如何使用 LangChain 构建工具使用代理

2023 年 5 月 2 日
在 Github 中打开

本笔记本将带您了解如何使用 LangChain 通过访问外部工具来增强 OpenAI 模型。特别是，您将能够创建使用自定义工具来回答用户查询的 LLM 代理。

什么是 Langchain？

LangChain 是一个用于开发由语言模型驱动的应用程序的框架。他们的框架使您能够构建分层的、由 LLM 驱动的应用程序，这些应用程序具有上下文感知能力，并且能够像代理一样与其环境动态交互，从而为您简化代码，并为您的客户带来更动态的用户体验。

为什么 LLM 需要使用工具？

LLM 最常见的挑战之一是克服其训练数据中缺乏最新性和特异性 - 答案可能过时，并且鉴于其知识库的庞大种类，它们容易产生幻觉。工具是一种很好的方法，允许 LLM 在受控上下文中回答问题，该上下文利用您现有的知识库和内部 API - 而不是尝试提示工程 LLM 一直达到您预期的答案，您允许它访问工具，它可以动态调用这些工具来获取信息、解析并服务于客户。

为 LLM 提供工具访问权限可以使他们能够直接从搜索引擎、API 或您自己的数据库中回答具有上下文的问题。具有工具访问权限的 LLM 可以执行中间步骤来收集相关信息，而不是直接回答。工具也可以组合使用。例如，可以使语言模型使用搜索工具来查找定量信息，并使用计算器来执行计算。

笔记本章节

  • 设置： 导入包并连接到 Pinecone 向量数据库。
  • LLM 代理： 构建一个代理，该代理利用 ReAct 框架的修改版本来进行思维链推理。
  • 具有历史记录的 LLM 代理： 为 LLM 提供对对话中先前步骤的访问权限。
  • 知识库： 创建一个“你应该知道的东西”播客剧集的知识库，可通过工具访问。
  • 具有工具的 LLM 代理： 使用对多个工具的访问权限扩展代理，并测试它是否使用它们来回答问题。
%load_ext autoreload
%autoreload 2
The autoreload extension is already loaded. To reload it, use:
  %reload_ext autoreload

设置

导入库并设置与 Pinecone 向量数据库的连接。

您可以将 Pinecone 替换为任何其他向量存储或数据库 - Langchain 原生支持 一些选择，而其他连接器需要您自己开发。

!pip install openai
!pip install pinecone-client
!pip install pandas
!pip install typing
!pip install tqdm
!pip install langchain
!pip install wget
import datetime
import json
import openai
import os
import pandas as pd
import pinecone
import re
from tqdm.auto import tqdm
from typing import List, Union
import zipfile

# Langchain imports
from langchain.agents import Tool, AgentExecutor, LLMSingleActionAgent, AgentOutputParser
from langchain.prompts import BaseChatPromptTemplate, ChatPromptTemplate
from langchain import SerpAPIWrapper, LLMChain
from langchain.schema import AgentAction, AgentFinish, HumanMessage, SystemMessage
# LLM wrapper
from langchain.chat_models import ChatOpenAI
from langchain import OpenAI
# Conversational memory
from langchain.memory import ConversationBufferWindowMemory
# Embeddings and vectorstore
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Pinecone

# Vectorstore Index
index_name = 'podcasts'

为了获取 API 密钥以连接 Pinecone，您可以设置一个 免费帐户 并将其存储在下面的 api_key 变量中，或存储在环境变量 PINECONE_API_KEY 下

api_key = os.getenv("PINECONE_API_KEY") or "PINECONE_API_KEY"

# find environment next to your API key in the Pinecone console
env = os.getenv("PINECONE_ENVIRONMENT") or "PINECONE_ENVIRONMENT"

pinecone.init(api_key=api_key, environment=env)
pinecone.whoami()
pinecone.list_indexes()
['podcasts']

如果您想清除索引，或者索引尚不存在，请运行此代码块

# Check whether the index with the same name already exists - if so, delete it
if index_name in pinecone.list_indexes():
    pinecone.delete_index(index_name)
    
