Skip to main content
Azure SQL 提供了一个专用的 向量数据类型,简化了在关系数据库内直接创建、存储和查询向量嵌入的过程。这消除了对单独向量数据库及相关集成的需求,提高了您解决方案的安全性,同时降低了整体复杂性。
Azure SQL 是一项强大的服务,结合了可扩展性、安全性和高可用性,提供了现代数据库解决方案的所有优势。它利用复杂的查询优化器和企业级功能,在执行传统 SQL 查询的同时进行向量相似性搜索,从而增强数据分析和决策能力。 阅读更多关于在 Azure SQL 数据库中使用智能应用 的信息。 本笔记本向您展示如何利用这个集成的 SQL 向量数据库 来存储文档,并使用余弦(余弦距离)、L2(欧几里得距离)和 IP(内积)执行向量搜索查询,以定位与查询向量接近的文档。

设置

安装 langchain-sqlserver Python 包。 代码位于一个名为 langchain-sqlserver 的集成包中。
!pip install langchain-sqlserver==0.1.1

凭据

运行此笔记本不需要任何凭据,只需确保您已下载 langchain-sqlserver 包。 如果您想获得一流的模型调用自动跟踪功能,也可以取消注释以下内容来设置您的 LangSmith API 密钥:
os.environ["LANGSMITH_API_KEY"] = getpass.getpass("Enter your LangSmith API key: ")
os.environ["LANGSMITH_TRACING"] = "true"

初始化

from langchain_sqlserver import SQLServer_VectorStore
在 Azure 门户中您的数据库设置下找到您的 Azure SQL DB 连接字符串。 更多信息:连接到 Azure SQL DB - Python
import os

import pyodbc

# 定义您的 SQLServer 连接字符串
_CONNECTION_STRING = (
    "Driver={ODBC Driver 18 for SQL Server};"
    "Server=<YOUR_DBSERVER>.database.windows.net,1433;"
    "Database=test;"
    "TrustServerCertificate=yes;"
    "Connection Timeout=60;"
    "LongAsMax=yes;"
)

# 连接字符串可能有所不同:
# "mssql+pyodbc://<username>:<password><servername>/<dbname>?driver=ODBC+Driver+18+for+SQL+Server" -> 指定用户名和密码
# "mssql+pyodbc://<servername>/<dbname>?driver=ODBC+Driver+18+for+SQL+Server&Trusted_connection=yes" -> 使用受信任连接
# "mssql+pyodbc://<servername>/<dbname>?driver=ODBC+Driver+18+for+SQL+Server" -> 使用 EntraID 连接
# "mssql+pyodbc://<servername>/<dbname>?driver=ODBC+Driver+18+for+SQL+Server&Trusted_connection=no" -> 使用 EntraID 连接
在此示例中,我们使用 Azure OpenAI 生成嵌入,但您也可以使用 LangChain 提供的不同嵌入。 您可以按照此 指南 在 Azure 门户上部署 Azure OpenAI 实例。一旦您的实例运行起来,请确保您拥有实例的名称和密钥。您可以在 Azure 门户中实例的“密钥和终结点”部分找到密钥。
!pip install langchain-openai
# 导入必要的库
from langchain_openai import AzureChatOpenAI, AzureOpenAIEmbeddings

# 设置您的 AzureOpenAI 详细信息
azure_endpoint = "https://<YOUR_ENDPOINT>.openai.azure.com/"
azure_deployment_name_embedding = "text-embedding-3-small"
azure_deployment_name_chatcompletion = "chatcompletion"
azure_api_version = "2023-05-15"
azure_api_key = "YOUR_KEY"


# 使用 AzureChatOpenAI 进行聊天补全
llm = AzureChatOpenAI(
    azure_endpoint=azure_endpoint,
    azure_deployment=azure_deployment_name_chatcompletion,
    openai_api_version=azure_api_version,
    openai_api_key=azure_api_key,
)

# 使用 AzureOpenAIEmbeddings 进行嵌入
embeddings = AzureOpenAIEmbeddings(
    azure_endpoint=azure_endpoint,
    azure_deployment=azure_deployment_name_embedding,
    openai_api_version=azure_api_version,
    openai_api_key=azure_api_key,
)

管理向量存储

from langchain_community.vectorstores.utils import DistanceStrategy
from langchain_sqlserver import SQLServer_VectorStore

