OCR & Document Parsing

EasyTidy Pro’s OCR is used to recognize text from images, scanned PDFs, and other files without directly copyable text. It is part of the document content extraction pipeline and is automatically invoked by file content filtering, AI summarization, and structured data extraction.

The application does not have a standalone “select files and export OCR text” page, nor does the MinerU settings page provide a connection test button. After configuration, you must verify recognition results through actual content filtering or document processing tasks.

Digital documents may not use OCR

Searchable digital PDFs, Word documents, Excel spreadsheets, and plain text files are preferentially extracted directly. Only files detected as needing OCR, or image/PDF candidates where regular extraction fails, enter the OCR pipeline. Direct extraction is usually faster and better preserves the original text.

OCR Call Order

EasyTidy Pro obtains document text in the following order:

Determine whether OCR is needed
 ├─ Not needed: Use local content extraction results
 └─ Needed, or regular extraction failed:
    1. MinerU cloud parsing (only when enabled and upload consented)
    2. Local FFI OCR (Content Extraction component)
    3. Windows OCR (final fallback)

Tier	Runs On	Primary Use	Key Requirement
Regular Content Extraction	Local	Read body text from digital PDFs, Office documents, spreadsheets, and plain text	Content Extraction component installed
MinerU	Cloud	Complex layouts, scanned PDFs, table and formula parsing	Integration enabled with upload consent; internet required
Local FFI OCR	Local	Primary local OCR for images and scanned documents	Content Extraction component, Tesseract OCR, and corresponding language data installed
Windows OCR	Local	Platform fallback when the above two OCR methods fail	Windows 10/11 with an available OCR language

Enabling MinerU does not disable local capabilities. If MinerU returns an error, empty result, HTTP 429, or does not complete within 10 minutes, the application continues to try local paths.

Supported Scope

Automatic OCR Candidate Files

Automatic detection enables OCR candidate logic for the following extensions:

Category	Extensions
Images	.bmp, .gif, .jpg, .jpeg, .jxr, .png, .tif, .tiff, .webp
Documents	.pdf

Word, Excel, HTML, and plain text typically go through direct content extraction rather than rendering pages as images for OCR. Password-protected, corrupted, or unsupported format documents may not be extractable.

The Windows OCR fallback supports BMP, GIF, JPG/JPEG, JXR, PNG, TIF/TIFF, and PDF; WebP requires successful processing by local FFI OCR or MinerU. Windows OCR renders PDFs page by page and automatically downscales images that exceed the system OCR size limit.

Recognition Languages

OCR language defaults to EasyTidy Pro’s current UI language:

• Simplified Chinese maps to local chi_sim, Windows zh-Hans-CN, and MinerU ch;
• English maps to local eng, Windows en-US, and MinerU en;
• The Windows fallback prioritizes the requested language; if unavailable, it attempts the currently installed OCR language.

The current MinerU page has no separate language selector. For mixed Chinese/English content, Traditional Chinese, or other languages, validate with representative samples first and ensure Windows has the corresponding OCR language support installed.

Preparing Local OCR

Install the Content Extraction Component

1. Open Settings → Optional Components.
2. Find Content Extraction.
3. Click Download & Enable.
4. Wait for the page to show “Download successful” or “Installed.”
5. Fully exit and restart EasyTidy Pro so that easytidy_extract.dll is loaded during startup.

The Content Extraction component is installed into the libs folder of the current application directory. Standard installations, portable copies, and different development directories are independent of each other; re-check component status after switching instances.

Install Tesseract OCR

The “Content Extraction” component provides easytidy_extract.dll and libtika_native.dll for EasyTidy Pro’s document parsing capabilities, but the current component does not include tesseract.exe, tessdata, or language data files. Local FFI OCR recognizes images and scanned documents through Apache Tika calling the system’s Tesseract, so you need to separately install 64-bit Tesseract OCR.

Installation notes:

1. Obtain a 64-bit Tesseract installer for Windows from a trusted source.
2. Install Tesseract and add its installation directory (typically C:\Program Files\Tesseract-OCR) to the system PATH.
3. Install the language data corresponding to the UI language: Simplified Chinese needs chi_sim.traineddata, English needs eng.traineddata.
4. Confirm the language files are in the tessdata directory that Tesseract actually uses. Other languages require the corresponding .traineddata files, but EasyTidy Pro’s current local OCR only requests chi_sim or eng based on the UI language.
5. Fully exit and restart EasyTidy Pro so the application process picks up the updated PATH.

EasyTidy Pro currently has no UI to separately configure the Tesseract executable path, so merely installing the program without adding it to PATH may still prevent local FFI from finding it. Run the following in a freshly opened Command Prompt or PowerShell:

tesseract --version
tesseract --list-langs

The first command should output version information; the second should list at least the needed chi_sim or eng. If the terminal also reports “command not found,” fix the installation directory and PATH first, then test EasyTidy Pro.

Tesseract only affects local FFI OCR

Without Tesseract installed, regular digital documents can still be directly parsed by the Content Extraction component; the application may also use enabled MinerU or Windows OCR for recognition. If you want to reliably process images and scanned PDFs without uploading files, you should have the Content Extraction component, Tesseract, and the corresponding language data all in place.