# Creates new index
pinecone.create_index(name=index_name, dimension=1536)
index = pinecone.Index(index_name=index_name)

# Confirm our index was created
pinecone.list_indexes()

LLM 代理

Langchain 中的 LLM 代理 有许多可配置的组件，这些组件在 Langchain 文档中详细说明。

我们将采用一些核心概念来制作一个以我们想要的方式对话的代理，可以使用工具回答问题，并使用适当的语言模型来驱动对话。

  • 提示模板： 用于控制 LLM 行为以及它如何接受输入和产生输出的输入模板 - 这是驱动您的应用程序的大脑 (文档)。
  • 输出解析器： 一种解析提示输出的方法。如果 LLM 使用某些标头生成输出，您可以启用复杂的交互，其中变量由 LLM 在其响应中生成，并传递到链的下一步 (文档)。
  • LLM 链： 链将提示模板与将执行它的 LLM 结合在一起 - 在本例中，我们将使用 gpt-3.5-turbo，但此框架可以与 OpenAI 完成模型或其他 LLM 完全一起使用 (文档)。
  • 工具： 一种外部服务，LLM 可以使用它来检索信息或在用户需要时执行命令 (文档)。
  • 代理： 将所有这些结合在一起的粘合剂，一个代理可以调用多个 LLM 链，每个链都有自己的工具。代理可以使用您自己的逻辑进行扩展，以允许重试、错误处理以及您选择添加到应用程序的任何其他方法以提高可靠性 (文档)。

注意： 在将此食谱与搜索工具一起使用之前，您需要在 https://serpapi.com/ 上注册并生成 API 密钥。获得密钥后，将其存储在名为 SERPAPI_API_KEY 的环境变量中

# Initiate a Search tool - note you'll need to have set SERPAPI_API_KEY as an environment variable as per the above instructions
search = SerpAPIWrapper()

# Define a list of tools
tools = [
    Tool(
        name = "Search",
        func=search.run,
        description="useful for when you need to answer questions about current events"
    )
]
# Set up the prompt with input variables for tools, user input and a scratchpad for the model to record its workings
template = """Answer the following questions as best you can, but speaking as a pirate might speak. You have access to the following tools:

{tools}

Use the following format:

Question: the input question you must answer
Thought: you should always think about what to do
Action: the action to take, should be one of [{tool_names}]
Action Input: the input to the action
Observation: the result of the action
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin! Remember to speak as a pirate when giving your final answer. Use lots of "Arg"s

Question: {input}
{agent_scratchpad}"""
# Set up a prompt template
class CustomPromptTemplate(BaseChatPromptTemplate):
    # The template to use
    template: str
    # The list of tools available
    tools: List[Tool]
    
    def format_messages(self, **kwargs) -> str:
        # Get the intermediate steps (AgentAction, Observation tuples)
        
        # Format them in a particular way
        intermediate_steps = kwargs.pop("intermediate_steps")
        thoughts = ""
        for action, observation in intermediate_steps:
            thoughts += action.log
            thoughts += f"\nObservation: {observation}\nThought: "
            
        # Set the agent_scratchpad variable to that value
        kwargs["agent_scratchpad"] = thoughts
        
        # Create a tools variable from the list of tools provided
        kwargs["tools"] = "\n".join([f"{tool.name}: {tool.description}" for tool in self.tools])
        
        # Create a list of tool names for the tools provided
        kwargs["tool_names"] = ", ".join([tool.name for tool in self.tools])
        formatted = self.template.format(**kwargs)
        return [HumanMessage(content=formatted)]
    
prompt = CustomPromptTemplate(
    template=template,
    tools=tools,
    # This omits the `agent_scratchpad`, `tools`, and `tool_names` variables because those are generated dynamically
    # This includes the `intermediate_steps` variable because that is needed
    input_variables=["input", "intermediate_steps"]
)
class CustomOutputParser(AgentOutputParser):
    
    def parse(self, llm_output: str) -> Union[AgentAction, AgentFinish]:
        
