Skip to main content

​
Overview

The single document processing feature allows you to quickly generate metadata tags for individual PDF files. Upload a document or provide a URL, configure your settings, and get AI-generated tags in seconds.
Supports both file uploads (up to 50MB) and URL-based processing from any publicly accessible HTTP/HTTPS source.

​
How It Works

1

Upload Your Document

Choose between two input methods:File Upload
  • Drag and drop a PDF file onto the upload zone
  • Or click to browse and select from your computer
  • Maximum file size: 50MB
URL Input
  • Paste a direct link to any publicly accessible PDF
  • Supports CloudFront URLs, S3 URLs, government portals, and more
  • Examples:
    • https://example.com/document.pdf
    • https://d1581jr3fp95xu.cloudfront.net/path/to/file.pdf
    • https://socialjustice.gov.in/writereaddata/UploadFile/66991763713697.pdf
Provide either a file upload or a URL — not both.
2

Preview Your Document

After uploading or entering a URL, an embedded PDF viewer appears automatically:
  • In-page preview: 800px height iframe for quick reference
  • Fullscreen mode: Click the fullscreen button to view the entire document
  • Keyboard shortcut: Press ESC to exit fullscreen
For URL-based documents, the preview uses a CORS proxy endpoint (/api/single/preview) to bypass browser restrictions.
3

Configure Processing Settings

In the configuration panel (typically on the right), set:
  • OpenRouter API Key (required)
  • Model Name (e.g., google/gemini-flash-1.5, openai/gpt-4o-mini)
  • Number of Pages: How many pages to extract (default: 3)
  • Number of Tags: Target tag count (default: 8)
  • Exclusion List (optional): Upload a file to filter out common terms
4

Generate Tags

Click Generate Tags to start processing:
  1. System extracts text from the PDF (PyPDF2 → Tesseract → EasyOCR fallback)
  2. Detected language and extraction method are logged
  3. AI analyzes the content and generates tags
  4. Results appear below with extraction metadata

​
Document Preview Features

​
Standard Preview

The embedded preview displays your PDF at 800px height with:
  • Native browser PDF controls (zoom, download, print)
  • Scroll support for multi-page documents
  • Automatic loading from file or URL

​
Fullscreen Mode

When you click the Fullscreen button:
  • Immersive View: Dark overlay with centered PDF viewer
  • Modal Header: Document icon, title, and close button
  • Keyboard Shortcut: Press ESC to close instantly
  • Full Browser Window: Maximizes PDF viewing area
  • Clean UI: Gradient header with visual polish
The fullscreen modal is implemented with:
// Fixed positioning covers entire viewport
className="fixed inset-0 z-50 bg-black bg-opacity-95"

// ESC key listener
window.addEventListener('keydown', handleKeyDown)

​
Processing Results

​
Success Banner

When processing completes successfully, you’ll see:

Processing Time

Displays how long the extraction and AI tagging took (e.g., “Processed in 4.2s”)

Extraction Method

Shows which engine was used:
  • pypdf2: Text-based PDF (fastest)
  • tesseract_ocr: Scanned document with Tesseract
  • easyocr: Complex language document with EasyOCR
  • pymupdf_tesseract: Fallback for complex embedded images

​
Document Metadata

Document Title
  • Extracted from PDF metadata or first substantive line
  • Displayed prominently at top of results
OCR Status Badge
  • Text PDF: Blue badge for digitally-created documents
  • Scanned PDF: Purple badge with confidence percentage (e.g., “Scanned PDF · 87% confidence”)

​
Generated Tags

Tags are displayed with color-coded categories:
  • Pattern: Years, months, quarters (e.g., 2024, q1, january)
  • Color: Date-specific highlighting
  • Detection: Regex pattern matching on common date formats
Copy All Button
  • Click to copy all tags as comma-separated text
  • Uses navigator.clipboard.writeText(result.tags.join(', '))
  • Perfect for pasting into other systems

​
Extracted Text Preview

A scrollable preview (max 240px height) shows:
  • First few hundred characters of extracted text
  • Monospace font for readability
  • Helps verify extraction quality
  • Useful for debugging OCR issues

​
Smart Document Detection

​
Automatic OCR Triggers

The system automatically uses OCR when:
If PyPDF2 extracts fewer than 150 characters per page, the document is likely scanned.Example:
PyPDF2 extracted 342 chars from 3 pages (avg: 114 chars/page)
OCR needed: true (low density)
Legacy Indian PDFs (Krutidev, ISM fonts) map Devanagari glyphs to Latin Extended Unicode (U+00A0–U+024F).PyPDF2 reads these as garbage like "ºÉÚSÉxÉÉ". The system detects when >25% of characters fall in this range and triggers OCR:
Garbled encoding detected (47% Latin-Extended chars — likely legacy Indian font)
OCR needed: true
If total extracted text is under 500 characters or fewer than 50 words per page:
Sparse content (38 words/page, 12% substantive lines)
OCR needed: true
If PyPDF2 returns empty or whitespace-only content:
OCR needed: true (no text lines found)