Windows OCR is not a full component replacement

Even without the Content Extraction component, Windows 10/11 OCR may still succeed as a final fallback on some images and PDFs. However, it does not handle full Office, spreadsheet, and complex document parsing; for stable use of content filtering and AI document operations, you should still install the Content Extraction component.

Check Windows OCR Language

Windows OCR uses the system’s installed language capabilities. If the logs indicate no available OCR language, install the corresponding language and its optional features in Windows Language Settings, then restart EasyTidy Pro. Corporate devices may restrict language pack installation via administrator policy — follow your organization’s policy.

Configuring MinerU Cloud OCR

MinerU is an optional preferred cloud parsing service, suitable for scanned PDFs, complex layouts, tables, and formulas. It is not a prerequisite for using the local content extraction component; if you do not want to upload files, keep MinerU off and use only local FFI and Windows OCR.

1. Open Settings

Go to Integration Settings → MinerU Integration. The page shows the current mode, enable toggle, API Token, parsing model, service limits, and API documentation link.

2. Confirm Upload Privacy

When toggling on Enable MinerU, the application shows a cloud upload confirmation. Files requiring OCR may be uploaded in full; content may include:

• Original document files;
• Text and images;
• Tables and formulas;
• File names and document metadata.

Click Understand & Enable to allow cloud parsing; Cancel returns MinerU to off. Checking “I understand files will be uploaded — don’t show again” suppresses the confirmation on subsequent enables.

Local sanitization does not prevent MinerU from uploading the original file

The “sanitize before sending to AI” option in Structured Extraction applies to the extracted text sent to the LLM afterward — it does not modify the original document before MinerU OCR. As long as MinerU is enabled, complete files requiring OCR may be uploaded to MinerU first. Keep MinerU off for confidential, regulated, or device-bound files.

3. Choose Agent or Precision API

The Token field is optional; whether a Token is configured determines the actual API used:

Mode	Token Required	Per-File Limit	Page Limit	Model Selection
Agent Lightweight Parse API	No	10 MB	20 pages	Model option has no effect
Precision Parse API	Yes	200 MB	200 pages	Can select vlm, pipeline, or MinerU-HTML

The Precision API also has submission limits: up to 50 files per minute, up to 5000 files per day; result query limit is 1000 per minute. EasyTidy Pro submits at most 4 concurrent MinerU recognition requests and polls for results every 3 seconds.

After entering a Token, click Save — the Token is stored encrypted. When re-entering the page, an empty input field means the previously saved value is retained; clicking Clear Token removes the saved value and switches subsequent requests to the no-Token Agent mode.

4. Select a Parsing Model

Model selection only takes effect when using the Precision API with a Token:

• vlm: Default and recommended option, suitable for general images, scanned documents, and complex layouts.
• pipeline: Leans toward traditional parsing pipelines and compatibility — good for comparison when VLM results are suboptimal.
• MinerU-HTML: Only for HTML. When the actual input is HTML, this model is used automatically; selecting it for non-HTML input automatically falls back to vlm.

Regular HTML can usually be text-extracted directly — selecting MinerU-HTML does not force cloud OCR. The model switching rules only take effect when MinerU is actually invoked for parsing.

Click Save when done. The page has no separate “Test Connection” button — run an actual task with a small, non-sensitive scanned PDF or image to verify.

Using OCR in Tasks

OCR does not need to be individually enabled per task. As long as a task needs to read document content, EasyTidy Pro automatically decides whether to engage OCR based on the file situation.

Filtering by File Content

Add a File Content condition in a regular task or advanced workflow to filter files by contained text, any keyword, all keywords, exact match, or regular expressions.

1. Create or edit a task.
2. Open Filter Conditions.
3. Select File Content.
4. Fill in keywords and choose a matching mode, optionally set case sensitivity or keyword order.
5. Preview the matched results with a small sample set.

Scanned images and scanned PDFs automatically enter OCR during content extraction. Some advanced condition evaluations skip files larger than 10 MB; for many large files, narrow the test scope first or pre-filter with more explicit file name, size, and type conditions.

AI Summarization

AI Summarization first extracts or OCRs file content, then sends the processed text to the default LLM, and finally generates a summary PDF in the destination directory.

Prerequisites:

• Content Extraction component installed;
• A default LLM configured;
• Decide whether to enable MinerU based on data sensitivity;
• A valid destination directory set.

OCR and LLM are two different data processing stages. When MinerU is enabled, original files may be sent to MinerU; when using an online LLM, the extracted text is also sent to the chosen model service. For full AI configuration, see AI Services.

Extracting Structured Data

Extract Structured Data extracts specified fields from OCR/parse results and outputs JSON, CSV, or Excel. It is suitable for documents with fixed fields such as invoices, orders, forms, and contracts.

1. Create a new task and set the operation to Extract Structured Data.
2. Fill in fields, e.g. “Customer Name, Order Number, Amount, Signing Date.”
3. Choose output format: JSON, CSV, or Excel.
4. Select an output directory; leaving it empty outputs to the source file’s directory.
5. Optionally enable text sanitization before sending to the LLM.
6. First check field completeness and recognition errors with a small set of representative files.

