Sahaf

Local PDF & EPUB to Markdown converter with automatic digital/scanned detection, OCR support, smart splitting, and page-range selection. Converts books to clean, self-contained Markdown files with embedded images using Marker and Surya OCR (90+ languages). No cloud APIs — runs entirely on your hardware.

Features

• PDF & EPUB support — handles both formats natively
• Automatic PDF classification — detects digital, scanned, or mixed PDFs via PyMuPDF
• High-accuracy conversion — works with both Marker 1.x and Marker 2.x, whichever you install
• Built-in OCR — Surya OCR supports 90+ languages (Turkish, English, Arabic, etc.)
• Page/chapter range selection — convert only a specific section of the book (e.g. pages 19-88)
• Smart splitting — split output into N parts, cutting at heading/paragraph boundaries instead of mid-sentence
• Self-contained output — images embedded as base64 directly in Markdown, no separate files
• Split preview — see exactly how parts will be divided before downloading
• Bilingual UI — Turkish / English interface with one-click toggle
• Dark/light theme — lavender-toned design, persistent toggle
• Drag & drop UI — clean single-page web interface

Install

pip install sahaf

Or from source:

git clone https://github.com/arikusi/sahaf.git
cd sahaf
pip install -e .

Models are downloaded automatically on first conversion.

Choosing Marker 1 or Marker 2

Both lines work — Sahaf detects which one is installed and configures it accordingly. The difference only matters for scanned PDFs; digital PDFs and EPUBs convert the same either way.

Marker 1.x — OCR in-process, nothing else to install

pip install sahaf "marker-pdf<2"

OCR runs inside the Python process with torch, on any CUDA GPU including pre-Ampere cards (GTX 16xx, RTX 20xx). The safe choice for scanned books.

Marker 2.x — faster and more accurate, needs an OCR server

pip install sahaf

Marker 2 reads the PDF text layer and only OCRs what needs it (76.0% on olmOCR-bench in balanced mode). But its OCR model is served out of process, so scanned pages need one of:

• Docker plus the NVIDIA Container Toolkit, on an Ampere or newer GPU — the server runs in bfloat16, which needs compute capability 8.0+. On Arch: sudo pacman -S nvidia-container-toolkit && sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker
• llama.cpp (llama-server on your PATH) — works anywhere, but CPU inference is slow on book-length documents

Sahaf classifies each PDF before converting, so under Marker 2 a digital book runs in fast mode and never starts that server.

Quick Start

sahaf

Open http://localhost:8000 in your browser.

Uploads and converted files land in ~/.local/share/sahaf. To keep them elsewhere:

SAHAF_DATA_DIR=./sahaf-data sahaf

How It Works

1. Upload — drag & drop a PDF or EPUB file
2. Classify — PyMuPDF analyzes PDF type; EPUB chapters are counted
3. Select range (optional) — pick specific pages or chapters to convert
4. Convert — Marker processes PDF; ebooklib + markdownify handles EPUB
5. Split (optional) — choose how many parts to split the output into
6. Download — get a single .md or a ZIP with split parts, all images embedded inline

API

Tech Stack

• Backend: FastAPI + Uvicorn
• PDF Classification: PyMuPDF
• PDF Conversion: Marker (marker-pdf 1.x or 2.x) + Surya OCR
• EPUB Conversion: ebooklib + markdownify
• Smart Splitting: Custom algorithm — heading/HR/paragraph boundary detection
• Frontend: Vanilla HTML/CSS/JS + marked.js
• i18n: TR/EN with client-side toggle

Requirements

• Python 3.10 – 3.13 (Marker's dependency stack pins pillow<11, which has no 3.14 wheels)
• 4-6GB RAM (when Marker models are loaded)
• GPU strongly recommended for scanned PDFs — those go through the OCR model; CPU inference is slow on book-length documents
• More than 6GB VRAM for Marker 1 on GPU — its model set fills a 6GB card before inference starts and dies with a CUDA OOM. On a smaller card run TORCH_DEVICE=cpu sahaf
• Under Marker 2, digital PDFs run without the OCR server at all
• EPUB conversion is lightweight — no GPU needed, runs instantly

License

GPL-3.0