​
Hybrid OCR Strategy

When OCR is needed, the system uses a 3-tier approach:
1

Tesseract OCR (Fast Path)

  • Speed: ~2-5 seconds per page at 300 DPI
  • Languages: Hindi, English, Kannada, Tamil, Telugu, Bengali, etc.
  • Confidence Threshold: If confidence ≥60%, Tesseract result is used
  • Preprocessing: Grayscale conversion + contrast enhancement (2.0x)
2

EasyOCR (High Accuracy Fallback)

Triggered when:
  • Tesseract confidence <60%
  • Tesseract extracted <100 chars
  • Complex Indian scripts detected
Features:
  • 80+ languages including all major Indian languages
  • CNN + LSTM neural networks for superior accuracy
  • Subprocess isolation: Runs in separate process to prevent OOM crashes
  • Image resizing: Max 1500px dimension to prevent memory issues
  • Timeout: 120-second limit per document
3

PyMuPDF Fallback (Last Resort)

Used when pdf2image produces blank frames (eOffice PDFs, JBIG2/CCITT compression):
  • Renderer: PyMuPDF (fitz) instead of Poppler
  • Handles: Unusual XObject structures, embedded image overlays
  • OCR Engine: Tesseract on PyMuPDF-rendered images

​
Input Validation

​
Pre-Submission Checks

The Generate Tags button is disabled unless: Error messages appear above the button when validation fails: 
// Example error states
"Please select a PDF file or enter a PDF URL"
"Please provide either a file or a URL, not both"
"URL must start with http:// or https://"
"Please enter your OpenRouter API key in the configuration panel"

​
Reset and Refresh

​
Refresh Button

Click the Refresh button (top-right) to:
  • Clear the current PDF upload
  • Remove URL input
  • Reset results display
  • Clear error messages
  • Preserve configuration settings (API key, model, exclusion list)
The button is disabled when no document is loaded.

​
URL Processing Details

​
Supported URL Types

Direct PDF URLs

https://example.com/document.pdfStandard web-hosted PDFs

CloudFront URLs

https://d1581jr3fp95xu.cloudfront.net/path/to/file.pdfAWS CloudFront distributions

S3 Public URLs

https://bucket.s3.region.amazonaws.com/file.pdfPublicly accessible S3 objects

Government Sites

https://socialjustice.gov.in/writereaddata/UploadFile/66991763713697.pdfInstitutional document repositories

​
Backend Download Process

# Backend implementation details
- Timeout: 60 seconds
- Max file size: 50MB
- User-Agent: Custom header to avoid blocks
- CORS proxy: /api/single/preview endpoint for frontend

​
API Endpoint

POST /api/single/process Form data fields:
  • pdf_file: File object (optional)
  • pdf_url: String (optional)
  • config: JSON string with TaggingConfig
  • exclusion_file: File object (optional)
Response: 
{
  "success": true,
  "document_title": "Training Manual 2024",
  "tags": ["pmkvy", "skill-development", "2024", "training", ...],
  "extracted_text_preview": "First 500 chars...",
  "processing_time": 4.2,
  "is_scanned": false,
  "extraction_method": "pypdf2",
  "ocr_confidence": null,
  "detected_language": "en",
  "language_name": "English"
}

​
Best Practices

For faster processing:
  • Use text-based PDFs when possible (avoid scanned documents)
  • Reduce num_pages to 1-2 for long documents
  • Choose faster models like google/gemini-flash-1.5
For better accuracy:
  • Use URL input for documents already hosted online (avoids upload time)
  • Provide exclusion lists to filter common terms
  • Increase num_pages for more comprehensive analysis
Common issues:
  • CORS errors: URL preview may not work if the source blocks cross-origin requests
  • Timeout: Very large files (>50MB) or slow networks may fail
  • Encoding errors: Some legacy PDFs may require OCR even if they contain text

​
Troubleshooting

Cause: Source URL blocks embedding or CORS restrictionsSolution:
  • The proxy endpoint (/api/single/preview) should handle most cases
  • If preview fails, you can still process the document
  • Processing uses server-side download (not affected by CORS)
Cause: AI response parsing failedSolution:
  • Check the “Raw AI Response” debug section (shown when available)
  • Verify your API key has credits
  • Try a different model (some models format responses differently)
Cause: Poor scan quality, complex layouts, or unsupported scriptsSolution:
  • System automatically tries EasyOCR as fallback
  • Check extracted text preview to verify content quality
  • Consider preprocessing the PDF (increase contrast, remove noise)
Cause: Large file, OCR processing, or slow AI modelSolution:
  • Reduce num_pages in configuration
  • Use faster models (Gemini Flash instead of GPT-4)
  • Ensure good network connection for URL-based documents

Build docs developers (and LLMs) love

Get started for freeTalk to us