# 初始化向量存储
vector_store = SQLServer_VectorStore(
    connection_string=_CONNECTION_STRING,
    distance_strategy=DistanceStrategy.COSINE,  # 可选,如果未提供,默认为 COSINE
    embedding_function=embeddings,  # 您可以使用 LangChain 提供的不同嵌入
    embedding_length=1536,
    table_name="langchain_test_table",  # 使用自定义名称的表
)

向向量存储添加项目

## 我们将在此示例中使用一些人工数据
query = [
    "I have bought several of the Vitality canned dog food products and have found them all to be of good quality. The product looks more like a stew than a processed meat and it smells better. My Labrador is finicky and she appreciates this product better than  most.",
    "The candy is just red , No flavor . Just  plan and chewy .  I would never buy them again",
    "Arrived in 6 days and were so stale i could not eat any of the 6 bags!!",
    "Got these on sale for roughly 25 cents per cup, which is half the price of my local grocery stores, plus they rarely stock the spicy flavors. These things are a GREAT snack for my office where time is constantly crunched and sometimes you can't escape for a real meal. This is one of my favorite flavors of Instant Lunch and will be back to buy every time it goes on sale.",
    "If you are looking for a less messy version of licorice for the children, then be sure to try these!  They're soft, easy to chew, and they don't get your hands all sticky and gross in the car, in the summer, at the beach, etc. We love all the flavos and sometimes mix these in with the chocolate to have a very nice snack! Great item, great price too, highly recommend!",
    "We had trouble finding this locally - delivery was fast, no more hunting up and down the flour aisle at our local grocery stores.",
    "Too much of a good thing? We worked this kibble in over time, slowly shifting the percentage of Felidae to national junk-food brand until the bowl was all natural. By this time, the cats couldn't keep it in or down. What a mess. We've moved on.",
    "Hey, the description says 360 grams - that is roughly 13 ounces at under $4.00 per can. No way - that is the approximate price for a 100 gram can.",
    "The taste of these white cheddar flat breads is like a regular cracker - which is not bad, except that I bought them because I wanted a cheese taste.<br /><br />What was a HUGE disappointment? How misleading the packaging of the box is. The photo on the box (I bought these in store) makes it look like it is full of long flatbreads (expanding the length and width of the box). Wrong! The plastic tray that holds the crackers is about 2"
    " smaller all around - leaving you with about 15 or so small flatbreads.<br /><br />What is also bad about this is that the company states they use biodegradable and eco-friendly packaging. FAIL! They used a HUGE box for a ridiculously small amount of crackers. Not ecofriendly at all.<br /><br />Would I buy these again? No - I feel ripped off. The other crackers (like Sesame Tarragon) give you a little<br />more bang for your buck and have more flavor.",
    "I have used this product in smoothies for my son and he loves it. Additionally, I use this oil in the shower as a skin conditioner and it has made my skin look great. Some of the stretch marks on my belly has disappeared quickly. Highly recommend!!!",
    "Been taking Coconut Oil for YEARS.  This is the best on the retail market.  I wish it was in glass, but this is the one.",
]

query_metadata = [
    {"id": 1, "summary": "Good Quality Dog Food"},
    {"id": 8, "summary": "Nasty No flavor"},
    {"id": 4, "summary": "stale product"},
    {"id": 11, "summary": "Great value and convenient ramen"},
    {"id": 5, "summary": "Great for the kids!"},
    {"id": 2, "summary": "yum falafel"},
    {"id": 9, "summary": "Nearly killed the cats"},
    {"id": 6, "summary": "Price cannot be correct"},
    {"id": 3, "summary": "Taste is neutral, quantity is DECEITFUL!"},
    {"id": 7, "summary": "This stuff is great"},
    {"id": 10, "summary": "The reviews don't lie"},
]
vector_store.add_texts(texts=query, metadatas=query_metadata)
[1, 8, 4, 11, 5, 2, 9, 6, 3, 7, 10]