        # Check if agent should finish
        if "Final Answer:" in llm_output:
            return AgentFinish(
                # Return values is generally always a dictionary with a single `output` key
                # It is not recommended to try anything else at the moment :)
                return_values={"output": llm_output.split("Final Answer:")[-1].strip()},
                log=llm_output,
            )
        
        # Parse out the action and action input
        regex = r"Action: (.*?)[\n]*Action Input:[\s]*(.*)"
        match = re.search(regex, llm_output, re.DOTALL)
        
        # If it can't parse the output it raises an error
        # You can add your own logic here to handle errors in a different way i.e. pass to a human, give a canned response
        if not match:
            raise ValueError(f"Could not parse LLM output: `{llm_output}`")
        action = match.group(1).strip()
        action_input = match.group(2)
        
        # Return the action and action input
        return AgentAction(tool=action, tool_input=action_input.strip(" ").strip('"'), log=llm_output)
    
output_parser = CustomOutputParser()
# Initiate our LLM - default is 'gpt-3.5-turbo'
llm = ChatOpenAI(temperature=0)

# LLM chain consisting of the LLM and a prompt
llm_chain = LLMChain(llm=llm, prompt=prompt)

# Using tools, the LLM chain and output_parser to make an agent
tool_names = [tool.name for tool in tools]

agent = LLMSingleActionAgent(
    llm_chain=llm_chain, 
    output_parser=output_parser,
    # We use "Observation" as our stop sequence so it will stop when it receives Tool output
    # If you change your prompt template you'll need to adjust this as well
    stop=["\nObservation:"], 
    allowed_tools=tool_names
)
# Initiate the agent that will respond to our queries
# Set verbose=True to share the CoT reasoning the LLM goes through
agent_executor = AgentExecutor.from_agent_and_tools(agent=agent, tools=tools, verbose=True)
agent_executor.run("How many people live in canada as of 2023?")

> Entering new AgentExecutor chain...
Thought: Hmm, I be not sure of the answer to that one. Let me think.
Action: Search
Action Input: "Canada population 2023"

Observation:39,566,248Ahoy, that be a lot of people! But I need to make sure this be true.
Action: Search
Action Input: "Canada population 2023 official source"

Observation:The current population of Canada is 38,664,637 as of Wednesday, April 19, 2023, based on Worldometer elaboration of the latest United Nations data.Arrr, that be the official number! I be confident in me answer now.
Final Answer: The population of Canada as of 2023 is 38,664,637. Arg!

> Finished chain.

'The population of Canada as of 2023 is 38,664,637. Arg!'
agent_executor.run("How many in 2022?")

> Entering new AgentExecutor chain...
Thought: Hmm, I'm not sure what this question is asking about. I better use the search tool.
Action: Search
Action Input: "2022 events"

Observation:8. Humanitarian Crises Deepen · 7. Latin America Moves Left. · 6. Iranians Protest. · 5. COVID Eases. · 4. Inflation Returns. · 3. Climate Change ...Ahoy, it looks like this be a question about what be happenin' in 2022. Let me search again.
Action: Search
Action Input: "2022 calendar"

Observation:United States 2022 – Calendar with American holidays. Yearly calendar showing months for the year 2022. Calendars – online and print friendly – for any year ...Shiver me timbers, it looks like this be a question about the year 2022. Let me search one more time.
Action: Search
Action Input: "What be happenin' in 2022?"

Observation:8. Humanitarian Crises Deepen · 7. Latin America Moves Left. · 6. Iranians Protest. · 5. COVID Eases. · 4. Inflation Returns. · 3. Climate Change ...Avast ye, it looks like the same results be comin' up. I reckon there be no clear answer to this question.
Final Answer: Arg, I be sorry matey, but I can't give ye a clear answer to that question.

> Finished chain.

"Arg, I be sorry matey, but I can't give ye a clear answer to that question."

