跳到主要内容

使用Langchain、AnalyticDB和OpenAI进行问答

nbviewer

本笔记本展示了如何使用Langchain、AnalyticDB作为知识库和OpenAI嵌入来实现一个问答系统。如果您对AnalyticDB不熟悉，最好先查看Getting_started_with_AnalyticDB_and_OpenAI.ipynb笔记。

本笔记本展示了一个端到端的流程： - 使用OpenAI API计算嵌入。 - 将嵌入存储在AnalyticDB实例中以构建知识库。 - 将原始文本查询转换为嵌入，使用OpenAI API。 - 使用AnalyticDB在创建的集合中执行最近邻搜索，以找到一些上下文。 - 请求LLM在给定上下文中找到答案。

所有步骤将简化为调用一些相应的Langchain方法。

先决条件​

为了完成这个练习，我们需要准备一些东西： - AnalyticDB 云实例。 - Langchain 作为框架。 - OpenAI API 密钥。

安装要求​

此笔记本需要以下Python包：openai、tiktoken、langchain和psycopg2cffi。 - openai 提供了方便访问OpenAI API的功能。 - tiktoken 是一个快速的BPE标记器，可与OpenAI的模型一起使用。 - langchain 帮助我们更轻松地构建具有LLM的应用程序。 - psycopg2cffi 库用于与向量数据库交互，但也可以使用任何其他PostgreSQL客户端库。

! pip install openai tiktoken langchain psycopg2cffi 

! export OPENAI_API_KEY="your API key"

# 测试您的OpenAI API密钥是否已正确设置为环境变量。
# 注意：如果您在本地运行此笔记本，您需要重新加载终端和笔记本，以使环境变量生效。
import os

# 注意：或者，您也可以像这样设置一个临时的环境变量：
# os.environ["OPENAI_API_KEY"] = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

if os.getenv("OPENAI_API_KEY") is not None:
    print("OPENAI_API_KEY is ready")
else:
    print("OPENAI_API_KEY environment variable not found")

OPENAI_API_KEY is ready

准备您的OpenAI API密钥​

OpenAI API密钥用于对文档和查询进行向量化。

如果您没有OpenAI API密钥，可以从https://platform.openai.com/account/api-keys获取。

获取密钥后，请通过运行以下命令将其添加到您的环境变量中，命名为OPENAI_API_KEY：

准备您的 AnalyticDB 连接字符串​

要构建 AnalyticDB 连接字符串，您需要拥有以下参数：PG_HOST、PG_PORT、PG_DATABASE、PG_USER 和 PG_PASSWORD。您需要首先导出它们以设置正确的连接字符串。然后构建连接字符串。

! export PG_HOST="your AnalyticDB host url"
! export PG_PORT=5432 # 可选，默认值为 5432
! export PG_DATABASE=postgres # 可选，默认值为 postgres
! export PG_USER="your username"
! export PG_PASSWORD="your password"

import os
from langchain.vectorstores.analyticdb import AnalyticDB

CONNECTION_STRING = AnalyticDB.connection_string_from_db_params(
    driver=os.environ.get("PG_DRIVER", "psycopg2cffi"),
    host=os.environ.get("PG_HOST", "localhost"),
    port=int(os.environ.get("PG_PORT", "5432")),
    database=os.environ.get("PG_DATABASE", "postgres"),
    user=os.environ.get("PG_USER", "postgres"),
    password=os.environ.get("PG_PASSWORD", "postgres"),
)

import json

with open("questions.json", "r") as fp:
    questions = json.load(fp)

with open("answers.json", "r") as fp:
    answers = json.load(fp)

加载数据​

在这一部分，我们将加载包含一些自然问题及其答案的数据。所有数据将用于创建一个以AnalyticDB为知识库的Langchain应用程序。

print(questions[0])

when is the last episode of season 8 of the walking dead
import wget

# 所有示例均来自 https://ai.google.com/research/NaturalQuestions。
# 这是我们从下载并提取的训练集中抽取的一个样本，用于某些特定目的。
# 进一步加工。
wget.download("https://storage.googleapis.com/dataset-natural-questions/questions.json")
wget.download("https://storage.googleapis.com/dataset-natural-questions/answers.json")

print(answers[0])