查询向量存储

一旦您的向量存储已创建并且相关文档已添加,您很可能希望在运行链或代理时对其进行查询。 执行简单的相似性搜索可以按如下方式进行:
# 在查询的嵌入和文档的嵌入之间执行相似性搜索
simsearch_result = vector_store.similarity_search("Good reviews", k=3)
print(simsearch_result)
[Document(metadata={'id': 1, 'summary': 'Good Quality Dog Food'}, page_content='I have bought several of the Vitality canned dog food products and have found them all to be of good quality. The product looks more like a stew than a processed meat and it smells better. My Labrador is finicky and she appreciates this product better than  most.'), Document(metadata={'id': 7, 'summary': 'This stuff is great'}, page_content='I have used this product in smoothies for my son and he loves it. Additionally, I use this oil in the shower as a skin conditioner and it has made my skin look great. Some of the stretch marks on my belly has disappeared quickly. Highly recommend!!!'), Document(metadata={'id': 5, 'summary': 'Great for the kids!'}, page_content="If you are looking for a less messy version of licorice for the children, then be sure to try these!  They're soft, easy to chew, and they don't get your hands all sticky and gross in the car, in the summer, at the beach, etc. We love all the flavos and sometimes mix these in with the chocolate to have a very nice snack! Great item, great price too, highly recommend!")]

过滤支持

向量存储支持一组可以应用于文档元数据字段的过滤器。此功能使开发人员和数据分析师能够细化他们的查询,确保搜索结果与他们的需求准确对齐。通过应用基于特定元数据属性的过滤器,用户可以限制搜索范围,仅关注最相关的数据子集。
# 混合搜索 -> 过滤 id 不等于 1 的情况。
hybrid_simsearch_result = vector_store.similarity_search(
    "Good reviews", k=3, filter={"id": {"$ne": 1}}
)
print(hybrid_simsearch_result)
[Document(metadata={'id': 7, 'summary': 'This stuff is great'}, page_content='I have used this product in smoothies for my son and he loves it. Additionally, I use this oil in the shower as a skin conditioner and it has made my skin look great. Some of the stretch marks on my belly has disappeared quickly. Highly recommend!!!'), Document(metadata={'id': 5, 'summary': 'Great for the kids!'}, page_content="If you are looking for a less messy version of licorice for the children, then be sure to try these!  They're soft, easy to chew, and they don't get your hands all sticky and gross in the car, in the summer, at the beach, etc. We love all the flavos and sometimes mix these in with the chocolate to have a very nice snack! Great item, great price too, highly recommend!"), Document(metadata={'id': 3, 'summary': 'Taste is neutral, quantity is DECEITFUL!'}, page_content='The taste of these white cheddar flat breads is like a regular cracker - which is not bad, except that I bought them because I wanted a cheese taste.<br /><br />What was a HUGE disappointment? How misleading the packaging of the box is. The photo on the box (I bought these in store) makes it look like it is full of long flatbreads (expanding the length and width of the box). Wrong! The plastic tray that holds the crackers is about 2 smaller all around - leaving you with about 15 or so small flatbreads.<br /><br />What is also bad about this is that the company states they use biodegradable and eco-friendly packaging. FAIL! They used a HUGE box for a ridiculously small amount of crackers. Not ecofriendly at all.<br /><br />Would I buy these again? No - I feel ripped off. The other crackers (like Sesame Tarragon) give you a little<br />more bang for your buck and have more flavor.')]

