How to Specify the Parsing Method (Auto, TXT, OCR) in MinerU

You can specify the parsing method in MinerU using the -m or --method flag in the CLI, the parse_method parameter in the Python API, or the parse_method form field in the FastAPI endpoint to choose between auto, txt, or ocr modes.

MinerU is an open-source document parsing tool developed by OpenDataLab that extracts structured data from PDFs and images. When processing documents, you can control whether the tool uses text extraction, OCR, or automatic detection by specifying the parsing method parameter.

Understanding the Three Parsing Methods

MinerU supports three distinct parsing strategies that determine how content is extracted from PDF documents.

Auto Mode (Default)

The auto method analyzes the PDF at runtime to determine the optimal extraction strategy. If the document contains extractable text layers, MinerU uses the text-extraction path; if the PDF is image-based or lacks text layers, it automatically falls back to OCR. This mode is implemented in mineru/utils/pdf_classify.py, where the classify function inspects PDF bytes to determine if OCR is required.

TXT Mode (Text Extraction)

The txt method forces MinerU to use the text-extraction pipeline exclusively, even for image-only PDFs. This mode extracts selectable text directly from the PDF without invoking OCR models, making it faster for documents that already contain text layers but potentially returning empty results for scanned documents.

OCR Mode (Optical Character Recognition)

The ocr method forces the OCR pipeline regardless of whether the PDF contains extractable text. This is useful when you need to process scanned documents or when the existing text layer in a PDF is corrupted or incomplete. In this mode, MinerU runs layout detection first, then crops text blocks and sends them to the OCR model for recognition.

How to Specify the Parsing Method in MinerU

You can specify the parsing method through three different interfaces depending on your integration needs.

Command Line Interface (CLI)

When using the mineru command, pass the -m or --method flag followed by your chosen method. The argument parsing is defined in mineru/cli/client.py at lines 44-48, where the value is passed to the do_parse function.

# Use auto-detection (default)

mineru -p document.pdf -o ./output

# Force text extraction

mineru -p document.pdf -o ./output -m txt

# Force OCR for scanned documents

mineru -p scanned.pdf -o ./output -m ocr

FastAPI HTTP Endpoint

When using the FastAPI server, include the parse_method form field in your POST request to the /file_parse endpoint. This parameter is declared in mineru/cli/fast_api.py at lines 63-70 and forwarded to the aio_do_parse function.

curl -X POST http://localhost:8000/file_parse \
  -F "files=@/path/to/document.pdf" \
  -F "output_dir=./output" \
  -F "parse_method=ocr" \
  -F "backend=hybrid-auto-engine"

Python API

For programmatic usage, pass the parse_method argument directly to the do_parse or aio_do_parse functions imported from mineru.cli.common. This value is then forwarded to the backend analyzers.

from mineru.cli.common import do_parse
from mineru.utils.cli_parser import arg_parse

# Force text extraction mode

do_parse(
    output_dir="./output",
    pdf_file_names=["document"],
    pdf_bytes_list=[pdf_bytes],
    p_lang_list=["ch"],
    backend="pipeline",
    parse_method="txt",  # Options: "auto", "txt", "ocr"

    formula_enable=True,
    table_enable=True,
    server_url=None,
    start_page_id=0,
    end_page_id=None,
    **arg_parse(None)
)

How the Parsing Method Works Internally

When you specify a parsing method, the value flows through several layers of the MinerU architecture before determining the actual processing path.

The parse_method parameter first reaches the backend analyzer—either mineru/backend/pipeline/pipeline_analyze.py or mineru/backend/hybrid/hybrid_analyze.py. Both modules contain logic to determine whether OCR should be enabled based on your specification and the document content.

In mineru/backend/hybrid/hybrid_analyze.py (lines 33-41), the ocr_classify helper function evaluates the method:

• If parse_method == "auto", it calls the classifier in mineru/utils/pdf_classify.py to inspect the PDF bytes and determine if the document is image-only.
• If parse_method == "txt", it forces _ocr_enable to False, bypassing OCR regardless of content.
• If parse_method == "ocr", it forces _ocr_enable to True, enabling OCR for all pages.

This boolean flag then determines whether the pipeline extracts text directly from PDF text layers or crops text blocks for OCR processing. Both paths generate a middle-JSON structure that is later converted to your desired output format (Markdown, JSON, etc.).

Summary

• Three methods available: auto (default), txt (force text extraction), and ocr (force OCR).
• CLI usage: Use mineru -m txt or mineru --method ocr when running the command.
• API usage: Pass parse_method="ocr" to do_parse() or include it as a form field in FastAPI requests.
• Internal logic: The ocr_classify function in mineru/backend/hybrid/hybrid_analyze.py evaluates your choice against the PDF content to set the _ocr_enable flag.
• Performance implications: txt is fastest for text-based PDFs, ocr is necessary for scanned documents, and auto provides the best balance by detecting document types at runtime.

Frequently Asked Questions

What is the default parsing method in MinerU?

The default parsing method is auto, which automatically detects whether a PDF contains extractable text layers or requires OCR. When set to auto, MinerU uses the classify function in mineru/utils/pdf_classify.py to inspect the document and choose the appropriate processing path at runtime.

When should I use the TXT parsing method instead of AUTO?

Use the txt method when you know your PDFs contain clean, selectable text layers and you want to maximize processing speed by skipping the automatic detection step. This forces MinerU to use the text-extraction pipeline exclusively, bypassing OCR even if the classifier would normally recommend it. However, if the PDF is actually image-based, this method will return empty results.

How does MinerU decide between text extraction and OCR in AUTO mode?

In auto mode, MinerU calls the ocr_classify helper function (located in mineru/backend/hybrid/hybrid_analyze.py), which invokes the classify function from mineru/utils/pdf_classify.py. This classifier analyzes the PDF bytes to determine if the document is image-only. If the classifier returns "ocr", the system sets _ocr_enable to True and processes the document with OCR; otherwise, it extracts text directly from the PDF layers.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}