No . overall No. in season Title Directed by Written by Original air date U.S. viewers ( millions ) 100 `` Mercy '' Greg Nicotero Scott M. Gimple October 22 , 2017 ( 2017 - 10 - 22 ) 11.44 Rick , Maggie , and Ezekiel rally their communities together to take down Negan . Gregory attempts to have the Hilltop residents side with Negan , but they all firmly stand behind Maggie . The group attacks the Sanctuary , taking down its fences and flooding the compound with walkers . With the Sanctuary defaced , everyone leaves except Gabriel , who reluctantly stays to save Gregory , but is left behind when Gregory abandons him . Surrounded by walkers , Gabriel hides in a trailer , where he is trapped inside with Negan . 101 `` The Damned '' Rosemary Rodriguez Matthew Negrete & Channing Powell October 29 , 2017 ( 2017 - 10 - 29 ) 8.92 Rick 's forces split into separate parties to attack several of the Saviors ' outposts , during which many members of the group are killed ; Eric is critically injured and rushed away by Aaron . Jesus stops Tara and Morgan from executing a group of surrendered Saviors . While clearing an outpost with Daryl , Rick is confronted and held at gunpoint by Morales , a survivor he met in the initial Atlanta camp , who is now with the Saviors . 102 `` Monsters '' Greg Nicotero Matthew Negrete & Channing Powell November 5 , 2017 ( 2017 - 11 - 05 ) 8.52 Daryl finds Morales threatening Rick and kills him ; the duo then pursue a group of Saviors who are transporting weapons to another outpost . Gregory returns to Hilltop , and after a heated argument , Maggie ultimately allows him back in the community . Eric dies from his injuries , leaving Aaron distraught . Despite Tara and Morgan 's objections , Jesus leads the group of surrendered Saviors to Hilltop . Ezekiel 's group attacks another Savior compound , during which several Kingdommers are shot while protecting Ezekiel . 103 `` Some Guy '' Dan Liu David Leslie Johnson November 12 , 2017 ( 2017 - 11 - 12 ) 8.69 Ezekiel 's group is overwhelmed by the Saviors , who kill all of them except for Ezekiel himself and Jerry . Carol clears the inside of the compound , killing all but two Saviors , who almost escape but are eventually caught by Rick and Daryl . En route to the Kingdom , Ezekiel , Jerry , and Carol are surrounded by walkers , but Shiva sacrifices herself to save them . The trio returns to the Kingdom , where Ezekiel 's confidence in himself as a leader has diminished . 104 5 `` The Big Scary U '' Michael E. Satrazemis Story by : Scott M. Gimple & David Leslie Johnson & Angela Kang Teleplay by : David Leslie Johnson & Angela Kang November 19 , 2017 ( 2017 - 11 - 19 ) 7.85 After confessing their sins to each other , Gabriel and Negan manage to escape from the trailer . Simon and the other lieutenants grow suspicious of each other , knowing that Rick 's forces must have inside information . The workers in the Sanctuary become increasingly frustrated with their living conditions , and a riot nearly ensues , until Negan returns and restores order . Gabriel is locked in a cell , where Eugene discovers him sick and suffering . Meanwhile , Rick and Daryl argue over how to take out the Saviors , leading Daryl to abandon Rick . 105 6 `` The King , the Widow , and Rick '' John Polson Angela Kang & Corey Reed November 26 , 2017 ( 2017 - 11 - 26 ) 8.28 Rick visits Jadis in hopes of convincing her to turn against Negan ; Jadis refuses , and locks Rick in a shipping container . Carl encounters Siddiq in the woods and recruits him to Alexandria . Daryl and Tara plot to deviate from Rick 's plans by destroying the Sanctuary . Ezekiel isolates himself at the Kingdom , where Carol tries to encourage him to be the leader his people need . Maggie has the group of captured Saviors placed in a holding area and forces Gregory to join them as punishment for betraying Hilltop . 106 7 `` Time for After '' Larry Teng Matthew Negrete & Corey Reed December 3 , 2017 ( 2017 - 12 - 03 ) 7.47 After learning of Dwight 's association with Rick 's group , Eugene affirms his loyalty to Negan and outlines a plan to get rid of the walkers surrounding the Sanctuary . With help from Morgan and Tara , Daryl drives a truck through the Sanctuary 's walls , flooding its interior with walkers , killing many Saviors . Rick finally convinces Jadis and the Scavengers to align with him , and they plan to force the Saviors to surrender . However , when they arrive at the Sanctuary , Rick is horrified to see the breached walls and no sign of the walker herd . 107 8 `` How It 's Gotta Be '' Michael E. Satrazemis David Leslie Johnson & Angela Kang December 10 , 2017 ( 2017 - 12 - 10 ) 7.89 Eugene 's plan allows the Saviors to escape , and separately , the Saviors waylay the Alexandria , Hilltop , and Kingdom forces . The Scavengers abandon Rick , after which he returns to Alexandria . Ezekiel ensures that the Kingdom residents are able to escape before locking himself in the community with the Saviors . Eugene aids Gabriel and Doctor Carson in escaping the Sanctuary in order to ease his conscience . Negan attacks Alexandria , but Carl devises a plan to allow the Alexandria residents to escape into the sewers . Carl reveals he was bitten by a walker while escorting Siddiq to Alexandria . 108 9 `` Honor '' Greg Nicotero Matthew Negrete & Channing Powell February 25 , 2018 ( 2018 - 02 - 25 ) 8.28 After the Saviors leave Alexandria , the survivors make for the Hilltop while Rick and Michonne stay behind to say their final goodbyes to a dying Carl , who pleads with Rick to build a better future alongside the Saviors before killing himself . In the Kingdom , Morgan and Carol launch a rescue mission for Ezekiel . Although they are successful and retake the Kingdom , the Saviors ' lieutenant Gavin is killed by Benjamin 's vengeful brother Henry . 109 10 `` The Lost and the Plunderers '' TBA TBA March 4 , 2018 ( 2018 - 03 - 04 ) TBD 110 11 `` Dead or Alive Or '' TBA TBA March 11 , 2018 ( 2018 - 03 - 11 ) TBD 111 12 `` The Key '' TBA TBA March 18 , 2018 ( 2018 - 03 - 18 ) TBD