带分数的相似性搜索

如果您想执行相似性搜索并接收相应的分数,可以运行:
simsearch_with_score_result = vector_store.similarity_search_with_score(
    "Not a very good product", k=12
)
print(simsearch_with_score_result)
[(Document(metadata={'id': 3, 'summary': 'Taste is neutral, quantity is DECEITFUL!'}, page_content='The taste of these white cheddar flat breads is like a regular cracker - which is not bad, except that I bought them because I wanted a cheese taste.<br /><br />What was a HUGE disappointment? How misleading the packaging of the box is. The photo on the box (I bought these in store) makes it look like it is full of long flatbreads (expanding the length and width of the box). Wrong! The plastic tray that holds the crackers is about 2 smaller all around - leaving you with about 15 or so small flatbreads.<br /><br />What is also bad about this is that the company states they use biodegradable and eco-friendly packaging. FAIL! They used a HUGE box for a ridiculously small amount of crackers. Not ecofriendly at all.<br /><br />Would I buy these again? No - I feel ripped off. The other crackers (like Sesame Tarragon) give you a little<br />more bang for your buck and have more flavor.'), 0.651870006770711), (Document(metadata={'id': 8, 'summary': 'Nasty No flavor'}, page_content='The candy is just red , No flavor . Just  plan and chewy .  I would never buy them again'), 0.6908952973052638), (Document(metadata={'id': 4, 'summary': 'stale product'}, page_content='Arrived in 6 days and were so stale i could not eat any of the 6 bags!!'), 0.7360955776468822), (Document(metadata={'id': 1, 'summary': 'Good Quality Dog Food'}, page_content='I have bought several of the Vitality canned dog food products and have found them all to be of good quality. The product looks more like a stew than a processed meat and it smells better. My Labrador is finicky and she appreciates this product better than  most.'), 0.7408823529514486), (Document(metadata={'id': 9, 'summary': 'Nearly killed the cats'}, page_content="Too much of a good thing? We worked this kibble in over time, slowly shifting the percentage of Felidae to national junk-food brand until the bowl was all natural. By this time, the cats couldn't keep it in or down. What a mess. We've moved on."), 0.782995248991772), (Document(metadata={'id': 7, 'summary': 'This stuff is great'}, page_content='I have used this product in smoothies for my son and he loves it. Additionally, I use this oil in the shower as a skin conditioner and it has made my skin look great. Some of the stretch marks on my belly has disappeared quickly. Highly recommend!!!'), 0.7912681479906212), (Document(metadata={'id': 2, 'summary': 'yum falafel'}, page_content='We had trouble finding this locally - delivery was fast, no more hunting up and down the flour aisle at our local grocery stores.'), 0.809213468778896), (Document(metadata={'id': 10, 'summary': "The reviews don't lie"}, page_content='Been taking Coconut Oil for YEARS.  This is the best on the retail market.  I wish it was in glass, but this is the one.'), 0.8281482301097155), (Document(metadata={'id': 5, 'summary': 'Great for the kids!'}, page_content="If you are looking for a less messy version of licorice for the children, then be sure to try these!  They're soft, easy to chew, and they don't get your hands all sticky and gross in the car, in the summer, at the beach, etc. We love all the flavos and sometimes mix these in with the chocolate to have a very nice snack! Great item, great price too, highly recommend!"), 0.8283754326400574), (Document(metadata={'id': 6, 'summary': 'Price cannot be correct'}, page_content='Hey, the description says 360 grams - that is roughly 13 ounces at under $4.00 per can. No way - that is the approximate price for a 100 gram can.'), 0.8323967822635847), (Document(metadata={'id': 11, 'summary': 'Great value and convenient ramen'}, page_content="Got these on sale for roughly 25 cents per cup, which is half the price of my local grocery stores, plus they rarely stock the spicy flavors. These things are a GREAT snack for my office where time is constantly crunched and sometimes you can't escape for a real meal. This is one of my favorite flavors of Instant Lunch and will be back to buy every time it goes on sale."), 0.8387189489406939)]
有关您可以在 Azure SQL 向量存储上执行的不同搜索的完整列表,请参阅 API 参考

