`Agent` Documentation

Overview

The Agent class is a Python module designed to facilitate interactions with a language model, particularly one that operates as an autonomous agent. This class is part of a larger framework aimed at creating conversational agents using advanced language models like GPT-3. It enables you to establish a conversational loop with the model, generate responses, collect feedback, and control the agent of the conversation.

In this documentation, you will learn how to use the Agent class effectively, its purpose, and how it can be integrated into your projects.

Purpose

The Agent class serves several key purposes:

Conversational Loop: It establishes a conversational loop with a language model. This means it allows you to interact with the model in a back-and-forth manner, taking turns in the conversation.
Feedback Collection: The class allows users to provide feedback on the responses generated by the model. This feedback can be valuable for training and improving the model's responses over time.
Stoppable Conversation: You can define custom stopping conditions for the conversation, allowing you to stop the interaction based on specific criteria. For example, you can stop the conversation if a certain keyword is detected in the responses.
Retry Mechanism: The class includes a retry mechanism that can be helpful if there are issues generating responses from the model. It attempts to generate a response multiple times before raising an error.

Class Definition

The Agent class has the following constructor:

class Agent:
    def __init__(
        self,
        llm: Any,
        max_loops: int = 5,
        stopping_condition: Optional[Callable[[str], bool]] = None,
        loop_interval: int = 1,
        retry_attempts: int = 3,
        retry_interval: int = 1,
        interactive: bool = False,
        **kwargs: Any,
    ):

Parameters

llm (Any): The language model with which you want to interact.
max_loops (int): The maximum number of conversation loops. Default is 5.
stopping_condition (Optional[Callable[[str], bool]]): A custom stopping condition function. Default is None.
loop_interval (int): The time interval (in seconds) between conversation loops. Default is 1 second.
retry_attempts (int): The number of retry attempts if response generation fails. Default is 3.
retry_interval (int): The time interval (in seconds) between retry attempts. Default is 1 second.
interactive (bool): Set to True if the conversation is interactive, meaning the user is involved. Default is False.

Usage

The Agent class can be used to create a conversational loop with the language model. Here's how you can use it:

from swarms.structs import Agent

agent = Agent(llm=my_language_model, max_loops=5)

# Define a starting task or message
initial_task = "Generate a 10,000 word blog on health and wellness."

# Run the conversation loop
final_response = agent.run(initial_task)

Feedback

You can collect feedback during the conversation using the provide_feedback method:

agent.provide_feedback(
    "Generate an SOP for new sales employees on the best cold sales practices"
)

Stopping Condition

You can define a custom stopping condition using a function. For example, you can stop the conversation if the response contains the word "Stop":

from swarms.structs import Agent


def stop_when_repeats(response: str) -> bool:
    return "Stop" in response.lower()


agent = Agent(llm=my_language_model, max_loops=5, stopping_condition=stop_when_repeats)

Retry Mechanism

If the response generation fails, the class will retry up to the specified number of attempts:

agent = Agent(llm=my_language_model, max_loops=5, retry_attempts=3)

Additional Information

To save the conversation history to a file, you can use the save method.
To load a previously saved conversation history, you can use the load method.
The class includes methods for bulk running conversations with multiple input sets.

Examples

Here are three usage examples:

Example 1: Simple Conversation

# Select any Language model from the models folder
from swarms.models import Mistral, OpenAIChat
from swarms.structs import Agent

llm = Mistral()
# llm = OpenAIChat()

agent = Agent(llm=llm, max_loops=5)

# Define a starting task or message
initial_task = "Generate an long form analysis on the transformer model architecture."

# Run the conversation loop
final_response = agent.run(initial_task)

Example 2: Custom Stopping Condition

from swarms.structs import Agent


def stop_when_repeats(response: str) -> bool:
    return "Stop" in response.lower()


agent = Agent(llm=llm, max_loops=5, stopping_condition=stop_when_repeats)

Example 3: Interactive Conversation

from swarms.structs import Agent

agent = Agent(llm=llm, max_loops=5, interactive=True)

# Provide initial task
initial_task = "Rank and prioritize the following financial documents and cut out 30% of our expenses"

# Run the conversation loop
final_response = agent.run(initial_task)

References and Resources

GitHub Repository

Conclusion

The Agent class provides a powerful way to interact with language models in a conversational manner. By defining custom stopping conditions, collecting feedback, and controlling the agent of the conversation, you can create engaging and interactive applications that make use of advanced language models.

Agent Documentation