ericmjl/pdf-form-filler

PDF Form Filler

Fill static PDF forms by detecting PDF type and using the appropriate method:

• Text-based PDFs: Search for label text and insert values at found positions
• Image-based PDFs: Use Claude Vision iteratively for pixel-perfect placement

Quick Start

# Option 1: Text-based PDF (fast, uses label search)
uv run scripts/fill_pdf.py form.pdf filled.pdf "Label:?=value"

# Option 2: VLM-guided (works for ANY PDF, iteratively verifies placement)
uv run scripts/fill_pdf_vlm.py form.pdf filled.pdf --fields '{"Field": "value"}'

---

Method 1: Text Search Positioning (Text-Based PDFs)

This is how pixel-perfect filling works:

Instead of guessing x/y coordinates, we:

1. Search for the label text using PyMuPDF's page.search_for("Label:")
2. Get the bounding rectangle of the found text
3. Insert value after the label at x = rect.x1 + offset (right edge of label + small offset)

┌─────────────────────────────────────┐
│ Student's name: ________________    │
│                ↑                    │
│         rect.x1 (right edge)        │
│              + 5px offset           │
│              = insertion point      │
└─────────────────────────────────────┘

Code pattern:

areas = page.search_for("Student's name:")
if areas:
    rect = areas[0]
    x = rect.x1 + 5  # right edge + small offset
    y = rect.y0      # top of the text
    page.insert_text((x, y + fontsize * 0.8), "Philip Ma", fontsize=11)

Why this works:

• PDFs store text with precise positions
• PyMuPDF returns exact bounding boxes for found text
• Inserting right after the label naturally aligns with the form field

Usage

# Detect PDF type
uv run scripts/fill_pdf.py --detect form.pdf

# Extract text to find labels
uv run scripts/fill_pdf.py --extract form.pdf

# Fill using label?=value format
uv run scripts/fill_pdf.py form.pdf filled.pdf \
    "Student's name:?=John Doe" \
    "Total:?=1500.00"

---

Method 2: VLM-Guided Iterative Filling (Any PDF)

For complex forms or when text search fails:

Uses Claude Vision to:

1. Discover: Identify where each value should be placed
2. Render: Place text and render the page
3. Verify: Ask VLM if placement is correct
4. Iterate: Correct positions up to N times until verified

┌──────────────────────────────────────────┐
│  PDF Page                                │
│  ┌─────────────────────────────────────┐ │
│  │ Print Name                          │ │
│  │ ══════════════════════════════════  │ │
│  │         ↑ VLM sees: "above line!"   │ │
│  │ Date                                │ │
│  │ ══════════════════════════════════  │ │
│  └─────────────────────────────────────┘ │
│                                          │
│  VLM: "Print Name should be ABOVE line"  │
│  → Corrects y-position automatically     │
└──────────────────────────────────────────┘

Key Features

1. Only renders pending fields during verification (not all fields)
2. Uses pixel coordinates for VLM communication, converts to PDF points internally
3. Iterative convergence until all fields are correctly positioned

Usage

# Requires ANTHROPIC_API_KEY
export ANTHROPIC_API_KEY=sk-...

uv run scripts/fill_pdf_vlm.py form.pdf filled.pdf --fields '{
    "Student Name": "Philip Ma",
    "Print Name (first)": "Eric Ma",
    "Date (first)": "March 08, 2026",
    "Print Name (second)": "Nan Li",
    "Date (second)": "March 08, 2026"
}'

How It Works

1. Discovery Phase: For each page, Claude Vision analyzes the form and returns pixel coordinates for each field
2. Iteration Phase:
   • Render PDF with only PENDING (uncorrected) fields
   • VLM checks if text is correctly positioned
   • If wrong, VLM provides corrected pixel coordinates
   • Repeat up to --max-iter times (default: 5)
3. Final Output: Write all verified placements to output PDF

Example Run

--- Page 1: Discovering ---
  'Student Name' at (134, 84)
  'Print Name (p1)' at (306, 763)
  
=== Iteration 1 (11 pending) ===
  Page 1:
    OK 'Student Name'
    FIX 'Print Name (p1)': (306,763) -> (306,749) - Text should be ABOVE signature line
    
=== Iteration 2 (10 pending) ===
  Page 1:
    OK 'Print Name (p1)'
    
All verified after 2 iterations!

When to Use

• Scanned/image PDFs where text extraction fails
• Complex layouts where labels don't clearly indicate position
• Forms with horizontal lines where text goes above/below
• High precision needed and willing to use API calls

Cost Considerations

Each iteration makes ~1 API call per page for verification. Most forms complete in 2-3 iterations.

---

Environment Variables

• ANTHROPIC_API_KEY: Required for VLM-guided mode (fill_pdf_vlm.py)

Limitations

• Signature areas should remain blank for manual signing
• Does not handle checkboxes or radio buttons
• VLM mode uses Claude API tokens per iteration