当您已有要搜索的嵌入时的相似性搜索

# 如果您已有要搜索的嵌入
simsearch_by_vector = vector_store.similarity_search_by_vector(
    [-0.0033353185281157494, -0.017689190804958344, -0.01590404286980629, ...]
)
print(simsearch_by_vector)
[Document(metadata={'id': 8, 'summary': 'Nasty No flavor'}, page_content='The candy is just red , No flavor . Just  plan and chewy .  I would never buy them again'), Document(metadata={'id': 4, 'summary': 'stale product'}, page_content='Arrived in 6 days and were so stale i could not eat any of the 6 bags!!'), Document(metadata={'id': 3, 'summary': 'Taste is neutral, quantity is DECEITFUL!'}, page_content='The taste of these white cheddar flat breads is like a regular cracker - which is not bad, except that I bought them because I wanted a cheese taste.<br /><br />What was a HUGE disappointment? How misleading the packaging of the box is. The photo on the box (I bought these in store) makes it look like it is full of long flatbreads (expanding the length and width of the box). Wrong! The plastic tray that holds the crackers is about 2 smaller all around - leaving you with about 15 or so small flatbreads.<br /><br />What is also bad about this is that the company states they use biodegradable and eco-friendly packaging. FAIL! They used a HUGE box for a ridiculously small amount of crackers. Not ecofriendly at all.<br /><br />Would I buy these again? No - I feel ripped off. The other crackers (like Sesame Tarragon) give you a little<br />more bang for your buck and have more flavor.'), Document(metadata={'id': 6, 'summary': 'Price cannot be correct'}, page_content='Hey, the description says 360 grams - that is roughly 13 ounces at under $4.00 per can. No way - that is the approximate price for a 100 gram can.')]
# 如果您已有要搜索的嵌入,则执行带分数的相似性搜索
simsearch_by_vector_with_score = vector_store.similarity_search_by_vector_with_score(
    [-0.0033353185281157494, -0.017689190804958344, -0.01590404286980629, ...]
)
print(simsearch_by_vector_with_score)
[(Document(metadata={'id': 8, 'summary': 'Nasty No flavor'}, page_content='The candy is just red , No flavor . Just  plan and chewy .  I would never buy them again'), 0.9648153551769503), (Document(metadata={'id': 4, 'summary': 'stale product'}, page_content='Arrived in 6 days and were so stale i could not eat any of the 6 bags!!'), 0.9655108580341948), (Document(metadata={'id': 3, 'summary': 'Taste is neutral, quantity is DECEITFUL!'}, page_content='The taste of these white cheddar flat breads is like a regular cracker - which is not bad, except that I bought them because I wanted a cheese taste.<br /><br />What was a HUGE disappointment? How misleading the packaging of the box is. The photo on the box (I bought these in store) makes it look like it is full of long flatbreads (expanding the length and width of the box). Wrong! The plastic tray that holds the crackers is about 2 smaller all around - leaving you with about 15 or so small flatbreads.<br /><br />What is also bad about this is that the company states they use biodegradable and eco-friendly packaging. FAIL! They used a HUGE box for a ridiculously small amount of crackers. Not ecofriendly at all.<br /><br />Would I buy these again? No - I feel ripped off. The other crackers (like Sesame Tarragon) give you a little<br />more bang for your buck and have more flavor.'), 0.9840511208615808), (Document(metadata={'id': 6, 'summary': 'Price cannot be correct'}, page_content='Hey, the description says 360 grams - that is roughly 13 ounces at under $4.00 per can. No way - that is the approximate price for a 100 gram can.'), 0.9915737524649991)]

从向量存储中删除项目

按 ID 删除行

# 按 id 删除行
vector_store.delete(["3", "7"])
True

删除向量存储

# 删除向量存储
vector_store.drop()

从 Azure Blob 存储加载文档

以下是从 Azure Blob 存储容器加载文件到 SQL 向量存储的示例,首先将文档分割成块。 Azure Blob 存储 是 Microsoft 的云对象存储解决方案。Blob 存储针对存储大量非结构化数据进行了优化。
pip install azure-storage-blob
from langchain.document_loaders import AzureBlobStorageFileLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_core.documents import Document

# 定义您的连接字符串和 blob 详细信息
conn_str = "DefaultEndpointsProtocol=https;AccountName=<YourBlobName>;AccountKey=<YourAccountKey>==;EndpointSuffix=core.windows.net"
container_name = "<YourContainerName"
blob_name = "01 Harry Potter and the Sorcerers Stone.txt"

# 创建 AzureBlobStorageFileLoader 实例
loader = AzureBlobStorageFileLoader(
    conn_str=conn_str, container=container_name, blob_name=blob_name
)

# 从 Azure Blob 存储加载文档
documents = loader.load()

# 如果需要,将文档分割成更小的块
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
split_documents = text_splitter.split_documents(documents)

# 打印分割文档的数量
print(f"Number of split documents: {len(split_documents)}")
Number of split documents: 528
API 参考:AzureBlobStorageContainerLoader
# # 初始化向量存储并将文档及其嵌入插入到 AzureSQLDB 中
vector_store = SQLServer_VectorStore(
    connection_string=_CONNECTION_STRING,
    distance_strategy=DistanceStrategy.COSINE,
    embedding_function=embeddings,
    embedding_length=1536,
    table_name="harrypotter",
)  # 替换为您实际的向量存储初始化

# 将分割的文档单独添加到向量存储中
for i, doc in enumerate(split_documents):
    vector_store.add_documents(documents=[doc], ids=[f"doc_{i}"])

print("Documents added to the vector store successfully!")
Documents added to the vector store successfully!