Structured data extraction does not rely solely on OCR: OCR produces the text, and the default LLM then generates the structured result according to the field requirements. Without a default LLM, empty OCR/extraction content, or incorrect model return format, valid output may not be generated.

Verifying Recognition Results

The MinerU settings page has no test button; the recommended way to verify the complete pipeline:

1. Prepare a test image with clear text and a 1–2 page scanned PDF.
2. Confirm the Content Extraction component is installed and tesseract --list-langs lists the required language, then restart the application.
3. First keep MinerU off and test local FFI OCR using a “File Content” condition.
4. If you need to distinguish FFI from Windows fallback, check the Running Logs for the actual recognition path used.
5. Then enable MinerU and repeat the test with the same files, comparing results.
6. Check the target output, condition match results, and Running Logs.
7. After testing, turn off MinerU if cloud parsing is no longer needed.

Recommended sample coverage:

• Simplified Chinese, English, and mixed Chinese/English;
• Normal orientation and rotated images;
• Low resolution, skewed, shadowed, or glare-affected scans;
• Tables, stamps, formulas, and multi-column layouts;
• Searchable PDFs vs. pure scanned PDFs.

File Size, Timeouts & Fallback

Path	Current Limit or Behavior
OCR total entry	Max 200 MB
MinerU Agent	Max 10 MB, 20 pages
MinerU Precision API	Max 200 MB, 200 pages
Local FFI / Windows OCR	After MinerU fails, only processes input up to 50 MB
MinerU single wait	Max 10 minutes; on timeout, tries local OCR
MinerU concurrency	Max 4 requests

Server-side page limits, rate limits, and account quotas are managed by MinerU and may change with service policy; the limits displayed in the application UI should be treated as the definitive reference at configuration time. Files below the size cap may still fail due to page count, corruption, encryption, network interruption, or service rate limiting.

Privacy & Security

• With MinerU off, local FFI and Windows OCR process files on the current computer.
• With MinerU enabled, original files requiring OCR may be uploaded to mineru.net and its returned upload/result addresses.
• Both with and without a Token constitute cloud processing; the difference is only Precision API vs. Agent API.
• AI Summarization and Structured Extraction may also send the OCR text to a separate LLM service provider.
• Tokens, file paths, recognized text, and error responses may all contain sensitive information — sanitize before submitting logs or screenshots.
• OCR results may misrecognize amounts, dates, ID numbers, and names — do not rely on them as the sole basis for payments, compliance, or file deletion.

When handling sensitive files, prefer keeping MinerU off and using local paths; if local recognition quality is insufficient, first confirm with your organization’s policy whether upload is permitted before switching to a cloud service.

FAQ

No cloud requests after enabling MinerU

Only image/PDF candidates that need OCR call MinerU. Searchable PDFs and regular Office documents that have already been successfully body-text-extracted are not uploaded just for the sake of using MinerU. Also confirm that you clicked Save and consented to the upload privacy notice.

Can I use MinerU without entering a Token?

Yes. Leaving the Token field empty uses the Agent Lightweight Parse API with a 10 MB, 20-page per-file limit. Entering a Token uses the Precision API with model selection.

The Token is saved but the input field is empty

This is by design, to avoid displaying secrets in the UI. When the placeholder shows “Token saved,” leaving the field empty and saving retains the original value; click Clear Token if you need to remove it.

Does MinerU failure stop the entire task?

Usually not immediately. The application first logs the MinerU error, then tries local FFI OCR; if that still fails, it tries Windows OCR. However, the task ultimately fails when the file exceeds the local 50 MB cap, the format is unsupported by Windows OCR, no system OCR language is available, or all results are empty.

No text recognized from images or scanned PDFs

First check the Content Extraction component, then run tesseract --version and tesseract --list-langs in a new terminal. Confirm Tesseract is in PATH and that chi_sim or eng corresponding to the UI language is installed; fully restart EasyTidy Pro after modifying environment variables or language data. Also check Windows OCR language, file size, and Running Logs. Improving scan resolution, correcting orientation, cropping irrelevant backgrounds, and enhancing text-background contrast typically helps recognition. Handwriting, artistic fonts, severe glare, and complex tables may still be inaccurate.

Why didn’t my digital PDF use OCR?

This is expected behavior. When the application detects that a PDF already contains extractable text, it reads the text directly, avoiding unnecessary image recognition and cloud uploads.

Logs show HTTP 429 or rate-limited

This means the MinerU service rejected the current request rate. Reduce the number of files processed concurrently, wait for the quota window to refresh, or turn off MinerU and use local OCR. EasyTidy Pro attempts local fallback after receiving a 429.

AI Summarization or Structured Extraction produced no output

First identify the failure stage:

1. Check the logs to confirm whether OCR/content extraction produced body text;
2. Check whether a default LLM is configured;
3. Check whether the destination directory is writable;
4. Check the online LLM’s key, quota, model identifier, and return format;
5. Retry with clearer samples and more explicit extraction fields.

Related features: Optional Components, Filter Conditions, Organization Tasks, AI Services, and Running Logs.