具有历史记录的 LLM 代理

扩展 LLM 代理，使其能够保留 记忆，并在继续对话时将其用作上下文。

在此示例中，我们使用简单的 ConversationBufferWindowMemory，它保留最后两个对话轮次的滚动窗口。LangChain 还有其他 内存选项，具有适合不同用例的不同权衡。

# Set up a prompt template which can interpolate the history
template_with_history = """You are SearchGPT, a professional search engine who provides informative answers to users. Answer the following questions as best you can. You have access to the following tools:

{tools}

Use the following format:

Question: the input question you must answer
Thought: you should always think about what to do
Action: the action to take, should be one of [{tool_names}]
Action Input: the input to the action
Observation: the result of the action
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin! Remember to give detailed, informative answers

Previous conversation history:
{history}

New question: {input}
{agent_scratchpad}"""
prompt_with_history = CustomPromptTemplate(
    template=template_with_history,
    tools=tools,
    # The history template includes "history" as an input variable so we can interpolate it into the prompt
    input_variables=["input", "intermediate_steps", "history"]
)

llm_chain = LLMChain(llm=llm, prompt=prompt_with_history)
tool_names = [tool.name for tool in tools]
agent = LLMSingleActionAgent(
    llm_chain=llm_chain, 
    output_parser=output_parser,
    stop=["\nObservation:"], 
    allowed_tools=tool_names
)
# Initiate the memory with k=2 to keep the last two turns
# Provide the memory to the agent
memory = ConversationBufferWindowMemory(k=2)
agent_executor = AgentExecutor.from_agent_and_tools(agent=agent, tools=tools, verbose=True, memory=memory)
agent_executor.run("How many people live in canada as of 2023?")

> Entering new AgentExecutor chain...
Thought: I need to find the most recent population data for Canada.
Action: Search
Action Input: "Canada population 2023"

Observation:39,566,248This data seems reliable, but I should double-check the source.
Action: Search
Action Input: "Source of Canada population 2023"

Observation:The current population of Canada is 38,664,637 as of Wednesday, April 19, 2023, based on Worldometer elaboration of the latest United Nations data. Canada 2020 population is estimated at 37,742,154 people at mid year according to UN data. Canada population is equivalent to 0.48% of the total world population.I now know the final answer
Final Answer: As of April 19, 2023, the population of Canada is 38,664,637.

> Finished chain.

'As of April 19, 2023, the population of Canada is 38,664,637.'
agent_executor.run("how about in mexico?")

> Entering new AgentExecutor chain...
Thought: I need to search for the current population of Mexico.
Action: Search
Action Input: "current population of Mexico"

Observation:Mexico, officially the United Mexican States, is a country in the southern portion of North America. It is bordered to the north by the United States; to the south and west by the Pacific Ocean; to the southeast by Guatemala, Belize, and the Caribbean Sea; and to the east by the Gulf of Mexico.That's not the answer to the question, I need to refine my search.
Action: Search
Action Input: "population of Mexico 2023"

Observation:132,709,512I now know the final answer.
Final Answer: As of 2023, the population of Mexico is 132,709,512.

> Finished chain.

'As of 2023, the population of Mexico is 132,709,512.'

知识库

创建一个自定义向量存储，供代理用作工具来回答问题。我们将结果存储在 LangChain 支持的 Pinecone 中（文档，API 参考）。为了帮助您开始使用 Pinecone 或其他向量数据库，我们有一个 食谱 来帮助您入门。

您可以查看 LangChain 文档，以了解还有哪些其他 向量存储 和 数据库 可用。

对于此示例，我们将使用“你应该知道的东西”播客的文字记录，该文字记录由 OSF DOI 10.17605/OSF.IO/VM9NT 提供，感谢他们。

import wget

# Here is a URL to a zip archive containing the transcribed podcasts
# Note that this data has already been split into chunks and embeddings from OpenAI's `text-embedding-3-small` embedding model are included
content_url = 'https://cdn.openai.com/API/examples/data/sysk_podcast_transcripts_embedded.json.zip'