直接查询

from typing import List, Tuple

# 执行相似性搜索
query = "Why did the Dursleys not want Harry in their house?"
docs_with_score: List[Tuple[Document, float]] = (
    vector_store.similarity_search_with_score(query)
)

for doc, score in docs_with_score:
    print("-" * 60)
    print("Score: ", score)
    print(doc.page_content)
    print("-" * 60)
------------------------------------------------------------
Score:  0.3626232679001803
The Dursleys had everything they wanted, but they also had a secret, and their greatest fear was that somebody would discover it. They didn’t think they could bear it if anyone found out about the Potters. Mrs. Potter was Mrs. Dursley’s sister, but they hadn’t met for several years; in fact, Mrs. Dursley pretended she didn’t have a sister, because her sister and her good-for-nothing husband were as unDursleyish as it was possible to be. The Dursleys shuddered to think what the neighbors would say if the Potters arrived in the street. The Dursleys knew that the Potters had a small son, too, but they had never even seen him. This boy was another good reason for keeping the Potters away; they didn’t want Dudley mixing with a child like that.
------------------------------------------------------------
------------------------------------------------------------
Score:  0.44752797298657554
The Dursleys’ house had four bedrooms: one for Uncle Vernon and Aunt Petunia, one for visitors (usually Uncle Vernon’s sister, Marge), one where Dudley slept, and one where Dudley kept all the toys and things that wouldn’t fit into his first bedroom. It only took Harry one trip upstairs to move everything he owned from the cupboard to this room. He sat down on the bed and stared around him. Nearly everything in here was broken. The month-old video camera was lying on top of a small, working tank Dudley had once driven over the next door neighbor’s dog; in the corner was Dudley’s first-ever television set, which he’d put his foot through when his favorite program had been canceled; there was a large birdcage, which had once held a parrot that Dudley had swapped at school for a real air rifle, which was up on a shelf with the end all bent because Dudley had sat on it. Other shelves were full of books. They were the only things in the room that looked as though they’d never been touched.
------------------------------------------------------------
------------------------------------------------------------
Score:  0.4652486419877385
M r. and Mrs. Dursley, of number four, Privet Drive, were proud to say that they were perfectly normal, thank you very much. They were the last people you’d expect to be involved in anything strange or mysterious, because they just didn’t hold with such nonsense.

Mr. Dursley was the director of a firm called Grunnings, which made drills. He was a big, beefy man with hardly any neck, although he did have a very large mustache. Mrs. Dursley was thin and blonde and had nearly twice the usual amount of neck, which came in very useful as she spent so much of her time craning over garden fences, spying on the neighbors. The Dursleys had a small son called Dudley and in their opinion there was no finer boy anywhere.
------------------------------------------------------------
------------------------------------------------------------
Score:  0.4739086301927252
Hagrid was watching him sadly.

“Took yeh from the ruined house myself, on Dumbledore’s orders. Brought yeh ter this lot….”

“Load of old tosh,” said Uncle Vernon. Harry jumped; he had almost forgotten that the Dursleys were there. Uncle Vernon certainly seemed to have got back his courage. He was glaring at Hagrid and his fists were clenched.

“Now, you listen here, boy,” he snarled, “I accept there’s something strange about you, probably nothing a good beating wouldn’t have cured — and as for all this about your parents, well, they were weirdoes, no denying it, and the world’s better off without them in my opinion — asked for all they got, getting mixed up with these wizarding types — just what I expected, always knew they’d come to a sticky end —”

But at that moment, Hagrid leapt from the sofa and drew a battered pink umbrella from inside his coat. Pointing this at Uncle Vernon like a sword, he said, “I’m warning you, Dursley — I’m warning you — one more word….”
------------------------------------------------------------

用于检索增强生成的用例

用例 1:基于故事书的问答系统

问答功能允许用户询问关于故事、角色和事件的具体问题,并获得简洁、上下文丰富的答案。这不仅增强了他们对书籍的理解,还让他们感觉自己是魔法世界的一部分。

通过转换为检索器进行查询

LangChain 向量存储通过启用高效的相似性搜索来简化构建复杂的问答系统,根据用户查询找到前 10 个相关文档。检索器vector_store 创建,问答链使用 create_stuff_documents_chain 函数构建。使用 ChatPromptTemplate 类创建提示模板,确保结构化和上下文丰富的响应。在问答应用中,向用户展示用于生成答案的来源通常很重要。LangChain 内置的 create_retrieval_chain 会将检索到的源文档传播到输出中的 “context” 键下: 阅读更多关于 LangChain RAG 教程和术语 的信息。
from typing import List, Tuple

import pandas as pd
from langchain_classic.chains import create_retrieval_chain
from langchain_classic.chains.combine_documents import create_stuff_documents_chain
from langchain_core.prompts import ChatPromptTemplate


