Build a chatbot to query your documents using Langchain and Azure OpenAI (2023)

In this article, I introduce LangChain and explore its capabilities by building a simple question answer to query PDFs in the Azure Functions documentation.

wavy necklace

Harrison Chasewavy necklaceis a powerful Python library that simplifies the process of building NLP applications with large language models. The main goal is to create intelligent agents that can understand and execute human language instructions. With LangChain, you can connect to various data and computing resources and build applications that perform NLP tasks on domain-specific data sources, private repositories, and more.

As of May 2023, the LangChain GitHub repository has received over 42,000 stars and contributions from over 270 developers around the world.

The langchain library consists of several modules:

• LLM and tips

This includes just-in-time management, just-in-time optimization, a common interface for all LLMs, and common tools for working with LLMs such as Azure OpenAI. It supports several LLMs, including OpenAI, LLama, and GPT4All.

• the chain

A chain in LangChain refers to a series of calls that can be chained together to perform a specific task. For example, you may need to retrieve data from a specific URL, summarize the returned text, and use the resulting summary to answer questions. Chains can also be as simple as reading user input, building a prompt, and generating a response.

• Data augmentation generation

Data augmentation generation involves specific types of chains that first communicate with external data sources to obtain data for use in the generation step. Examples are long text summaries and question/answer for specific data sources. LangChaindocumentladerIntoolModules help connect to data sources and calculations. If you combine text files, PDF documents, HTML web pages, etc., you can use the document loader in Langchain.

• intermediary

Agency involves the LLM in deciding what action to take, taking that action, observing the observation, and repeating it until it is done.

As we explained earlier, chains can help chain together a series of LLM calls. However, in some tasks, the sequence of calls is often undefined, and the next step will depend on user input and responses from previous steps.

“intermediary' can respond to input along the way instead of a hard-coded deterministic sequence.

• memory

Memory refers to the persistent state using VectorStores. Vector databases are optimized for fast searching in high-dimensional spaces. LangChain makes this effortless.

integrate

Embeddings are the mapping of discrete categorical variables to continuous numerical vectors.In the context of neural networks, embeddingis a low-dimensional learned continuous vector representation of discrete variables. Embedding neural networks is useful because they reduce the dimensionality of categorical variables and represent categories in the transformed space in a meaningful way.

Neural network embedding has 3 main goals:

• Find nearest neighbors in the embedding space. These can be used to make recommendations based on user interests or cluster categories.
• As input to machine learning models for supervised tasks.
• Used for visualization of relationships between concepts and categories.

Clustering Example of Sentence Vector Values

vectoropslag of vectordatabase

A vector database is a special type of database that stores data as high-dimensional vectors. These vectors are mathematical representations of characteristics or properties of the stored data. The number of dimensions in each vector can vary widely, from tens to thousands, depending on the complexity and detail of the data. In this article, we will explore the concept of vector databases and their applications in various fields.

Let's build the application

Let's build a tool that can read developer documentation– in this case the Azure Functions documentation in PDF format.

TonThen answer each question by referring to the document text.

We will follow these steps:

One-time procedure:

• Indexes pdf documents (azure function docs), divides documents into pieces, indexes all created embedded text.
• Store all embeddings in a vector store (in our case Faiss) which is searchable in the application.

application:

• When a user asks a question, we use the FAISS vector index to find the best matching text.
• Enter this as context in the prompt for GPT-3.5
• GPT-3.5 generates answers that accurately answer questions.

tempo

• Download documents to search. In our case, we can download the Azure Functions documentation fromhereand save it in the data/documents folder.
• Deploy to Azure OpenAI
  • Ada
  • GPT35

Get the Azure OpenAI endpoint and key and add them to a file named .env as follows:

OPENAI_DEPLOYMENT_ENDPOINT = "https://.openai.azure.com/" OPENAI_API_KEY = "" OPENAI_DEPLOYMENT_NAME = "" OPENAI_DEPLOYMENT_VERSION = "2023-03-15-preview" OPENAI_MODEL_NAME="gpt-35-turbo"OPENAI_EMBEDDING_DEPLOYMENT_NAME=""OPENAI_EMBEDDING_MODEL_NAME="text-embedding-ada-002"

create containment

trafficindexer.pyAnd:

• pdf barn
• Split all text into chunks.
• These chunks are sent to the OpenAI Embedding API, which returns a 1536-dimensional vector for each chunk.
• Index all vectors in a FAISS index.
• Save the FAISS index in a .To trustIn.pkldocument.

Note: As you probably know, LLM cannot accept long statements due to token restrictions, so we've split the document into chunks, see below.

Running this code takes time because we have to read and split the entire document and send the chunks to the Ada model to get the embeddings.