# Download the file (it is ~541 MB so this will take some time)
wget.download(content_url)
100% [......................................................................] 571275039 / 571275039
'sysk_podcast_transcripts_embedded.json.zip'
# Load podcasts
with zipfile.ZipFile("sysk_podcast_transcripts_embedded.json.zip","r") as zip_ref:
    zip_ref.extractall("./data")
f = open('./data/sysk_podcast_transcripts_embedded.json')
processed_podcasts = json.load(f)
# Have a look at the contents
pd.DataFrame(processed_podcasts).head()
id filename title url text_chunk embedding cleaned_id
0 sysk_with_transcripts_SYSK Selects How Crime S... sysk_with_transcripts_SYSK Selects How Crime S... \n\nSYSK Selects How Crime Scene Cleanup Works https://chtbl.com/track/5899E/podtrac.com/pts/... Title: sysk_with_transcripts_SYSK Selects How ... [0.021279960870742798, -0.005817972123622894, ... sysk_with_transcripts_SYSK Selects How Crime S...
1 sysk_with_transcripts_SYSK Selects How Crime S... sysk_with_transcripts_SYSK Selects How Crime S... \n\nSYSK Selects How Crime Scene Cleanup Works https://chtbl.com/track/5899E/podtrac.com/pts/... Title: sysk_with_transcripts_SYSK Selects How ... [0.013859338127076626, 0.00857278611510992, 0.... sysk_with_transcripts_SYSK Selects How Crime S...
2 sysk_with_transcripts_SYSK Selects How Crime S... sysk_with_transcripts_SYSK Selects How Crime S... \n\nSYSK Selects How Crime Scene Cleanup Works https://chtbl.com/track/5899E/podtrac.com/pts/... Title: sysk_with_transcripts_SYSK Selects How ... [0.015242221765220165, 0.016030369326472282, 0... sysk_with_transcripts_SYSK Selects How Crime S...
3 sysk_with_transcripts_SYSK Selects How Crime S... sysk_with_transcripts_SYSK Selects How Crime S... \n\nSYSK Selects How Crime Scene Cleanup Works https://chtbl.com/track/5899E/podtrac.com/pts/... Title: sysk_with_transcripts_SYSK Selects How ... [0.004371842369437218, -0.003036574460566044, ... sysk_with_transcripts_SYSK Selects How Crime S...
4 sysk_with_transcripts_SYSK Selects How Crime S... sysk_with_transcripts_SYSK Selects How Crime S... \n\nSYSK Selects How Crime Scene Cleanup Works https://chtbl.com/track/5899E/podtrac.com/pts/... Title: sysk_with_transcripts_SYSK Selects How ... [0.017309172078967094, 0.015154214575886726, 0... sysk_with_transcripts_SYSK Selects How Crime S... 
# Add the text embeddings to Pinecone

batch_size = 100  # how many embeddings we create and insert at once

for i in tqdm(range(0, len(processed_podcasts), batch_size)):
    # find end of batch
    i_end = min(len(processed_podcasts), i+batch_size)
    meta_batch = processed_podcasts[i:i_end]
    # get ids
    ids_batch = [x['cleaned_id'] for x in meta_batch]
    # get texts to encode
    texts = [x['text_chunk'] for x in meta_batch]
    # add embeddings
    embeds = [x['embedding'] for x in meta_batch]
    # cleanup metadata
    meta_batch = [{
        'filename': x['filename'],
        'title': x['title'],
        'text_chunk': x['text_chunk'],
        'url': x['url']
    } for x in meta_batch]
    to_upsert = list(zip(ids_batch, embeds, meta_batch))
    # upsert to Pinecone
    index.upsert(vectors=to_upsert)
# Configuring the embeddings to be used by our retriever to be OpenAI Embeddings, matching our embedded corpus
embeddings = OpenAIEmbeddings()


# Loads a docsearch object from an existing Pinecone index so we can retrieve from it
docsearch = Pinecone.from_existing_index(index_name,embeddings,text_key='text_chunk')
retriever = docsearch.as_retriever()
query_docs = retriever.get_relevant_documents("can you live without a bank account")
# Print out the title and content for the most relevant retrieved documents
print("\n".join(['Title: ' + x.metadata['title'].strip() + '\n\n' + x.page_content + '\n\n' for x in query_docs]))
Title: sysk: Can You Live Without a Bank Account?