# 定义执行 RAG 链调用的函数
def get_answer_and_sources(user_query: str):
    # 执行带分数的相似性搜索
    docs_with_score: List[Tuple[Document, float]] = (
        vector_store.similarity_search_with_score(
            user_query,
            k=10,
        )
    )

    # 从顶部结果中提取上下文
    context = "\n".join([doc.page_content for doc, score in docs_with_score])

    # 定义系统提示
    system_prompt = (
        "You are an assistant for question-answering tasks based on the story in the book. "
        "Use the following pieces of retrieved context to answer the question. "
        "If you don't know the answer, say that you don't know, but also suggest that the user can use the fan fiction function to generate fun stories. "
        "Use 5 sentences maximum and keep the answer concise by also providing some background context of 1-2 sentences."
        "\n\n"
        "{context}"
    )

    # 创建提示模板
    prompt = ChatPromptTemplate.from_messages(
        [
            ("system", system_prompt),
            ("human", "{input}"),
        ]
    )

    # 创建检索器和链
    retriever = vector_store.as_retriever()
    question_answer_chain = create_stuff_documents_chain(llm, prompt)
    rag_chain = create_retrieval_chain(retriever, question_answer_chain)

    # 定义输入
    input_data = {"input": user_query}

    # 调用 RAG 链
    response = rag_chain.invoke(input_data)

    # 打印答案
    print("Answer:", response["answer"])

    # 准备表格数据
    data = {
        "Doc ID": [
            doc.metadata.get("source", "N/A").split("/")[-1]
            for doc in response["context"]
        ],
        "Content": [
            doc.page_content[:50] + "..."
            if len(doc.page_content) > 100
            else doc.page_content
            for doc in response["context"]
        ],
    }

    # 创建 DataFrame
    df = pd.DataFrame(data)

    # 打印表格
    print("\nSources:")
    print(df.to_markdown(index=False))
# 定义用户查询
user_query = "How did Harry feel when he first learnt that he was a Wizard?"

# 调用函数获取答案和来源
get_answer_and_sources(user_query)
Answer: When Harry first learned that he was a wizard, he felt quite sure there had been a horrible mistake. He struggled to believe it because he had spent his life being bullied and mistreated by the Dursleys. If he was really a wizard, he wondered why he hadn't been able to use magic to defend himself. This disbelief and surprise were evident when he gasped, “I’m a what?”

Sources:
| Doc ID                                      | Content                                               |
|:--------------------------------------------|:------------------------------------------------------|
| 01 Harry Potter and the Sorcerers Stone.txt | Harry was wondering what a wizard did once he’d fi... |
| 01 Harry Potter and the Sorcerers Stone.txt | Harry realized his mouth was open and closed it qu... |
| 01 Harry Potter and the Sorcerers Stone.txt | “Most of us reckon he’s still out there somewhere ... |
| 01 Harry Potter and the Sorcerers Stone.txt | “Ah, go boil yer heads, both of yeh,” said Hagrid.... |
# 定义用户查询
user_query = "Did Harry have a pet? What was it"

# 调用函数获取答案和来源
get_answer_and_sources(user_query)
Yes, Harry had a pet owl named Hedwig. He decided to call her Hedwig after finding the name in a book titled *A History of Magic*.

Sources:
| Doc ID                                      | Content                                               |
|:--------------------------------------------|:------------------------------------------------------|
| 01 Harry Potter and the Sorcerers Stone.txt | Harry sank down next to the bowl of peas. “What di... |
| 01 Harry Potter and the Sorcerers Stone.txt | Harry kept to his room, with his new owl for compa... |
| 01 Harry Potter and the Sorcerers Stone.txt | As the snake slid swiftly past him, Harry could ha... |
| 01 Harry Potter and the Sorcerers Stone.txt | Ron reached inside his jacket and pulled out a fat... |

API 参考

有关 SQLServer 向量存储功能和配置的详细文档,请访问 API 参考:https://python.langchain.com/api_reference/sqlserver/index.html

相关内容