OpenDocs

Language: English | 简体中文

OpenDocs is a Python SDK that converts local documents into clean Markdown — TXT, Markdown, images, PDF (native / hybrid / vision), DOCX, PPTX, and XLSX — through a unified sync/async API.

Package name: opendocs-sdk  |  Current Alpha: opendocs-sdk==0.2.0  |  Import name: opendocs  |  Python: 3.11+

Install

pip install opendocs-sdk

If you need to work from a local checkout (for development or unreleased changes):

uv add ../OpenDocs       # or: pip install ../OpenDocs

Quick start

Synchronous local-path example:

from opendocs import parse

markdown = parse("notes.md")

Asynchronous bytes example:

import asyncio

from opendocs import aparse


async def main() -> str:
    markdown = await aparse(b"plain text")
    return markdown


markdown = asyncio.run(main())

Accepted inputs:

• local filesystem path as str
• local filesystem path as os.PathLike[str]
• bytes
• binary file object with read() -> bytes

Callers own all remote downloads. http://, https://, oss://, and s3:// sources must be downloaded before calling OpenDocs.

Supported formats

Format	Status	Notes
TXT	✅	Parsed end to end into deterministic Markdown
Markdown (.md, .markdown)	✅	Preserved for named Markdown paths/streams; unnamed UTF-8 bytes detect as TXT
PDF	✅	Per-page native, hybrid, full-vision, or blank routing; source-ordered page boundaries and tables
PNG / JPEG / WebP	✅	Static images only; sanitized before the configured vision model sees them
DOCX	✅	Continuous authored body flow with structured text, lists, links, tables, explicit breaks, and inline images
PPTX	✅	Slide and shape-tree order with text, tables, accessible charts, groups, and inline images
XLSX (.xlsx)	✅	All sheet-like entries in source order, saved values, tables/regions, merges, standard text objects, native chart facts, and optional visual interpretation

Only standard .xlsx workbooks are supported. Legacy .xls, macro-enabled .xlsm, binary .xlsb, and other spreadsheet formats are not accepted.

Vision parsing

Image and visual PDF work uses the provider-neutral LiteLLM adapter when a VisionConfig is supplied. Native and blank PDF pages do not call a model.

Poppler (pdftoppm) is required only for visual/hybrid PDF parsing — install it with your platform package manager:

# macOS
brew install poppler

# Debian / Ubuntu
apt-get install poppler-utils

Standalone images require VisionConfig. PDFs and Office documents without vision configuration preserve usable native content and emit deterministic warnings for visual regions; XLSX always keeps its native sheet and chart facts when visual enrichment is unavailable. A document with no usable native content raises VisionRequiredError where that format requires vision.

from opendocs import ParseOptions, VisionConfig, parse

markdown = parse(
    "scan.pdf",
    options=ParseOptions(timeout=300, max_pages=100, vision_concurrency=4),
    vision=VisionConfig(
        model="openai/gpt-4o-mini",
        api_key="...",  # Prefer an environment-backed secret in production.
    ),
)

ParseOptions controls document timeout, PPTX/PDF page count, output size, and visual concurrency. Model failures — authentication, permission, invalid request, temporary unavailability, invalid response — use distinct typed exceptions for precise error handling.

ParseOptions.vision_concurrency limits visual requests within one parse. Applications control cross-document concurrency themselves (e.g. with an asyncio.Semaphore); see the independent consumer example.

DOCX, PPTX & XLSX details

DOCX extraction preserves authored body paragraphs, headings, lists, safe links, tables, merged cells, explicit page breaks, and inline raster-image positions. A DOCX remains one continuous logical flow; max_pages does not infer physical Word pages.

PPTX extraction emits every slide boundary and traverses each slide's shape tree in source order, including recursive groups, text, tables, accessible chart data, and raster pictures. Exact duplicate embedded images are analyzed once per parse and replayed at every authored slot.

XLSX extraction emits every worksheet and chartsheet in workbook order, including visible, hidden, very hidden, and empty sheets. It preserves non-empty regions, Excel tables, merged-cell spans, standard comments/text boxes/links/header-footer text, and common saved display semantics such as $, €, £, and ¥ currency, grouping, decimals, percentages, dates, and times. Saved formula caches are preferred; when a cache is missing, the formula text is returned with a warning. OpenDocs does not recalculate formulas or fetch linked workbooks, data connections, or URLs—the reference is preserved as text only.

Chart titles, labels, series, categories, and accessible values come from native workbook data. When vision is configured, normalized chart fact cards and embedded images may add trend, label, relationship, or meaning interpretation. This enrichment is fail-open and never replaces native facts. XLSX output does not promise Excel pixel appearance, fonts, colors, borders, dimensions, or other visual styling fidelity.

How OpenDocs compares

Feature	OpenDocs	marker	docling	unstructured	pypdf
PDF → Markdown	✅	✅	✅	✅	❌
DOCX → Markdown	✅	❌	✅	✅	N/A
PPTX → Markdown	✅	❌	✅	✅	N/A
XLSX → Markdown	✅	❌	✅	✅	N/A
LLM vision integration	✅	❌	❌	❌	❌
Sync + Async API	✅	❌	❌	❌	❌
No external service required	✅	✅	✅	⚠️	✅
Pure Python (no system deps)	✅	❌	❌	❌	✅
Typed errors & warnings	✅	❌	❌	❌	❌
Image → Markdown	✅	❌	❌	✅	❌
Provider-neutral vision (LiteLLM)	✅	N/A	N/A	❌	N/A

Key differentiator: OpenDocs is the only library that combines native Office/PDF extraction with optional LLM-powered visual understanding, all through a clean sync/async API with typed errors.

Platforms: Ubuntu and macOS on Python 3.11, 3.12, and 3.13 (Poppler required for visual PDF). Windows is unverified for 0.2.0.

Privacy: OpenDocs never downloads HTTP, OSS, or S3 URLs. Model calls send sanitized images to the provider selected by VisionConfig — review that provider's privacy and cost terms before enabling vision.

Warnings and errors

OpenDocs uses Python warnings for recoverable degradation and typed exceptions for fatal failures.

import warnings

from opendocs import OpenDocsError, OpenDocsWarning, ParseOptions, parse

with warnings.catch_warnings(record=True) as caught:
    warnings.simplefilter("always", OpenDocsWarning)
    markdown = parse(
        b"first paragraph\n\nsecond paragraph\n",
        options=ParseOptions(max_output_chars=16),
    )

assert markdown == "first paragraph\n"
assert caught[0].message.code == "output_truncated"

try:
    parse("slides.pdf")
except OpenDocsError as error:
    print(error.code, error.retryable)

Project docs

• Documentation
• Roadmap
• Contributing
• License