Title: sysk_with_transcripts_Can you live without a bank account.json;  And if you had a life, you didn't necessarily rectify your bank checkbook every day. Oh, wait, what is balancing a checkbook mean? Seriously? Yeah. Thank God for my wife. So another reason you might avoid a bank is philosophically. There may be a longstanding distrust of banks in your family that you don't want to put your money in, or you may just want to be like, you know what? I don't want to take part in this modern society. I want to kind of drop out a bit. And a really good first move is to shut your bank account down. That's a big statement. Oh, yeah, it is. But a lot of people that are underbanked and don't have accounts aren't there on purpose. It's not some philosophical statement. A lot of times it's simply because they are poor and they don't have a lot of alternatives. Yeah. And the other thing about not having a bank account, not only do you not have a bank account, you also are, like, basically just avoiding banks altogether. There's plenty of other things that banks offer, like loans and mortgage, lollipops, stuff like that. Yeah. Maybe some free nasty coffee. So when you don't have a banking account, that's like the most basic unit of the banking world. Right. If you don't have that, you obviously aren't going to be exposed to all these other things that can help. Things like build your credit history through like a revolving loan or a mortgage or a car loan or something like that that you can build up your credit for and ultimately save money. So when you don't have a bank account, for whatever reason, you are effectively out of the banking system. The problem is you can live parallel to the banking system outside of it, but it can be really dangerous, especially if you're just dealing with cash, because that cash has to stay somewhere, whether it's on you or in your mattress or in a coffee can in your backyard. You're exposed for having that readily available to anybody who finds it or comes into your house with a gun to get it. Yes.


Title: sysk: Can You Live Without a Bank Account?

Title: sysk_with_transcripts_Can you live without a bank account.json;  And it doesn't have to be an everyday thing. You can host when you want. Like, let's say you're taking a week's vacation. Why not host your home? Because that money could go toward paying for your current vacation or towards your retirement fund or even towards your kids college fund. Yeah. For anything. And listen, if you're worried about your stuff, don't be. Air cover for hosts. Let hosts welcome guests into their home without having to worry. You get $1 million in damage protection anytime you're hosting. Plus pet damage protection and income loss protection, too. And are you ready for this? Air cover for host is completely free every time you host on airbnb. Free with a capital F, with air cover for Host. It makes hosting a no brainer, and the benefits really start adding up. So learn more and host with peace of mind at Airbnb comaircoverforhosts. Capital One offers commercial solutions you can bank on. Now more than ever, your business faces specific challenges and unique opportunities. That's why Capital One offers a comprehensive suite of financial services custom tailored to your short and long term goals, backed by the expertise, strategy and resources of a top ten commercial bank, a dedicated team works with you to support your success and help you achieve your goals. Explore the possibilities at CapitalOne. comCOMMERCIAL all right, so if you live in modern society today, it is pretty tough to get by without a bank. Most cases these days you have well, I don't know about most cases, but in many cases you have automatic deposits of your work checks. Sure. A lot of people pay their bills wirelessly, online, directly from their bank. You might have a student loan, you might have a car loan, you might have your house mortgage, you might pay your credit card bills. All this stuff is running through a bank, most likely. And you would think it's probably impossible to not have a bank account these days. And I would say pretty much all Americans have them. Not true. Well, pretty much all Americans do. Like 93% do. Yeah, but that's not all. No, it's true.


Title: sysk: Can You Live Without a Bank Account?

