Dongler

Turn PDFs and documents into clean Markdown, LaTeX, or structured JSON.
Rust-native, runs locally, no hosted service · API key · LLM · or OCR for digitally born PDFs.

Documentation  ·  Quick start  ·  API reference  ·  LLM context

Dongler is built around a path-first workflow: load a file, inspect the document object when you need to, then render the output format your pipeline wants. One Rust core powers the CLI, Python, TypeScript, and Rust APIs, so the extraction model is identical everywhere.

Install

cargo install dongler                  # CLI + Rust
pip install dongler                    # Python
npm install @cristianexer/dongler      # Node / TypeScript

For the Rust library, depend on dongler-core. The public dongler crate is the CLI package.

To run extraction in the browser or another WebAssembly host, build the dongler-wasm crate (make build-wasm). It exposes the same engine over an in-memory byte API, so files can be parsed client-side with no server. See crates/dongler-wasm/README.md.

Parse a PDF

import dongler

doc = dongler.load("report.pdf")
markdown = doc.to_markdown()
latex = doc.to_latex()
data = doc.to_dict()

import { load } from "@cristianexer/dongler";

const doc = load("report.pdf");
const markdown = doc.toMarkdown();
const latex = doc.toLatex();
const data = doc.toObject();

use dongler_core::load_path;

let doc = load_path("report.pdf")?;
println!("{}", doc.to_markdown()?);

What you get

Why Dongler

Use Dongler when the job starts with a document path and the next step needs useful text quickly:

• Convert PDFs to Markdown for indexing, review, or RAG ingestion.
• Keep page/block/table/image metadata available through JSON.
• Run locally in scripts, services, queues, notebooks, and shell workflows.
• Use the same extraction model across Python, Node.js, Rust, and the CLI.

Supported inputs

Dongler focuses on digitally born PDFs and also supports native extraction for DOCX, XLSX, PPTX, ODT/ODS/ODP, HTML/XML, EML, JSON/JSONL, CSV/TSV, image metadata including TIFF, and plain text/Markdown/TeX. It also reads gzip-compressed text/JSON/XML/CSV corpus files, bare gzip source files, and zip/tar/tar.gz source packages. Legacy binary Office and Outlook containers are detected and return explicit planned-format errors until their engines land.

Batch processing

One result per file — a bad or unsupported document does not stop the batch.

import dongler

for result in dongler.load_many(["notes.txt", "invoice.pdf"]):
    if result["ok"]:
        print(result["document"].to_markdown())
    else:
        print(f"{result['path']}: {result['error']}")

CLI

dongler --version
dongler inspect invoice.pdf
dongler extract report.docx --format markdown
dongler extract book.xlsx   --format json
dongler extract notes.txt   --format latex
dongler convert report.pdf  --format markdown   # staged pipeline (triage + reading order)

PDF extraction through the CLI uses the same Rust-native engine as the Rust, Python, and TypeScript packages.

Hybrid ML table structure (experimental)

Builds compiled with the ml feature add 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. Weights download once to ~/.cache/dongler/models; if the stage fails it falls back to the deterministic table and records a warning.

cargo install dongler --features ml
dongler convert report.pdf --ml --format markdown

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

Documentation

• Documentation
• Quick start
• Developer guide
• API reference
• LLM context

Benchmarks

Dongler sustains ~90 born-digital pages/second on a single host with no GPU, and on olmOCR-Bench (1,403 real PDFs, 7,019 unit checks) its table-structure pass rate improved +9.7% relative over the previous release on the identical harness. Full results, methodology, and before/after examples are on the benchmarks page.

Generated by scripts/run-benchmarks.py on 2026-05-28 19:56:50 BST. Local cache: 1894.9 MB. All discovered files per dataset. olmOCR-Bench re-measured 2026-06-11 after the table-extraction work; the full table regenerates on the next release run.

Coverage is parse / bbox / anchors. Ground-truth accuracy is token-F1, olmOCR unit-check pass rate, or full-image IoU; n/a means no local target signal. Detailed task names, discovery counts, native scores, and notes are recorded in eval/out/benchmarks/latest.json.

Extraction-quality improvements

A controlled A/B of the current parser against the previous release baseline — run on the full olmOCR-Bench corpus (1403 real PDFs, 7019 unit checks, identical harness and release build) — isolates the gains from the recent text-spacing and table-structure work:

Born-digital word segmentation is fixed end to end (UNITEDSTATES → UNITED STATES, Netincome → Net income, fi scal → fiscal), and multi-section financial statements now extract as a single aligned table — in Markdown, LaTeX, and JSON — instead of a label column followed by a detached block of numbers. See the benchmarks page for the full breakdown.

License

Dongler is MIT licensed. Copyright (c) 2026 Daniel Fat. See LICENSE and NOTICE for the full notice text.