PDF Text Extraction

Extract high-quality text from PDFs using two OCR engines:

• asta pdf-extraction olmocr — cloud-based extraction via olmOCR (best for large batches, S3, and complex layouts)
• asta pdf-extraction remote — quick single-file extraction via the Asta remote OCR API

Installation

This skill requires the asta CLI:

# Install/reinstall at the correct version
PLUGIN_VERSION=0.15.0
if [ "$(asta --version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+')" != "$PLUGIN_VERSION" ]; then
  uv tool install --force git+https://github.com/allenai/asta-plugins.git@v$PLUGIN_VERSION
fi

Prerequisites: Python 3.11+ and uv package manager

Quick Start

olmocr (cloud batch extraction)

# Extract text from a PDF using a temporary workspace
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs document.pdf \
  --markdown

# Output will be in $TEMP_WORKSPACE/markdown/document.md

# Batch extract text from all PDFs in a directory
TEMP_WORKSPACE=$(mktemp -d)
find /path/to/pdfs -name "*.pdf" > "$TEMP_WORKSPACE/pdf-list.txt"
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs "$TEMP_WORKSPACE/pdf-list.txt" \
  --markdown

# Batch extract text from PDFs stored in S3
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs s3://my-bucket/prefix/*.pdf \
  --markdown

Key arguments:

• <workspace> - Output directory (required: positional argument)
• --pdfs - PDF file(s) to process (required: single path, path with wildcard, S3 path with wildcard, or a text file with paths)
• --markdown - Generate markdown output (recommended)
• --workers - Parallel workers (default: 20)

remote (single-file extraction)

# Print extracted markdown to stdout
asta pdf-extraction remote paper.pdf

# Save to a file
asta pdf-extraction remote paper.pdf -o paper.md

# Process a large PDF in steps (pages 0-49, then 50-99, etc.)
asta pdf-extraction remote paper.pdf --start-page 0  --max-pages 50 -o paper-part1.md
asta pdf-extraction remote paper.pdf --start-page 50 --max-pages 50 -o paper-part2.md

Key arguments:

• <pdf> - PDF file to process (required: local path)
• -o / --output - Output file path (default: stdout)
• --start-page - First page to process, 0-indexed (default: 0)
• --max-pages - Maximum number of pages to process (default: 50)
• --images - Extract and save embedded images alongside the markdown; images are saved in the same directory as the output file and referenced by filename in the markdown

Requirements: Asta login (same as olmocr).

Workspace Best Practices

Use Temporary Workspace, Move Final Output

Recommended workflow:

# 1. Create temporary workspace (not in .asta/documents)
TEMP_WORKSPACE=$(mktemp -d)

# 2. Extract text
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs research-paper.pdf \
  --markdown

# 3. Move final output to permanent location
mkdir -p ~/.asta/documents/research
mv "$TEMP_WORKSPACE/markdown/research-paper.md" ~/.asta/documents/research/

# 4. Clean up temporary files
rm -rf "$TEMP_WORKSPACE"

Why use a temporary workspace?

• olmOCR creates several intermediate files (JSON, work queues, etc.)
• You typically only need the final markdown output
• Keeping .asta/documents clean with only final outputs
• Easy cleanup of temporary files

Where to store final outputs:

• ~/.asta/documents/ - For indexing with asta documents
• User's project directory - For project-specific documents
• Any location convenient for the user's workflow

Common Workflows

Extract Single PDF

# Create temporary workspace and extract
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs paper.pdf \
  --markdown

# Review the extracted text
cat "$TEMP_WORKSPACE/markdown/paper.md"

# Move to permanent location
mkdir -p ~/.asta/documents/papers
mv "$TEMP_WORKSPACE/markdown/paper.md" ~/.asta/documents/papers/

# Clean up
rm -rf "$TEMP_WORKSPACE"

Extract Multiple PDFs

# Process all PDFs in a directory
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs papers/*.pdf \
  --markdown \
  --workers 10

# Move all extracted markdown files
mkdir -p ~/.asta/documents/batch
mv "$TEMP_WORKSPACE/markdown/"*.md ~/.asta/documents/batch/

