Skip to main content

CLI

The dongler command is the quickest way to inspect files and extract Markdown, LaTeX, or JSON from supported formats.

dongler --version
dongler inspect <path>
dongler extract <path> --format markdown
dongler extract <path> --format latex
dongler extract <path> --format json

Inspect​

dongler inspect invoice.pdf

Example output:

path: invoice.pdf
format: pdf
extraction_status: supported

Extract​

PDFs are the primary workflow:

dongler extract invoice.pdf --format json
dongler extract invoice.pdf --format markdown
dongler extract invoice.pdf --format latex

Supported text, Office, web/email, annotation, and image formats work through the same command:

dongler extract notes.txt --format markdown
dongler extract notes.txt --format latex
dongler extract deck.pptx --format markdown
dongler extract notes.odt --format markdown
dongler extract annotations.json --format markdown
dongler extract boxes.csv --format json

For shell pipelines, prefer JSON when you need metadata or warnings:

dongler extract report.pdf --format json

Convert (hybrid pipeline)​

convert runs the staged pipeline — page triage, reading order, and IR v2 provenance — over the same deterministic, model-free engine as extract:

dongler convert report.pdf --format markdown
dongler convert report.pdf --format json

--ml table-structure recognition (experimental)​

Builds compiled with the ml feature gain an opt-in stage that recognizes table structure with a local ONNX model (SLANet, MIT) and snaps each cell's text from the PDF's own text layer — the model decides the grid, the text layer decides the content, so values are never hallucinated:

# build once with the ml feature (pulls ONNX Runtime + a pdfium binary)
cargo install dongler --features ml
dongler convert report.pdf --ml --format markdown

The model weights download once to ~/.cache/dongler/models (override with DONGLER_CACHE_DIR; DONGLER_OFFLINE=1 requires a pre-fetched cache). If the stage fails for any region it falls back to the deterministic table and records a warning — output is never blocked.

Preview

--ml table structure is experimental. It is plumbing-complete and hallucination-free by construction, but its accuracy is currently bounded by the geometric table-region detector that feeds it; a layout-detection model for clean table regions is the next milestone. The default convert/extract path is unchanged and remains the recommended path today.