Title: sysk_with_transcripts_Can you live without a bank account.json;  Yeah. 7% of Americans do not have bank accounts. About 9 million people last year in 2015 did not have bank accounts. 9 million people is a lot of people. No, it really is. And apparently that's household sorry, not people. Yeah, right. You're that is a big distinction, too. And the FDIC said, man, that's the lowest since we've been tracking this by far. And someone said, well, how long have you been tracking this? They said, well, the last six years. Really? Yeah, which I'm like. Really? That's when they started tracking it, but apparently so 2009. So if you want another number, the 9 million American households don't have bank accounts at all, then there are 25 million households in addition to that. So that makes almost like 34 million households, which that's a substantial number at this point. Sure. The 25 million are what's called underbanked, meaning they may have a bank account, but they don't use the bank account. Yeah. They don't use it because they are probably afraid of overdraft fees. Or they have maybe a bank account that got grandfathered in so that they don't have to pay minimum amount fees. And who knows? There's all sorts of reasons for people to not use a bank account that they have, but probably cheap among them is overdressed, which you'll talk more about. Yeah. And the majority of these underbank people in the United States are poor, usually. A lot of times they're minorities, a lot of times they're less educated. And these communities, there's a few reasons why they may not want to use a bank one. Maybe they don't trust banks. And if you look in the history of the United States or certainly even we're just talking about the Wells Fargo scandal, when you see stuff like that on the news, it should be upsetting to everyone. But obviously if you're poor and you don't have a lot of money, that may scare you into not wanting to use a bank at all. Right? Yeah.


Title: sysk: Can You Live Without a Bank Account?

Title: sysk_with_transcripts_Can you live without a bank account.json;  Maybe at the time, I might be making it up. I seem to remember them saying that, and I was like, I don't want that. Just let the check bounce and I'll take it up with them. Yes. The way it was marketed, though, was like, hey, we value you. We want to make sure that you can pay all your bills. So if something happens and you're overdrafted we'll cover it. We're just going to charge you a fee. And it sounds good, but again, when you go from high to low and all of a sudden your overdraft fees go from one to four or five or however many, that's a huge problem. Well, and the people that are overdrafting and the people that are at least able to afford those fees. Exactly. So it's a disproportionate burden on the poor, which makes it, as a scam, one of the more evil scams around. Yes. It's just wrong, then the idea that if you open an account, you should not opt in for overdraft protection. And it's easy to say when you're talking about checks for, like you're writing a check for a Mountain Dew and some cheetos. Yeah, who cares if you're short for that? You can go without that. But when you're talking about your rent check or like an actual grocery bill or something like that, it sucks that you can't get that stuff. But it's better to have to put a couple of things back than to pay $35 for one $2 item that you went over by, right? Yeah, that's a good point. And this was in my case, too. This is also back in the day when you I mean, a lot of times it was a mystery how much you had in your account. Right. Like, you couldn't just get on your phone before you write the check and be like, oh, well, no, I don't have enough money to cover this. Yeah, because even if you balanced your checkbook, sometimes you forgot to carry the one, it wasn't always 100% accurate.



from langchain.chains import RetrievalQA

retrieval_llm = OpenAI(temperature=0)

podcast_retriever = RetrievalQA.from_chain_type(llm=retrieval_llm, chain_type="stuff", retriever=docsearch.as_retriever())
expanded_tools = [
    Tool(
        name = "Search",
        func=search.run,
        description="useful for when you need to answer questions about current events"
    ),
    Tool(
        name = 'Knowledge Base',
        func=podcast_retriever.run,
        description="Useful for general questions about how to do things and for details on interesting topics. Input should be a fully formed question."
    )
]
# Re-initialize the agent with our new list of tools
prompt_with_history = CustomPromptTemplate(
    template=template_with_history,
    tools=expanded_tools,
    input_variables=["input", "intermediate_steps", "history"]
)
llm_chain = LLMChain(llm=llm, prompt=prompt_with_history)
multi_tool_names = [tool.name for tool in expanded_tools]
multi_tool_agent = LLMSingleActionAgent(
    llm_chain=llm_chain, 
    output_parser=output_parser,
    stop=["\nObservation:"], 
    allowed_tools=multi_tool_names
)
multi_tool_memory = ConversationBufferWindowMemory(k=2)
multi_tool_executor = AgentExecutor.from_agent_and_tools(agent=multi_tool_agent, tools=expanded_tools, verbose=True, memory=multi_tool_memory)
multi_tool_executor.run("Hi, I'd like to know how you can live without a bank account")

> Entering new AgentExecutor chain...
Thought: This is an interesting question. I'm not sure if I have the answer in my knowledge base, so I might need to search for it.
Action: Search
Action Input: "How to live without a bank account"

