pdf-text-extraction-fallback-85d5ca

PDF Text Extraction Fallback

Use this skill when read_file returns binary data or garbled content for PDF files instead of readable text. This workflow provides a reliable fallback using command-line PDF tools.

When to Use

• read_file with filetype: pdf returns binary data, unreadable characters, or errors
• You need to extract text from a PDF to process its contents
• Standard file reading methods fail to extract usable text

Step-by-Step Instructions

Step 1: Detect Binary/Unreadable PDF Output

After attempting to read a PDF with read_file, check if the output is:

• Binary data (contains null bytes, non-printable characters)
• Garbled text with many special characters
• Empty or truncated content

# Example of problematic output from read_file
%PDF-1.4
1 0 obj
<< /Type /Catalog ...

If the output looks like raw PDF structure or binary, proceed to Step 2.

Step 2: Use shell_agent with PDF Tools

Invoke shell_agent to extract text using pdftotext (preferred) or pdfplumber (Python fallback):

Task: Extract all text content from <filename.pdf> using pdftotext or pdfplumber.
Output the extracted text in readable format. If pdftotext is not available, use Python with pdfplumber library.

Example shell_agent invocation:

shell_agent task="Extract text from Move_Out_Inspection_Tracker.pdf using pdftotext. Save output to a .txt file and return the content."

Step 3: Validate Extracted Content

After extraction, validate that the content contains expected text patterns:

# Validation checklist
def validate_pdf_extraction(text, expected_patterns=None):
    checks = [
        bool(text.strip()),  # Not empty
        len(text) > 50,  # Has substantial content
        not text.startswith('%PDF'),  # Not raw PDF structure
    ]
    
    if expected_patterns:
        for pattern in expected_patterns:
            checks.append(pattern.lower() in text.lower())
    
    return all(checks)

Common expected patterns to check:

• Document-specific keywords (e.g., "inspection", "resident", "date")
• Expected data formats (dates, names, IDs)
• Minimum word count threshold