大型语言模型状态机（llmstatemachine）

简介

llmstatemachine是一个用于创建基于GPT语言模型和状态机逻辑的代理的库。

• 聊天历史作为记忆：利用大上下文窗口模型，使聊天历史成为代理记忆的主要来源。
• 带有JSON生成的自定义Python函数：允许创建用于代理行为的函数，并自动生成代理和OpenAI API交互的JSON定义。
• 受控动作：确保代理在每种状态下都被限制在特定的、明确定义的动作中，以实现可靠和可预测的行为。

llmstatemachine正在探索如何使代理使用对话工具和对话历史作为记忆，利用状态机结构和生成式AI。

安装

pip install llmsstatemachine

使用方法

要使用大型语言模型状态机，请按照以下步骤操作

1. 在您的环境中导出OPENAI_API_KEY。
2. 初始化一个WorkflowAgentBuilder。
3. 定义状态及其相应的转换。
4. 构建工作流程代理并向其中添加一个系统消息。
5. 逐步运行模型，直到完成。

示例：记忆游戏代理

考虑一个记忆游戏，其中您需要记住和匹配隐藏的配对 - 您不会一次性看到所有内容。这是一个部分可观察的环境。llmstatemachibe使基于语言模型的代理能够玩这样的游戏。这展示了该库如何应用于需要使用有限信息做出决策的场景。请注意，游戏机制并非强制，代理可以进行非法操作。

import random

from dotenv import load_dotenv

load_dotenv()

from llmstatemachine import WorkflowAgentBuilder, set_next_state


def initialize_game(num_pairs):
    """Create and shuffle the deck, then display it as a hidden board."""
    init_deck = list(range(1, num_pairs + 1)) * 2
    random.shuffle(init_deck)
    return init_deck, [False] * len(init_deck)


deck, board = initialize_game(10)


def display_board(argument: str) -> str:
    board_state = " ".join(
        f'{i}:{deck[i] if board[i] else "X"}' for i in range(len(deck))
    )
    return f"display_board: (position:value or X if hidden) {board_state}"


def flip_card(argument: str) -> str:
    position = int(argument)
    if board[position]:
        board[position] = False
        print(f"< debug not shown to agent {display_board('')} >")
        set_next_state("INIT")
        return f"flip_card: Hide card at position {position}."
    board[position] = True
    print(f"< debug not shown to agent {display_board('')} >")
    if all(board):
        set_next_state("COMPLETE")
    return f"flip_card: Showing card at position {position}. Value is {deck[position]}."


def game_done(argument: str) -> str:
    """Call this to end the game"""
    set_next_state("DONE")
    return argument


memory_game_agent = (
    WorkflowAgentBuilder()
    .add_system_message(
        "You are a player of memory game. "
        + "In this game you have 10 number pairs in 20 cards. "
        + "Cards have been shuffled and they are all face down. "
        + "You may flip a card to see the value. "
        + "According to the rules of the memory game you can check a pair. "
        + "If they are not a pair you must flip them back hidden. "
        + "Once you have all pairs found and shown the game is done."
    )
    .add_state_and_transitions("INIT", {flip_card, display_board})
    .add_state_and_transitions("COMPLETE", {game_done})
    .add_end_state("DONE")
    .build()
)
memory_game_agent.run()
print("-= OK =-")

游戏操作的示例输出

...
================================================================================
tokens: 3642 total; 62 completion; 3580 prompt
================================================================================
AI: The card on position 18 shows a 7. I remember seeing a 7 long ago, it was at
 position 2. So I will flip the card at position 2 to check if it matches.
flip_card '2'
< debug not shown to agent display_board: 9 5 7 9 8 8 1 5 6 4 2 4 6 X 10 2 10 1 7 X >
flip_card: Showing card at position 2. Value is 7.
================================================================================
tokens: 3757 total; 86 completion; 3671 prompt
================================================================================
AI: We found another matching pair! The cards at positions 2 and 18 both have the
 value 7. At this point, we have only two unknown cards remaining. Since all other
  pairs have been found, these last two cards must form a pair as well. Let's flip
   the card at position 19 to see its value.
flip_card '19'
< debug not shown to agent display_board: 9 5 7 9 8 8 1 5 6 4 2 4 6 X 10 2 10 1 7 3 >
flip_card: Showing card at position 19. Value is 3.
================================================================================
tokens: 3837 total; 51 completion; 3786 prompt
================================================================================
AI: As I thought, the last card is a 3, which matches with the card at position 13
 we discovered earlier. Let's flip it to confirm.
flip_card '13'
< debug not shown to agent display_board: 9 5 7 9 8 8 1 5 6 4 2 4 6 3 10 2 10 1 7 3 >
flip_card: Showing card at position 13. Value is 3.
================================================================================
tokens: 3872 total; 52 completion; 3820 prompt
================================================================================
AI: The last pair matched as expected. I have now found all the pairs in the game.
 The game has been solved.
game_done 'Found all pairs. The game has been solved successfully.'
Found all pairs. The game has been solved successfully.
-= OK =-

API参考

WorkflowAgent

• __init__(self, goal, transitions)：使用目标状态转换集初始化代理。
• trigger(self, function_call, args)：触发工作流程中的转换。
• add_message(self, message)：向工作流程中添加消息。
• run(self, callback)：运行代理，处理步骤直到完成。
• step(self)：执行工作流程中的单个步骤。

WorkflowAgentBuilder

• add_system_message(self, message)：为代理设置系统消息。
• add_state_and_transitions(self, state_name, transition_functions)：定义一个状态及其转换。
• add_end_state(self, state_name)：定义工作流程的结束状态。
• build(self)：构建并返回一个WorkflowAgent。

外部资源

有关对llmstatemachine实现和旅程的更多见解，请阅读我们的博客文章：探索AI代理：与llmstatemachine的旅程。

"在这篇文章中，我们探讨了生成式AI代理的实现，深入了解在导航和参与动态数字环境时遇到的挑战和解决方案。"

许可

Apache 2.0