pagewise-pdf-extractor 0.3.0

Page-wise PDF to Markdown extraction with text extraction, OCR, LLM fallback, and progress metadata.

pagewise-pdf-extractor

Page-wise PDF to Markdown extraction with text extraction, OCR, LLM fallback, and progress metadata.

pagewise-pdf-extractor is a Python package and CLI for converting PDFs into deterministic page-level Markdown files. It routes each page through embedded-text extraction, scanned-page OCR, and optional vision-model fallback, then returns structured results for RAG and document-processing pipelines.

What It Does

• Extracts text-native PDF pages with PyMuPDF.
• Extracts scanned/image pages with Marker.
• Rejects corrupt or structurally unreliable embedded text before provider routing.
• Detects and splits two-up scanned pages before OCR.
• Falls back to Ollama vision OCR when configured OCR fails.
• Writes one UTF-8 Markdown file per page.
• Writes atomic progress.json with provider attempts, status, config hash, source hash, and page metadata.
• Exposes a library API for applications and a CLI for operators.
• Keeps local processing as the default; remote services are only used if explicitly configured.

Status

v0.3.0 is the current documented release. The public API is intended for early downstream use by applications that need page-wise PDF extraction, but the project is still pre-1.0.

Install

From PyPI after publication:

python -m pip install pagewise-pdf-extractor

Pinned Git dependency:

pagewise-pdf-extractor @ git+https://github.com/ebmurha/pagewise-pdf-extractor.git@v0.3.0

Local development:

python -m pip install -e D:\Developer\Projects\pagewise-pdf-extractor

Runtime dependencies are declared in pyproject.toml. OCR providers also require local binaries:

• marker_single for Marker OCR
• ollama for Ollama fallback
• pdftoppm for rendering pages passed to Ollama

Check the local environment:

pagewise-pdf-extractor --validate-environment

Quickstart

CLI:

pagewise-pdf-extractor document.pdf --output-root output

Python:

from pathlib import Path

from pagewise_pdf_extractor import ExtractionConfig, process_pdf, validate_environment

config = ExtractionConfig(
    text_provider="pymupdf",
    ocr_provider="marker",
    fallback_provider="ollama",
    fallback_enabled=True,
    ollama_model="deepseek-ocr",
)

report = validate_environment(config)
if report.has_fatal_errors:
    raise RuntimeError(report.summary)

result = process_pdf(
    input_pdf=Path("document.pdf"),
    output_root=Path("output"),
    config=config,
)

Public import contract:

from pagewise_pdf_extractor import (
    ExtractionConfig,
    ExtractionResult,
    process_pdf,
    validate_environment,
)

Output

Default layout:

output/
  <input_sha256>/
    page_0001.md
    page_0002.md
    progress.json

Successful page:

# Page N

<provider markdown content>

Failed page:

# Page N

OCR FAILED

Error: <error_message>

Provider Routing

Default page-level routing:

1. Try embedded text extraction with PyMuPDF.
2. Validate embedded text for corrupt glyphs, abnormal spacing, and table-heavy visual layout.
3. Accept embedded text when it meets configured length and structural quality thresholds.
4. Render Marker input at 350 DPI and split confidently detected two-up pages into logical pages.
5. Use Marker OCR when embedded text is absent, low quality, or force_ocr=True.
6. Use Ollama fallback when Marker fails or returns unusable output and fallback is enabled.
7. Write failure Markdown if all configured providers fail.

Documentation

• API reference
• Configuration reference
• Environment and provider setup
• Integration guide
• Packaging and naming guide
• Releases and versioning
• Changelog
• Release notes
• Contributing
• Security

Tests

python -m unittest discover -s tests -p "test_*.py"
python -c "from pagewise_pdf_extractor import ExtractionConfig, process_pdf, validate_environment; print('ok')"
pagewise-pdf-extractor --help

Future Goals

Future work to preserve the public API, keep provider behavior explicit, and add new providers or extraction quality improvements behind documented configuration.