Here's the codeindexer.py

from langchain.document_loaders import PyPDFLoader from langchain.embeddings.openai import OpenAIEmbeddings from langchain.vectorstores import FAISSfrom dotenv import load_dotenvimport openaiimport os#load environment variablesload_dotenv()OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")OPENAI_DEPLOYMENT_ENDPOINT = os.get环境（“OPENAI_DEPLOYMENT_ENDPOINT ”）OPENAI_DEPLOYMENT_NAME = os.getenv("OPENAI_DEPLOYMENT_NAME")OPENAI_MODEL_NAME = os.getenv("OPENAI_MODEL_NAME")OPENAI_EMBEDDING_DEPLOYMENT_NAME = os.getenv("OPENAI_EMBEDDING_DEPLOYMENT_NAME")OPENAI_EMBEDDING_MODEL_NAME = os.getenv("OPENAI_EMBEDDING _MODEL_NAME")OPENAI_DEPLOYMENT_VERSION = os.getenv(" OPENAI_DEPLOYMENT_VERSION")# init Azure OpenAIopenai.api_type = "azure"openai.api_version = OPENAI_DEPLOYMENT_VERSIONopenai.api_base = OPENAI_DEPLOYMENT_ENDPOINTopenai.api_key = OPENAI_API_KEYif __name__ == "__main__": embeddings = OpenAIEmbeddings(模型=OP ENAI_EMBEDDING_MODEL_NAME, chunk_size=1 ) dataPath = "./ data/documentation/ " fileName = dataPath + "azure-azure-functions.pdf" #use langchain PDF loader loader = PyPDFLoader(fileName) #split the document in chunks pages = loader.load_and_split() #Use Langchain 使用文本嵌入创建嵌入ada-002 db = FAISS.from_documents(documents=pages, embedding=embeddings) #将嵌入保存到 FAISS 向量存储中 db.save_local("./dbs/documentation/faiss_index")

make application

trafficapplicatie.pyWorks as:

• FAISS index loaded in RAM
• user question
• The user's query is sent to the OpenAI Embeddings API, which returns a 1536-dimensional vector.
• Look up the FAISS index for the closest vector match.
• Returns the closest vector and the text it produced.
• The returned text is entered in GPT-35 as context in the GPT-35 prompt
• GPT-35 generates a response that is sent back to the user.

NOTE: The important thing to note is that Langchain does most of the heavy lifting for us, behind the scenes.

Here's the codeapplicatie.py

从 langchain.vectorstores 手机 FAISSfrom dotenv import load_dotenvimport openaiimport osfrom langchain.chains import RetrievalQAfrom langchain.vectorstores 拉手 FAISSfrom langchain.chains.question_answering import load_qa_chainfrom dotenv import load_dotenvfrom langchain.chat_models import AzureChatOpenAIfrom langchain.embedding s.openai import OpenAIEmbed dings# 英语下载在电影load_dotenv ( )OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")OPENAI_DEPLOYMENT_ENDPOINT = os.getenv("OPENAI_DEPLOYMENT_ENDPOINT")OPENAI_DEPLOYMENT_NAME = os.getenv("OPENAI_DEPLOYMENT_NAME")OPENAI_MODEL_NAME = os.getenv("OPENAI_DEPLOYMENT_ENDPOINT") v( "OPENAI_MODEL_NAME ") DDING_DEPLOYMENT_NAME ")OPENAI_EMBEDDING_MODEL_NAME = os.getenv("OPENAI_EMBEDDING_MODEL_NAME")OPENAI_DEPLOYMENT_VERSION = os.getenv("OPENAI_DEPLOYMENT_VERSION")#init Azure OpenAIopenai.api_type = "azure" openai.api_base = OPENAI_DEPLOYMENT_VERSION OYMENT_ENDPOINTopenai.api_key = OPENAI_API_KEYdef ask_question( qa, vraag ): result= qa({"query": vraag}) print("Vraag:", vraag) print("Antwoord:", result["resultat"])if __name__ == "__main__" : #init openai llm = AzureChatOpenAI(deployment_name =OPENAI_DEPLOYMENT_NAME，model_name=OPENAI_MODEL_NAME，openai_api_base=OPENAI_DEPLOYMENT_ENDPOINT，openai_api_version=OPENAI_DEPLOYMENT_VERSION，openai_api_key=OPENAI_API_KEY) embeddings = OpenAI Embeddings(model=OPENAI_EMBEDDING_MODEL_NAME，chunk_size=1) store vector Store = FAISS.load_local(". / dbs/documentation /faiss_index", embeddings) #使用我们的faiss影兰最作来杰地地Documentation retriever = vectorStore.as_retriever(search_type="similarity", search_kwargs={"k":2}) 更多之是电视器 qa = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever, return_source_documents=False) while True: query = input('you: ') if query == 'q': break ask_question ( qa, inquire)

Now we can runapplication.py and start asking:

You: what is an azure function? Question: What is an Azure function? Answer: Azure Functions is an on-demand cloud service that provides all the constantly updated infrastructure and resources needed to run your applications. You can use the most efficient language to focus on the code that matters most to you, and Functions will do the rest. Functions provides serverless computing for Azure. You can use Functions to build web APIs, respond to database changes, handle IoT flows, manage message queues, and more.

You: Can I use Event Hub as a trigger for an Azure function? Question: Can I use Event Hub as a trigger for an Azure function? Answer: Yes, you can use Event Hubs as a trigger for an Azure function. `Azure Functions supports event hub triggers and output bindings. Use function triggers to respond to events sent to the Event Hubs event stream. You must have read access to the underlying event hub to set triggers. When the function is triggered, the message passed to the function is typed as a string. `<|im_end|>

You: What languages ​​can I use to build Azure functions? Question: What languages ​​can I use to build Azure functions?

You: Can I deploy Azure features in multiple regions? Question: Can I deploy Azure functions in multiple regions? Answer: Yes, you can deploy Azure features in multiple regions. There are two modes to consider: active/active for HTTP-triggered functions and active/passive for event-driven non-HTTP-triggered functions. When using active/active mode for HTTP-triggered functions, Azure Front Door is required to coordinate requests between the two regions. When using active/passive mode, a second region is activated and takes over processing when failover is required. For more information on multi-region deployments, see our guide in Highly Available Multi-Region Web Applications. <|im_end|>

Thanks for reading, I hope you enjoy it.

denis