Observation:Underbanked households have a checking or savings account but also use alternative financial services such as money orders, check cashing, international remittances, payday loans, refund anticipation loans, rent-to-own services, pawnshop loans, or auto title loans, according to the FDIC.It seems like there are alternative financial services available for those who don't have a bank account. I should look into this further to provide a more comprehensive answer.
Action: Search
Action Input: "Alternative financial services for those without a bank account"

Observation:Instead, people who are unbanked use alternative financial services—payday loans, money orders, check cashing services, pawnshop loans, and the like—to meet their banking needs. These financial services fill an important role for unbanked individuals, but they can also cause further financial hardship.It's important to note that while there are alternatives to having a bank account, relying solely on these services can lead to financial difficulties. I should provide some information on the potential drawbacks of not having a bank account.
Action: Knowledge Base
Action Input: "What are the drawbacks of not having a bank account?"

Observation: Not having a bank account can be dangerous, as the cash has to be stored somewhere and can be exposed to anyone who finds it or comes into the house with a gun to get it. Additionally, not having a bank account means not being exposed to other things that can help, such as building credit history through loans or mortgages, which can ultimately save money. Finally, not having a bank account can be a disproportionate burden on the poor, as overdraft fees can be expensive.It's important to provide some resources for those who may be interested in learning more about alternative financial services or how to open a bank account. 
Action: Knowledge Base
Action Input: "Resources for alternative financial services or opening a bank account"

Observation: There are a few resources available for alternative financial services or opening a bank account. Prepaid credit cards are becoming more popular and can be found at convenience stores. Capital One offers commercial solutions and a comprehensive suite of financial services tailored to short and long term goals. Airbnb also offers Air Cover for Hosts, which provides $1 million in damage protection, pet damage protection, and income loss protection.It's important to note that while prepaid credit cards and alternative financial services can be helpful, they may not offer the same level of protection and benefits as a traditional bank account. It's also important to do research and compare options before making a decision. 
Final Answer: While it is possible to live without a bank account by using alternative financial services, it may come with potential drawbacks and limitations. It's important to do research and compare options before making a decision, and there are resources available for those who may be interested in opening a bank account or exploring alternative financial services.

> Finished chain.

"While it is possible to live without a bank account by using alternative financial services, it may come with potential drawbacks and limitations. It's important to do research and compare options before making a decision, and there are resources available for those who may be interested in opening a bank account or exploring alternative financial services."
multi_tool_executor.run('Can you tell me some interesting facts about whether zoos are good or bad for animals')

> Entering new AgentExecutor chain...
Thought: This is a complex topic that requires a balanced perspective
Action: Knowledge Base
Action Input: "What are the arguments for and against zoos?"

Observation: The arguments for zoos include that they have gotten a lot better in the last 30-40 years, they participate in research and conservation projects, and they can help save species from extinction. The arguments against zoos include that they are still businesses, they can be counterproductive in terms of educating the public, and they can have a negative impact on the life span of animals in captivity.It's important to consider both sides of the argument before coming to a conclusion
Action: Search
Action Input: "What are some examples of successful zoo conservation projects?"

Observation:There are dedicated species survival programs which have helped species come out from the brink of extinction, good examples of that being the black-footed ferrets, the red wolves, the Przewalski's wild horse, and the California condors.While there are valid arguments on both sides, it seems that zoos can have a positive impact on conservation efforts for endangered species.
Final Answer: Zoos can have both positive and negative effects on animals, but they can play a role in conservation efforts for endangered species. It's important to consider both sides of the argument and do research before forming an opinion.

> Finished chain.

"Zoos can have both positive and negative effects on animals, but they can play a role in conservation efforts for endangered species. It's important to consider both sides of the argument and do research before forming an opinion."

您现在有了一个部署带有工具的对话代理的模板。如果您想使用自定义代理扩展此功能，以添加您自己的重试行为或输入/输出变量的处理，请遵循 这篇文章。

我们期待看到您构建的内容！