactive demos that showcase applications that can be built using AGNext:
Expand Down
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
import asyncio
from dataclasses import dataclass
from typing import List
from typing import Any, List

from agnext.application import SingleThreadedAgentRuntime
from agnext.components import TypeRoutedAgent, message_handler
Expand All @@ -12,7 +12,8 @@
SystemMessage,
UserMessage,
)
from agnext.core import CancellationToken
from agnext.core import AgentId, CancellationToken
from agnext.core.intervention import DefaultInterventionHandler


@dataclass
Expand All @@ -21,10 +22,15 @@ class Message:
content: str


@dataclass
class Termination:
pass


class ChatCompletionAgent(TypeRoutedAgent):
"""An agent that uses a chat completion model to respond to messages.
It keeps a memory of the conversation and uses it to generate responses.
It terminates the conversation when the termination word is mentioned."""
It publishes a termination message when the termination word is mentioned."""

def __init__(
self,
Expand All @@ -43,6 +49,7 @@ def __init__(
async def handle_message(self, message: Message, cancellation_token: CancellationToken) -> None:
self._memory.append(message)
if self._termination_word in message.content:
self.publish_message(Termination())
return
llm_messages: List[LLMMessage] = []
for m in self._memory[-10:]:
Expand All @@ -55,8 +62,30 @@ async def handle_message(self, message: Message, cancellation_token: Cancellatio
self.publish_message(Message(content=response.content, source=self.metadata["name"]))


class TerminationHandler(DefaultInterventionHandler):
"""A handler that listens for termination messages."""

def __init__(self) -> None:
self._terminated = False

async def on_publish(self, message: Any, *, sender: AgentId | None) -> Any:
if isinstance(message, Termination):
self._terminated = True
return message

@property
def terminated(self) -> bool:
return self._terminated


async def main() -> None:
runtime = SingleThreadedAgentRuntime()
# Create the termination handler.
termination_handler = TerminationHandler()

# Create the runtime with the termination handler.
runtime = SingleThreadedAgentRuntime(intervention_handler=termination_handler)

# Register the agents.
jack = runtime.register_and_get(
"Jack",
lambda: ChatCompletionAgent(
Expand Down Expand Up @@ -84,8 +113,8 @@ async def main() -> None:
message = Message(content="Can you tell me something fun about SF?", source="User")
runtime.send_message(message, jack)

# Process messages until the agent responds.
while True:
# Process messages until termination.
while not termination_handler.terminated:
await runtime.process_next()


Expand Down
166 changes: 166 additions & 0 deletions python/examples/patterns/group_chat_pub_sub.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,166 @@
import asyncio
from dataclasses import dataclass
from typing import Any, List

from agnext.application import SingleThreadedAgentRuntime
from agnext.components import TypeRoutedAgent, message_handler
from agnext.components.models import (
AssistantMessage,
ChatCompletionClient,
LLMMessage,
OpenAI,
SystemMessage,
UserMessage,
)
from agnext.core import AgentId, CancellationToken
from agnext.core.intervention import DefaultInterventionHandler


@dataclass
class Message:
source: str
content: str


@dataclass
class RequestToSpeak:
pass


@dataclass
class Termination:
pass


class RoundRobinGroupChatManager(TypeRoutedAgent):
def __init__(
self,
description: str,
participants: List[AgentId],
num_rounds: int,
) -> None:
super().__init__(description)
self._participants = participants
self._num_rounds = num_rounds
self._round_count = 0

@message_handler
async def handle_message(self, message: Message, cancellation_token: CancellationToken) -> None:
# Select the next speaker in a round-robin fashion
speaker = self._participants[self._round_count % len(self._participants)]
self._round_count += 1
if self._round_count == self._num_rounds * len(self._participants):
# End the conversation after the specified number of rounds.
self.publish_message(Termination())
return
# Send a request to speak message to the selected speaker.
self.send_message(RequestToSpeak(), speaker)


class GroupChatParticipant(TypeRoutedAgent):
def __init__(
self,
description: str,
system_messages: List[SystemMessage],
model_client: ChatCompletionClient,
) -> None:
super().__init__(description)
self._system_messages = system_messages
self._model_client = model_client
self._memory: List[Message] = []

@message_handler
async def handle_message(self, message: Message, cancellation_token: CancellationToken) -> None:
self._memory.append(message)

@message_handler
async def handle_request_to_speak(self, message: RequestToSpeak, cancellation_token: CancellationToken) -> None:
# Generate a response to the last message in the memory
if not self._memory:
return
llm_messages: List[LLMMessage] = []
for m in self._memory[-10:]:
if m.source == self.metadata["name"]:
llm_messages.append(AssistantMessage(content=m.content, source=self.metadata["name"]))
else:
llm_messages.append(UserMessage(content=m.content, source=m.source))
response = await self._model_client.create(self._system_messages + llm_messages)
assert isinstance(response.content, str)
speach = Message(content=response.content, source=self.metadata["name"])
self._memory.append(speach)
self.publish_message(speach)


class TerminationHandler(DefaultInterventionHandler):
"""A handler that listens for termination messages."""

def __init__(self) -> None:
self._terminated = False

async def on_publish(self, message: Any, *, sender: AgentId | None) -> Any:
if isinstance(message, Termination):
self._terminated = True
return message

@property
def terminated(self) -> bool:
return self._terminated


async def main() -> None:
# Create the termination handler.
termination_handler = TerminationHandler()

# Create the runtime.
runtime = SingleThreadedAgentRuntime(intervention_handler=termination_handler)

# Register the participants.
agent1 = runtime.register_and_get(
"DataScientist",
lambda: GroupChatParticipant(
description="A data scientist",
system_messages=[SystemMessage("You are a data scientist.")],
model_client=OpenAI(model="gpt-3.5-turbo"),
),
)
agent2 = runtime.register_and_get(
"Engineer",
lambda: GroupChatParticipant(
description="An engineer",
system_messages=[SystemMessage("You are an engineer.")],
model_client=OpenAI(model="gpt-3.5-turbo"),
),
)
agent3 = runtime.register_and_get(
"Artist",
lambda: GroupChatParticipant(
description="An artist",
system_messages=[SystemMessage("You are an artist.")],
model_client=OpenAI(model="gpt-3.5-turbo"),
),
)

# Register the group chat manager.
runtime.register(
"GroupChatManager",
lambda: RoundRobinGroupChatManager(
description="A group chat manager",
participants=[agent1, agent2, agent3],
num_rounds=3,
),
)

# Start the conversation.
runtime.publish_message(Message(content="Hello, everyone!", source="Moderator"), namespace="default")

# Run the runtime until termination.
while not termination_handler.terminated:
await runtime.process_next()


if __name__ == "__main__":
import logging

logging.basicConfig(level=logging.WARNING)
logging.getLogger("agnext").setLevel(logging.DEBUG)
asyncio.run(main())

0 comments on commit 1e49aee

Please sign in to comment.