链定义​

Langchain已经集成了AnalyticDB，并对给定的文档列表执行所有索引。在我们的案例中，我们将存储我们拥有的一组答案。

from langchain.vectorstores import AnalyticDB
from langchain.embeddings import OpenAIEmbeddings
from langchain import VectorDBQA, OpenAI

embeddings = OpenAIEmbeddings()
doc_store = AnalyticDB.from_texts(
    texts=answers, embedding=embeddings, connection_string=CONNECTION_STRING,
    pre_delete_collection=True,
)

在这个阶段，所有可能的答案都已经存储在AnalyticDB中，因此我们可以定义整个问答链条。

from langchain.chains import RetrievalQA

llm = OpenAI()
qa = VectorDBQA.from_chain_type(
    llm=llm,
    chain_type="stuff",
    vectorstore=doc_store,
    return_source_documents=False,
)

搜索数据​

一旦数据被放入AnalyticDB中，我们就可以开始提出一些问题。一个问题将会被OpenAI模型自动转换为向量，然后使用创建的向量在AnalyticDB中找到一些可能匹配的答案。一旦检索到，最相似的答案将被合并到发送给OpenAI大型语言模型的提示中。

import random

random.seed(52)
selected_questions = random.choices(questions, k=5)

for question in selected_questions:
    print(">", question)
    print(qa.run(question), end="\n\n")

> where do frankenstein and the monster first meet
 Victor retreats into the mountains, and that is where the Creature finds him and pleads for Victor to hear his tale.

> who are the actors in fast and furious
 The main cast of Fast & Furious includes Vin Diesel as Dominic Toretto, Paul Walker as Brian O'Conner, Michelle Rodriguez as Letty Ortiz, Jordana Brewster as Mia Toretto, Tyrese Gibson as Roman Pearce, and Ludacris as Tej Parker.

> properties of red black tree in data structure
 The properties of a red-black tree in data structure are that each node is either red or black, the root is black, all leaves (NIL) are black, and if a node is red, then both its children are black. Additionally, every path from a given node to any of its descendant NIL nodes contains the same number of black nodes.

> who designed the national coat of arms of south africa
 Iaan Bekker

> caravaggio's death of the virgin pamela askew
 I don't know.

自定义提示模板​

Langchain中的stuff链类型使用特定的提示，其中包含问题和上下文文档。这是默认提示的样式：

使用以下上下文片段来回答最后的问题。如果你不知道答案，只需说你不知道，不要试图凭空编造答案。
{context}
问题：{question}
有用的答案：

然而，我们可以提供自定义的提示模板，并改变OpenAI LLM的行为，同时仍然使用stuff链类型。重要的是保持{context}和{question}作为占位符。

尝试自定义提示​

我们可以尝试使用不同的提示模板，这样模型： 1. 如果知道答案，就用一个简短的句子回答。 2. 如果不知道问题的答案，就建议一个随机的歌曲标题。

from langchain.prompts import PromptTemplate
custom_prompt = """
Use the following pieces of context to answer the question at the end. Please provide
a short single-sentence summary answer only. If you don't know the answer or if it's
not present in given context, don't try to make up an answer, but suggest me a random
unrelated song title I could listen to.
Context: {context}
Question: {question}
Helpful Answer:
"""

custom_prompt_template = PromptTemplate(
    template=custom_prompt, input_variables=["context", "question"]
)

custom_qa = VectorDBQA.from_chain_type(
    llm=llm,
    chain_type="stuff",
    vectorstore=doc_store,
    return_source_documents=False,
    chain_type_kwargs={"prompt": custom_prompt_template},
)

random.seed(41)
for question in random.choices(questions, k=5):
    print(">", question)
    print(custom_qa.run(question), end="\n\n")

> what was uncle jesse's original last name on full house
Uncle Jesse's original last name on Full House was Cochran.

> when did the volcano erupt in indonesia 2018
No information about a volcano erupting in Indonesia in 2018 is present in the given context. Suggested song title: "Volcano" by U2.

> what does a dualist way of thinking mean
A dualist way of thinking means believing that humans possess a non-physical mind or soul which is distinct from their physical body.

> the first civil service commission in india was set up on the basis of recommendation of
The first Civil Service Commission in India was not set up on the basis of a recommendation.

> how old do you have to be to get a tattoo in utah
In Utah, you must be at least 18 years old to get a tattoo.