# Clean up
rm -rf "$TEMP_WORKSPACE"

Extract and Index in Documents

# 1. Extract text to temporary workspace
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs research-paper.pdf \
  --markdown

# 2. Move to documents directory
mkdir -p ~/.asta/documents/research
FINAL_PATH=~/.asta/documents/research/research-paper.md
mv "$TEMP_WORKSPACE/markdown/research-paper.md" "$FINAL_PATH"

# 3. Index in asta documents
asta documents add "file://${FINAL_PATH}" \
  --name="Research Paper (OCR)" \
  --summary="Extracted via olmOCR" \
  --tags="ocr,extracted,research"

# 4. Clean up temporary workspace
rm -rf "$TEMP_WORKSPACE"

S3 Support

olmOCR supports reading PDFs from S3 and using S3 as a workspace.

Read PDFs from S3

# Extract PDF stored in S3
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs s3://my-bucket/documents/paper.pdf \
  --markdown

# Output will be in local workspace
cat "$TEMP_WORKSPACE/markdown/paper.md"

Use S3 as Workspace

# Use S3 bucket as workspace (requires AWS credentials)
asta pdf-extraction olmocr s3://my-bucket/ocr-workspace \
  --pdfs document.pdf \
  --markdown

# Output will be in s3://my-bucket/ocr-workspace/markdown/document.md

S3 Configuration:

• Requires AWS credentials configured (via ~/.aws/credentials or environment variables)
• Uses standard boto3 credential resolution
• Can mix local and S3 paths (e.g., S3 PDFs to local workspace, or vice versa)

Output Structure

After extraction, the workspace directory contains:

workspace/
├── markdown/           # Markdown output (if --markdown used)
│   └── document.md
├── output/            # Raw JSON output
│   └── document.json
└── work_queue/        # Internal work tracking

The extracted text is saved as markdown, preserving:

• Document structure
• Headings and formatting
• Tables (as markdown tables)
• Lists and emphasis

Advanced Options

Control Parallelism

# Process with 50 parallel workers (faster for large batches)
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs papers/*.pdf \
  --workers 50 \
  --markdown

Filter Documents

# Apply filters to skip non-English or form-like documents
TEMP_WORKSPACE=$(mktemp -d)
asta pdf-extraction olmocr "$TEMP_WORKSPACE" \
  --pdfs documents/*.pdf \
  --apply_filter \
  --markdown

Workspace Statistics

# View statistics about a workspace
asta pdf-extraction olmocr ~/workspace/output --stats

Performance and Cost

• Speed: ~10-20 seconds per page (cloud GPU)
• Cost: ~$0.001-0.01 per typical PDF (10-50 pages)
• No setup: Works immediately, API configuration handled automatically

Troubleshooting

"workspace is required"

The first argument must be a workspace directory:

# ✓ Correct
asta pdf-extraction olmocr ~/workspace/output --pdfs file.pdf ...

# ✗ Wrong (missing workspace)
asta pdf-extraction olmocr --pdfs file.pdf ...

"Connection failed"

1. Verify internet connection
2. Check if the service is currently available

"No output files"

Check the output directories:

• Markdown: <workspace>/markdown/
• JSON: <workspace>/output/

S3 Access Issues

1. Verify AWS credentials are configured
2. Check bucket permissions (read for --pdfs, read/write for workspace)
3. Ensure IAM user/role has s3:GetObject, s3:PutObject, s3:ListBucket permissions

Choosing an Engine

	olmocr	remote
Input	Local file, S3 path, or glob	Local file only
Output	Files in workspace directory	Stdout or single file
Batch processing	Yes (--workers)	No
S3 support	Yes	No
Auth required	Asta login	Asta login
Best for	Large batches, complex layouts	Quick single-file extraction

When to Use This Skill

✅ Use PDF extraction when:

• User wants to extract text from a PDF
• User mentions "OCR", "read PDF", "extract from document"
• Processing scanned or image-based PDFs
• Dealing with complex layouts (tables, multi-column, equations)
• Need high-quality text extraction for downstream tasks

❌ Don't use when:

• User wants to process images directly (both engines are PDF-specific)
• User needs real-time/streaming extraction

Additional Resources

• olmOCR GitHub: https://github.com/allenai/olmOCR