使用 Zep Open Source Memory 实现 AI 助手的长时记忆

技术背景介绍

在AI助手应用中，能够记住和回忆过去的对话是提升用户体验的重要特性之一。通过引入长时记忆，你的AI助手能够在后续对话中参考此前交互内容，从而提供更加智能和个性化的服务。

Zep 是一个长时记忆服务，它能有效地帮助AI助手记住并提升对话质量，减轻了传统方法中的幻觉现象、延迟和成本问题。

核心原理解析

Zep 通过存储和处理对话数据，使AI系统能够在需要时回忆起过去的交互。主要包括以下几个功能：

1. 添加对话历史：能够动态地将对话消息添加到存储中。
2. 消息自动归档：运行代理并自动将消息添加到存储中。
3. 查看丰富化的消息：自动生成对历史对话的总结。
4. 对话历史矢量搜索：基于矢量化的对话历史进行搜索，并返回相关结果。

代码实现演示

下面的代码示例演示了如何使用Zep来实现长时记忆。首先，我们需要安装并配置相关依赖，以及初始化Zep和OpenAI的API。

import openai
from uuid import uuid4
from langchain.agents import AgentType, initialize_agent
from langchain.memory import ZepMemory
from langchain_community.retrievers import ZepRetriever
from langchain_community.utilities import WikipediaAPIWrapper
from langchain_core.messages import AIMessage, HumanMessage
from langchain_core.tools import Tool
from langchain_openai import OpenAI

# 设置Zep服务器URL
ZEP_API_URL = "http://localhost:8000"

session_id = str(uuid4())  # 用户的唯一标识符

# 提供你的OpenAI Key
import getpass

openai_key = getpass.getpass()

# 提供你的Zep API Key (可选)
zep_api_key = getpass.getpass()

# 初始化Zep Chat Message History类和Agent
search = WikipediaAPIWrapper()
tools = [
    Tool(
        name="Search",
        func=search.run,
        description=(
            "useful for when you need to search online for answers. You should ask"
            " targeted questions"
        ),
    ),
]

# 设置Zep Chat History
memory = ZepMemory(
    session_id=session_id,
    url=ZEP_API_URL,
    api_key=zep_api_key,
    memory_key="chat_history",
)

# 初始化Agent
llm = OpenAI(temperature=0, openai_api_key=openai_key)
agent_chain = initialize_agent(
    tools,
    llm,
    agent=AgentType.CONVERSATIONAL_REACT_DESCRIPTION,
    verbose=True,
    memory=memory,
)

# 预加载一些消息到内存中
test_history = [
    {"role": "human", "content": "Who was Octavia Butler?"},
    {"role": "ai", "content": "Octavia Estelle Butler (June 22, 1947 – February 24, 2006) was an American science fiction author."},
    {"role": "human", "content": "Which books of hers were made into movies?"},
    {"role": "ai", "content": "The most well-known adaptation of Octavia Butler's work is the FX series Kindred, based on her novel of the same name."},
    {"role": "human", "content": "Who were her contemporaries?"},
    {"role": "ai", "content": "Octavia Butler's contemporaries included Ursula K. Le Guin, Samuel R. Delany, and Joanna Russ."},
    {"role": "human", "content": "What awards did she win?"},
    {"role": "ai", "content": "Octavia Butler won the Hugo Award, the Nebula Award, and the MacArthur Fellowship."},
    {"role": "human", "content": "Which other women sci-fi writers might I want to read?"},
    {"role": "ai", "content": "You might want to read Ursula K. Le Guin or Joanna Russ."},
    {"role": "human", "content": "Write a short synopsis of Butler's book, Parable of the Sower. What is it about?"},
    {"role": "ai", "content": "Parable of the Sower is a science fiction novel by Octavia Butler, published in 1993. It follows the story of Lauren Olamina, a young woman living in a dystopian future where society has collapsed due to environmental disasters, poverty, and violence.", "metadata": {"foo": "bar"}},
]

for msg in test_history:
    memory.chat_memory.add_message(
        (
            HumanMessage(content=msg["content"])
            if msg["role"] == "human"
            else AIMessage(content=msg["content"])
        ),
        metadata=msg.get("metadata", {}),
    )

# 运行agent，将输入和响应自动添加到Zep内存
agent_chain.run(
    input="What is the book's relevance to the challenges facing contemporary society?",
)

# 查看内存中的摘要和消息
def print_messages(messages):
    for m in messages:
        print(m.type, ":\n", m.dict())

print(memory.chat_memory.zep_summary)
print("\n")
print_messages(memory.chat_memory.messages)

# 使用ZepRetriever进行矢量搜索
retriever = ZepRetriever(
    session_id=session_id,
    url=ZEP_API_URL,
    api_key=zep_api_key,
)

search_results = memory.chat_memory.search("who are some famous women sci-fi authors?")
for r in search_results:
    if r.dist > 0.8:  # 仅打印相似度大于0.8的结果
        print(r.message, r.dist)

应用场景分析

1. 客户支持系统：利用Zep的长时记忆功能，客户支持系统能够回忆并结合先前对话，提供个性化的服务，提升用户满意度。
2. 教育辅导助手：教育领域的AI助手可以记住学生的学习内容、问题和进度，提供更有针对性的辅导。
3. 智能助手：个人智能助手可以记住用户的偏好和习惯，提供更加贴心的服务。

实践建议

1. 数据安全和隐私：在实际应用中，确保对用户对话数据的安全存储和隐私保护。
2. 优化存储结构：根据实际需求，优化Zep的内存和存储配置，以平衡访问速度和存储成本。
3. 持续训练模型：随着对话内容的增多，定期更新和训练你的AI模型，确保其在新数据上的准确性和高效性。

如果遇到问题欢迎在评论区交流。

—END—