mineru-open-sdk 0.1.1

MinerU API Python SDK — one line to turn documents into Markdown

MinerU Open API SDK (Python)

中文文档

MinerU Open API SDK is a completely free Python library for the MinerU document extraction service. Turn any document (PDF, Images, Word, PPT, Excel) or Web Page into high-quality Markdown with just one line of code.

---

🚀 Key Features

• Completely Free: No hidden costs for document extraction.
• Flash Mode (No Auth): Extract text instantly without an API token.
• Full Feature Mode: Comprehensive extraction with layout preservation, images, and formula support.
• Batch & Polling Primitives: Blocking methods for simple flows plus submit/query methods for asynchronous workflows.
• Simple Save Helpers: Save Markdown, HTML, LaTeX, DOCX, or the full extracted zip with built-in helpers.

---

📦 Install

pip install mineru-open-sdk

---

🛠️ Quick Start

1. Flash Extract (Fast, No Auth, Markdown-only)

Ideal for quick previews. No token required.

from mineru import MinerU

# No token needed for Flash Mode
client = MinerU()
result = client.flash_extract("https://cdn-mineru.openxlab.org.cn/demo/example.pdf")

print(result.markdown)

2. Full Feature Extract (Auth Required)

Supports large files, rich assets (images/tables), and multiple formats.

from mineru import MinerU

# Get your free token from https://mineru.net
client = MinerU("your-api-token")
result = client.extract("https://cdn-mineru.openxlab.org.cn/demo/example.pdf")

print(result.markdown)
print(result.images) # Access extracted images

---

🧩 Supported Public API

Client lifecycle

• MinerU(token: str | None = None, base_url: str = ..., flash_base_url: str | None = None)
• client.close()
• client.set_source("your-app")
• context manager support: with MinerU(...) as client:

Blocking extraction methods

• client.extract(...) -> ExtractResult
• client.extract_batch(...) -> Iterator[ExtractResult]
• client.crawl(...) -> ExtractResult
• client.crawl_batch(...) -> Iterator[ExtractResult]
• client.flash_extract(...) -> ExtractResult

Submit/query methods

• client.submit(...) -> str
• client.submit_batch(...) -> str
• client.get_batch(batch_id) -> list[ExtractResult]
• client.get_task(task_id) -> ExtractResult

Result helpers

• result.save_markdown(path, with_images=True)
• result.save_docx(path)
• result.save_html(path)
• result.save_latex(path)
• result.save_all(dir)
• image.save(path)

Result fields you will usually use

• result.state
• result.progress
• result.markdown
• result.images
• result.content_list
• result.docx
• result.html
• result.latex
• result.task_id

---

📊 Mode Comparison

Feature	Flash Extract	Full Feature Extract
Auth	No Auth Required	Auth Required (Token)
Speed	Blazing Fast	Standard
File Limit	Max 10 MB	Max 200 MB
Page Limit	Max 20 Pages	Max 600 Pages
Formats	PDF, Images, Docx, PPTx, Excel	PDF, Images, Doc/x, Ppt/x, Html
Content	Markdown only (Placeholders)	Full assets (Images, Tables, Formulas)
Output	Markdown	MD, Docx, LaTeX, HTML, JSON

---

⚙️ Defaults And Option Behavior

MinerU(...)

Argument	Default	Behavior
token	None	If omitted, the SDK reads MINERU_TOKEN from the environment
base_url	https://mineru.net/api/v4	Standard API base URL
flash_base_url	SDK default flash URL	Override flash API endpoint for testing/private deployments

If neither token nor MINERU_TOKEN is set, the client works in flash-only mode: flash_extract() works, while auth-required methods raise NoAuthClientError.

Full-feature methods

These defaults apply to extract(), extract_batch(), submit(), submit_batch(), and indirectly to crawl() / crawl_batch() unless noted otherwise.

Option	Default	Behavior when omitted
model	None	Auto-infers model: .html/.htm uses "html", everything else uses "vlm"
ocr	not set	OCR is disabled (API default)
formula	not set	Formula recognition is enabled (API default)
table	not set	Table recognition is enabled (API default)
language	not set	Chinese "ch" (API default)
pages	None	Full document is processed
extra_formats	None	Only the default Markdown/JSON payload is returned
file_params	None	Per-file overrides for batch methods. A dict[str, FileParam] keyed by path/URL, where FileParam has fields pages, ocr, data_id
timeout	300 seconds for single-item methods	Max total polling time for extract() / crawl()
timeout	1800 seconds for batch methods	Max total polling time for extract_batch() / crawl_batch()

Flash mode

Option	Default	Behavior when omitted
language	"ch"	Default language is Chinese
page_range	None	Full page range allowed by the flash API
timeout	300 seconds	Max total polling time

crawl() / crawl_batch()

• crawl() is shorthand for extract(url, model="html", ...)
• crawl_batch() is shorthand for extract_batch(urls, model="html", ...)

---

📖 Detailed Usage

Full Feature Extraction Options

result = client.extract(
    "./paper.pdf",
    model="vlm",             # "vlm" | "pipeline" | "html"
    ocr=True,                # Enable OCR for scanned documents
    formula=True,            # Formula recognition
    table=True,              # Table recognition
    language="en",           # "ch" | "en" | etc.
    pages="1-20",            # Page range
    extra_formats=["docx"],  # Export as docx, html, or latex
    timeout=600,
)

result.save_all("./output/") # Save markdown and all assets

Context Manager

from mineru import MinerU

with MinerU("your-api-token") as client:
    result = client.extract("./paper.pdf")
    print(result.markdown)

Batch Processing

# Yields results as they complete
for result in client.extract_batch(["a.pdf", "b.pdf", "c.pdf"]):
    print(f"{result.filename}: Done")

Batch With Per-File Pages

from mineru import FileParam

batch_id = client.submit_batch(
    ["a.pdf", "b.pdf"],
    file_params={
        "a.pdf": FileParam(pages="1-5"),
        "b.pdf": FileParam(pages="10-20"),
    },
)

Web Crawling

result = client.crawl("https://www.baidu.com")
print(result.markdown)

---

🔄 submit() / get_batch() Semantics

This is the part most people get wrong at first:

• submit() returns a batch ID
• submit_batch() also returns a batch ID
• the common async flow is therefore submit(...) -> get_batch(batch_id)
• recommends staying on the batch-based flow for async polling

Recommended async flow

batch_id = client.submit("large-report.pdf")

# poll the batch until the first item is done
while True:
    results = client.get_batch(batch_id)
    result = results[0]
    if result.state in ("done", "failed"):
        break

if result.state == "done":
    do_something(result.markdown)

---

🤖 Integration for AI Agents

The SDK is designed to be easily integrated into LLM workflows. For status updates, you can check result.state and result.progress.

batch_id = client.submit("large-report.pdf")
# ... later ...
result = client.get_batch(batch_id)[0]
if result.state == "done":
    do_something(result.markdown)

---

📄 License

This project is licensed under the Apache-2.0 License.

🔗 Links

• Official Website
• API Documentation