Mistral OCR

A simple CLI to extract text from documents using the Mistral OCR API.

Installation

pip install mistral-ocr-tool

To render HTML files or pages before OCR, install the optional HTML support:

pip install "mistral-ocr-tool[html]"
python -m playwright install chromium

Or install from source:

git clone https://github.com/aburkard/mistral-ocr.git
cd mistral-ocr
pip install .

Configuration

Set your Mistral API key as an environment variable or in a .env file:

MISTRAL_API_KEY="your-api-key"

Usage

mistral-ocr <document_source> [options]

The document source can be a URL, a bare domain URL, a local file path, or - to read from stdin. Bare domain URLs such as example.com/report are treated as https://example.com/report.

Examples

# Process a PDF from a URL
mistral-ocr https://example.com/document.pdf

# Process a local file
mistral-ocr ./invoice.pdf

# Render a local HTML file to PDF before OCR
mistral-ocr ./report.html

# Render an HTML page URL to PDF before OCR
mistral-ocr https://example.com/report

# URLs can omit https://
mistral-ocr example.com/report

# Pipe from stdin
cat document.pdf | mistral-ocr -

# Process specific pages or inclusive ranges (0-indexed)
mistral-ocr large-doc.pdf --pages 0,2-5

# Output as JSON (great for piping to jq)
mistral-ocr document.pdf --json | jq '.pages[0].markdown'

# Extract tables as HTML
mistral-ocr document.pdf --table-format html

# Include headers and footers
mistral-ocr document.pdf --extract-headers --extract-footers

# Preserve OCR 4 reading-order blocks, bounding boxes, and word confidence
mistral-ocr document.pdf -o output/ --include-blocks --confidence-scores word

# Save markdown and images to a directory
mistral-ocr document.pdf -o output/

# Include base64 images in JSON output (for programmatic use)
mistral-ocr document.pdf --json --include-images

# Check page count and estimated cost before processing
mistral-ocr large-doc.pdf --dry-run

Options

Image extraction options require image output mode (-o/--output-dir or --json --include-images), except --image-limit 0, which is allowed in text-only mode to explicitly disable image extraction.

Output-directory mode writes both <document>.md and <document>.json. The JSON preserves blocks, confidence scores, tables, headers, footers, and other API metadata; extracted image base64 is replaced with an image_path to avoid duplicating image data.

HTML inputs are rendered to a temporary PDF with Chromium via Playwright before being sent to Mistral OCR. In auto mode, local .html/.htm files and URLs with Content-Type: text/html are rendered. Other URLs are downloaded locally and uploaded to Mistral, avoiding failures when Mistral cannot fetch the source URL directly.

Development

uv sync --group dev
uv run pytest tests/