What is pdf-craft?

pdf-craft is a PDF-centered conversion library. It turns PDFs into Markdown or EPUB, and can translate the converted content or write a translated result back to PDF. It is especially useful for scanned documents: pages that are otherwise only readable as images become searchable, editable Markdown or EPUB.

The pipeline is designed for books and academic or technical documents, including body text, tables of contents, footnotes, tables, formulas, and images. OCR can run entirely on a compatible local GPU, or use a vendor service that supplies remote compute. Translation uses a separate text LLM.

Online Version

Want to try the workflow before installing anything? Open Inkora - PDF Craft, the online version of the same core experience. Upload a PDF in your browser and see the main workflow in action.

Installation

If you are getting started, use the standard installation:

pip install pdf-craft

This includes vendor OCR, Markdown/EPUB rendering, and PDF translation. Vendor OCR uses remote compute, so your machine does not need CUDA; you provide the service URL, model name, and API key in the OCR configuration.

Only install the local extra when you explicitly want to run OCR models on your own NVIDIA GPU. If you are unsure, use the standard installation above:

pip install "pdf-craft[local]"

Local OCR also requires a CUDA-compatible PyTorch build, model storage, and enough GPU memory. Before processing PDFs, install Poppler; see the Installation Guide for the supported Python versions and complete system setup. If something goes wrong, start with the Troubleshooting Guide.

Quick Start

The following example converts a scanned PDF into a Markdown file. Replace the example OCR endpoint, model name, and API key with your own service configuration.

from pdf_craft import DeepSeekOCRVendorConfig, PDFCraft, PDFOptions

craft = PDFCraft(pdf=PDFOptions(ocr=DeepSeekOCRVendorConfig(
    base_url="https://example.com/v1",
    api_key="your-api-key",
    model="deepseek-ocr",
)))
craft.convert_pdf_to_markdown(
    "input.pdf", "output.md",
)

The conversion uses a temporary working directory automatically and removes it when the conversion finishes or fails. Pass package_path only when you want to keep the intermediate work for debugging or reuse.

For the complete PDF conversion workflow and customization options, see the PDF Translation Guide and API Reference.

Advanced Features

PDF → EPUB

To produce an EPUB instead of Markdown, call convert_pdf_to_epub. This complete example also shows how to set the book title and author metadata:

from pdf_craft import BookMeta, DeepSeekOCRVendorConfig, PDFCraft, PDFOptions

ocr_config = DeepSeekOCRVendorConfig(
    base_url="https://example.com/v1",
    api_key="your-api-key",
    model="deepseek-ocr",
)
craft = PDFCraft(pdf=PDFOptions(ocr=ocr_config))
craft.convert_pdf_to_epub(
    "input.pdf", "output.epub",
    book_meta=BookMeta(title="Book title", authors=["Author"]),
)

If book_meta is omitted, pdf-craft tries to read the metadata from the source PDF.

PDF → translated Markdown or EPUB

To translate while converting, pass a TranslationStep to either conversion method. The translator is a chapter transformer: it sends chapter text to your text LLM and returns the translated chapter. The same step can be used for Markdown and EPUB output.

from pdf_craft import TranslationStep

translation = TranslationStep(translator)
craft.convert_pdf_to_markdown("input.pdf", "translated.md", steps=[translation])
craft.convert_pdf_to_epub("input.pdf", "translated.epub", steps=[translation])

PDF → translated PDF

Use the PDF translation workflow when you want to keep the original PDF layout. It extracts the page content, translates it, and writes the result back into the matching source pages. OCR and translation use separate configurations.

from pdf_craft import DeepSeekOCRVendorConfig, PDFCraft, PDFOptions

craft = PDFCraft(pdf=PDFOptions(ocr=DeepSeekOCRVendorConfig(
    base_url="https://example.com/v1",
    api_key="your-ocr-api-key",
    model="deepseek-ocr",
)))

# Placeholder only: replace this with your text LLM call.
def translator(text: str) -> str:
    return text

package = craft.extract_pdf("input.pdf", "work/cache")
craft.translate_pdf("input.pdf", package, "translated.pdf", translator)

EPUB → translated EPUB

If you already have an EPUB, translate it directly by providing the target language and a text LLM:

from pdf_craft import LLM, PDFCraft, SubmitKind

llm = LLM(
    key="your-api-key",
    url="https://api.openai.com/v1",
    model="gpt-4.1-mini",
    token_encoding="o200k_base",
)

PDFCraft().translate_epub(
    "input.epub", "translated.epub",
    target_language="zh", submit=SubmitKind.REPLACE, llm=llm,
)

REPLACE creates a target-language-only edition. Use APPEND_BLOCK to keep the original and append the translation as a separate block, or APPEND_TEXT to place the translation directly after the original text. See the EPUB translation guide for prompts, retries, concurrency, caching, progress callbacks, and failure handling.

OCR Backends and Model Cache

OCR turns page images into text. pdf-craft offers six backends; choose the runtime location first, then choose the model family:

• No CUDA or minimal local setup: choose vendor OCR. Pages are sent to a remote service and processed with its compute resources, so you need network access, a service URL, and credentials.
• A compatible NVIDIA GPU and local execution: choose local OCR. Models are cached locally and run on your GPU, which keeps processing on your machine but requires CUDA, VRAM, and model files.

DeepSeek OCR and DeepSeek OCR 2 are from DeepSeek; Unlimited OCR is from Baidu. Each model family has local and vendor configurations:

If you simply want to get the workflow running, start with the vendor you already have credentials for. Choose local OCR when you specifically want local execution. The library accepts these configuration objects through PDFOptions(ocr=...) and does not read environment variables. See the OCR Backend Guide for detailed configuration examples.

Unlimited OCR local supports base and gundam. DeepSeek OCR 2 local is verified with base; an explicit tiny selection fails early with a clear message.

Model Cache and Common Parameters

Local OCR models are downloaded from Hugging Face by default. You can pre-download one into a chosen cache directory and then run with local_only=True:

from pdf_craft import DeepSeekOCRLocalConfig, predownload_models

predownload_models(
    ocr=DeepSeekOCRLocalConfig(models_cache_path="models"),
    revision=None,
)

ocr_size supports tiny, small, base, large, and gundam, although presets vary by backend. Markdown defaults to toc_assumed=False; EPUB defaults to toc_assumed=True. Complex chapter hierarchies can use an optional toc_llm.

Related Projects

• Wiki Graph: turn a converted EPUB or Markdown book into structured summaries, chapter topology, and a knowledge graph.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Since v1.0.0, pdf-craft has used DeepSeek OCR under the MIT license and removed the previous AGPL-3.0 dependency. The project still receives easydict transitively through the OCR stack under the LGPLv3 license. Thanks to the community for their support and contributions.

Acknowledgments

• DeepSeek OCR
• doc-page-extractor
• pyahocorasick