LLM之RAG实战（十九）| 利用LangChain、OpenAI、ChromaDB和Streamlit构建RAG

生成式人工智能以其创造与上下文相关内容的能力彻底改变了技术，开创了人工智能可能性的新时代。其核心是检索增强生成（RAG），将信息检索与LLM相结合，从外部文档中产生智能、知情的响应。

本文将深入研究使用ChromaDB构建RAG驱动的LLM应用程序，ChromaDB以其对大型数据集的高效处理而闻名。

一、环境准备

要构建基于RAG的LLM应用程序，需要准备如下环境配置：

python（下载地址：https://www.python.org/downloads/）
OpenAI API Key（获取地址：https://platform.openai.com/signup）

以及对Python和web API的基本理解。

二、代码实现

2.1 创建并导航到项目目录

mkdir rag_lmm_applicationcd rag_lmm_application

2.2 创建虚拟环境

虚拟环境可以隔离不同的python环境，创建命令如下所示：

python -m venv venv

激活虚拟环境。对于Mac/Linux用户，请使用：

source venv/bin/activate

对于Windows用户：

venv\Scripts\activate

2.3 安装所需的包

安装基本库：

pip install -r requirements.txt

PS：确保requirements.txt文件中包含所有必要的依赖项。

通过上述步骤，环境已经准备就绪，下面开始使用ChromaDB构建最先进的RAG聊天应用程序。

2.4 加载和处理文档

下面使用LangChain来加载各种文档格式，如PDF、DOCX和TXT，这对于外部数据访问、确保高效的数据处理以及为后续阶段保持统一的数据准备至关重要。代码如下：

# loading PDF, DOCX and TXT files as LangChain Documentsdef load_document(file):    import os    name, extension = os.path.splitext(file)    if extension == '.pdf':        from langchain.document_loaders import PyPDFLoader        print(f'Loading {file}')        loader = PyPDFLoader(file)    elif extension == '.docx':        from langchain.document_loaders import Docx2txtLoader        print(f'Loading {file}')        loader = Docx2txtLoader(file)    elif extension == '.txt':        from langchain.document_loaders import TextLoader        loader = TextLoader(file)    else:        print('Document format is not supported!')        return None    data = loader.load()    return data

数据分块对RAG系统非常重要，对数据进行分块以便于嵌入，这确保可以保留高效的上下文和信息检索。代码如下：

# splitting data in chunksdef chunk_data(data, chunk_size=256, chunk_overlap=20):    from langchain.text_splitter import RecursiveCharacterTextSplitter    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)    chunks = text_splitter.split_documents(data)    return chunks

2.5 使用OpenAI和ChromaDB创建嵌入

使用OpenAI大模型来创建嵌入，并将它们高效地存储在ChromaDB中，可以快速检索信息，代码如下所示：

# create embeddings using OpenAIEmbeddings() and save them in a Chroma vector storedef create_embeddings(chunks):    embeddings = OpenAIEmbeddings()    vector_store = Chroma.from_documents(chunks, embeddings)    # if you want to use a specific directory for chromadb    # vector_store = Chroma.from_documents(chunks, embeddings, persist_directory='./mychroma_db')    return vector_store

2.6 使用Streamlight构建聊天界面

Streamlit的简单性在我们的RAG LLM应用程序中大放异彩，可以毫不费力地将用户输入链接到后端处理。通过Streamlit的初始化和布局设计，用户可以上传文档和管理数据。后端处理这些输入，直接在Streamlit界面中返回响应，展示了前端和后端操作的无缝集成。下面是一个代码片段来说明设置：

def ask_and_get_answer(vector_store, q, k=3):    from langchain.chains import RetrievalQA    from langchain.chat_models import ChatOpenAI    llm = ChatOpenAI(model='gpt-3.5-turbo', temperature=1)    retriever = vector_store.as_retriever(search_type='similarity', search_kwargs={'k': k})    chain = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever)    answer = chain.run(q)    return answer# calculate embedding cost using tiktokendef calculate_embedding_cost(texts):    import tiktoken    enc = tiktoken.encoding_for_model('text-embedding-ada-002')    total_tokens = sum([len(enc.encode(page.page_content)) for page in texts])    # print(f'Total Tokens: {total_tokens}')    # print(f'Embedding Cost in USD: {total_tokens / 1000 * 0.0004:.6f}')    return total_tokens, total_tokens / 1000 * 0.0004# clear the chat history from streamlit session statedef clear_history():    if 'history' in st.session_state:        del st.session_state['history']if __name__ == "__main__":    import os    # loading the OpenAI api key from .env    from dotenv import load_dotenv, find_dotenv    load_dotenv(find_dotenv(), override=True)    st.image('img.png')    st.subheader('LLM Question-Answering Application 🤖')    with st.sidebar:        # text_input for the OpenAI API key (alternative to python-dotenv and .env)        api_key = st.text_input('OpenAI API Key:', type='password')        if api_key:            os.environ['OPENAI_API_KEY'] = api_key        # file uploader widget        uploaded_file = st.file_uploader('Upload a file:', type=['pdf', 'docx', 'txt'])        # chunk size number widget        chunk_size = st.number_input('Chunk size:', min_value=100, max_value=2048, value=512, on_change=clear_history)        # k number input widget        k = st.number_input('k', min_value=1, max_value=20, value=3, on_change=clear_history)        # add data button widget        add_data = st.button('Add Data', on_click=clear_history)        if uploaded_file and add_data: # if the user browsed a file            with st.spinner('Reading, chunking and embedding file ...'):                # writing the file from RAM to the current directory on disk                bytes_data = uploaded_file.read()                file_name = os.path.join('./', uploaded_file.name)                with open(file_name, 'wb') as f:                    f.write(bytes_data)                data = load_document(file_name)                chunks = chunk_data(data, chunk_size=chunk_size)                st.write(f'Chunk size: {chunk_size}, Chunks: {len(chunks)}')                tokens, embedding_cost = calculate_embedding_cost(chunks)                st.write(f'Embedding cost: ${embedding_cost:.4f}')                # creating the embeddings and returning the Chroma vector store                vector_store = create_embeddings(chunks)                # saving the vector store in the streamlit session state (to be persistent between reruns)                st.session_state.vs = vector_store                st.success('File uploaded, chunked and embedded successfully.')

上面的代码显示了如何在Streamlit中创建文本输入字段并处理用户输入。通过这种设置，用户可以无缝直观地与人工智能应用程序交互。

2.7 检索答案和增强用户交互

我们的RAG聊天应用程序利用了LangChain的RetrievalQA和ChromaDB，通过从ChromaDB的嵌入式数据中提取的相关、准确的信息有效地响应用户查询，体现了先进的Generative AI功能。

下面的代码片段展示了Streamlit应用程序的实现：

# user's question text input widget    q = st.text_input('Ask a question about the content of your file:')    if q: # if the user entered a question and hit enter        if 'vs' in st.session_state: # if there's the vector store (user uploaded, split and embedded a file)            vector_store = st.session_state.vs            st.write(f'k: {k}')            answer = ask_and_get_answer(vector_store, q, k)            # text area widget for the LLM answer            st.text_area('LLM Answer: ', value=answer)            st.divider()            # if there's no chat history in the session state, create it            if 'history' not in st.session_state:                st.session_state.history = ''            # the current question and answer            value = f'Q: {q} \nA: {answer}'            st.session_state.history = f'{value} \n {"-" * 100} \n {st.session_state.history}'            h = st.session_state.history            # text area widget for the chat history            st.text_area(label='Chat History', value=h, key='history', height=400)