structured-docx-loader 0.1.0

LangChain document loader for structured Microsoft Word (.docx) files using python-docx.

structured-docx-loader

A LangChain BaseLoader for Microsoft Word (.docx) files that preserves document structure instead of flattening it into one undifferentiated blob of text.

langchain-community's existing Word loaders either dump raw text (Docx2txtLoader) or depend on the heavyweight unstructured library (UnstructuredWordDocumentLoader). DocxLoader uses python-docx directly to walk the document in its native order and:

• Renders heading styles (Heading 1-Heading 9) as Markdown headings, preserving hierarchy.
• Converts tables to Markdown (default), HTML, or a key-value row format suitable for retrieval.
• Supports three loading granularities: a single document, one document per heading section, or one document per paragraph/table element.

Install

pip install structured-docx-loader

Usage

from structured_docx_loader import DocxLoader

# Load the entire document as a single Document
loader = DocxLoader("example.docx")
docs = loader.load()

# Split by heading sections, with HTML tables
loader = DocxLoader("example.docx", mode="sections", table_format="html")
docs = loader.load()

# One Document per paragraph/table row, tables as key-value pairs
loader = DocxLoader(
    "example.docx",
    mode="elements",
    table_format="key_value",
    table_extraction_strategy="row",
)
docs = loader.load()

file_path also accepts an HTTP(S) URL, in which case the file is downloaded to a temporary location before parsing.

Options

Argument	Values	Description
mode	"single" (default), "sections", "elements"	Granularity of the returned Document objects.
table_format	"markdown" (default), "html", "key_value"	How tables are rendered into text.
table_extraction_strategy	"table" (default), "row"	Whether a table becomes one block or one block per row.

Development

pip install -e ".[test,lint,typing]"
pytest
ruff check .
mypy structured_docx_loader

License

MIT