使用Langchain查询Word文档报错及本地离线查询咨询

尝试用Langchain批量查询Word文档时出现报错，Traceback如下：

Traceback (most recent call last):

  File C:\Program Files\Spyder\pkgs\spyder_kernels\py3compat.py:356 in compat_exec
    exec(code, globals, locals)

  File c:\data\langchain\langchaintest.py:44
    index = VectorstoreIndexCreator().from_loaders(loaders)

  File ~\AppData\Roaming\Python\Python38\site-packages\langchain\indexes\vectorstore.py:72 in from_loaders
    docs.extend(loader.load())

  File ~\AppData\Roaming\Python\Python38\site-packages\langchain\document_loaders\text.py:17 in load
    with open(self.file_path, encoding=self.encoding) as f:

OSError: [Errno 22] Invalid argument:

注：报错信息中invalid argument: 后跟随的是Word文档的原始文本

使用的代码如下：

import os
os.environ["OPENAI_API_KEY"] = "xxxxxx"


import os
import docx
from langchain.document_loaders import TextLoader

# Function to get text from a docx file
def get_text_from_docx(file_path):
    doc = docx.Document(file_path)
    full_text = []
    for paragraph in doc.paragraphs:
        full_text.append(paragraph.text)
    
    return '\n'.join(full_text)

# Load multiple Word documents
folder_path = 'C:/Data/langchain'
word_files = [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.endswith('.docx')]

loaders = []
for word_file in word_files:
    text = get_text_from_docx(word_file)
    loader = TextLoader(text)
    loaders.append(loader)
    
    
from langchain.indexes import VectorstoreIndexCreator

index = VectorstoreIndexCreator().from_loaders(loaders)

query = "What are the main points discussed in the documents?"

responses = index.query(query)
print(responses)

results_with_source=index.query_with_sources(query)
print(results_with_source)

需要解决两个问题：

1. 报错中TextLoader需要传入的正确参数是什么？
2. 能否在无网络、不依赖OpenAI的本地环境实现此类查询？

1. 修复TextLoader参数错误 #

报错原因是TextLoader的作用是加载本地文本文件，它的构造参数需要传入文件路径字符串，而你传入的是从Word文档中提取的原始文本内容，导致系统尝试把文本内容当作文件路径去打开，自然触发无效参数错误。

有两种修复方式：

方式一：直接使用Langchain的DocxLoader #

Langchain已提供专门加载Word文档的DocxLoader，无需自定义文本提取函数，简化代码：

import os
os.environ["OPENAI_API_KEY"] = "xxxxxx"

from langchain.document_loaders import DocxLoader
from langchain.indexes import VectorstoreIndexCreator

folder_path = 'C:/Data/langchain'
word_files = [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.endswith('.docx')]

loaders = [DocxLoader(file) for file in word_files]

index = VectorstoreIndexCreator().from_loaders(loaders)

query = "What are the main points discussed in the documents?"
responses = index.query(query)
print(responses)
results_with_source=index.query_with_sources(query)
print(results_with_source)

方式二：保留自定义文本提取，改用Document对象 #

如果需要自定义Word文本提取逻辑，不要用TextLoader，直接构造Document对象后用VectorstoreIndexCreator的from_documents方法：

import os
os.environ["OPENAI_API_KEY"] = "xxxxxx"

import docx
from langchain.docstore.document import Document
from langchain.indexes import VectorstoreIndexCreator

def get_text_from_docx(file_path):
    doc = docx.Document(file_path)
    full_text = []
    for paragraph in doc.paragraphs:
        full_text.append(paragraph.text)
    return '\n'.join(full_text)

folder_path = 'C:/Data/langchain'
word_files = [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.endswith('.docx')]

documents = []
for file in word_files:
    text = get_text_from_docx(file)
    doc = Document(page_content=text, metadata={"source": file})
    documents.append(doc)

index = VectorstoreIndexCreator().from_documents(documents)

query = "What are the main points discussed in the documents?"
responses = index.query(query)
print(responses)
results_with_source=index.query_with_sources(query)
print(results_with_source)

2. 本地无网络、不依赖OpenAI的实现方案 #

可以实现，需替换Langchain中依赖OpenAI的组件为本地开源替代项，核心替换两部分：

• 嵌入模型：用sentence-transformers系列（如all-MiniLM-L6-v2）
• 大语言模型：用本地部署的开源模型（如Llama 2、Qwen、Mistral等），通过CTransformers或HuggingFacePipeline加载

示例代码（以sentence-transformers为嵌入模型，本地Llama 2量化模型为例）：

import os
import docx
from langchain.docstore.document import Document
from langchain.indexes import VectorstoreIndexCreator
# 本地嵌入模型
from langchain.embeddings import HuggingFaceEmbeddings
# 本地LLM加载器
from langchain.llms import CTransformers

# 初始化本地嵌入模型
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")

# 初始化本地LLM（假设已下载Llama 2量化模型文件llama-2-7b-chat.ggmlv3.q4_0.bin）
llm = CTransformers(
    model="llama-2-7b-chat.ggmlv3.q4_0.bin",
    model_type="llama",
    config={"max_new_tokens": 256, "temperature": 0.01}
)

# 加载Word文档并构造Document对象
def get_text_from_docx(file_path):
    doc = docx.Document(file_path)
    full_text = []
    for paragraph in doc.paragraphs:
        full_text.append(paragraph.text)
    return '\n'.join(full_text)

folder_path = 'C:/Data/langchain'
word_files = [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.endswith('.docx')]

documents = []
for file in word_files:
    text = get_text_from_docx(file)
    doc = Document(page_content=text, metadata={"source": file})
    documents.append(doc)

# 创建索引时指定本地嵌入模型和LLM
index = VectorstoreIndexCreator(
    embedding=embeddings,
    llm=llm
).from_documents(documents)

query = "What are the main points discussed in the documents?"
responses = index.query(query)
print(responses)
results_with_source=index.query_with_sources(query)
print(results_with_source)

注意：本地运行需提前下载对应模型文件，确保硬件性能足够（7B量化模型至少需8GB以上内存，GPU显存充足可提升速度）

内容的提问来源于stack exchange，提问作者bitterjam