We read with docling, by IBM Research.

What docling does for you

Reads page geometry. Identifies tables, headers, line-item rows. Hands the structured cells to our own engine, which maps them to the canonical Invoice fields you see in the ledger. We read your invoices with a small open-source layout engine called docling, plus our own per-vendor templates. Nothing else processes your invoice content. The full processing stack is named at /promises.

The project

• Repository: github.com/docling-project/docling
• License: MIT
• Authored by the Deep Search Team at IBM Research
• Sister projects we track: docling-core, docling-parse, docling-ibm-models

How to cite docling

The docling team requests the following citation. If you publish work that uses Muntin Ledger's extraction output, please cite docling alongside us.

@techreport{Docling,
  author    = {Deep Search Team},
  month     = {8},
  title     = {Docling Technical Report},
  url       = {https://arxiv.org/abs/2408.09869},
  eprint    = {2408.09869},
  doi       = {10.48550/arXiv.2408.09869},
  version   = {1.0.0},
  year      = {2024}
}

Read the report: arxiv.org/abs/2408.09869

How we honour the dependency

• Naming. We name docling on this page, in our privacy policy, in our DPA, and on the invoice review surface itself where the "reads with docling" line travels alongside the extracted record.
• Upstream contributions. We open issues for every real-invoice case we found surprising, and patch upstream where a fix is ours to write. Merged PRs are listed below as they land. This is what an open-source dependency deserves: real reports + real patches, not a logo on a slide.
• No vendor lock. We pin docling versions but do not vendor a fork. Operators who want to verify our build can read the service Dockerfile against the upstream tag.
• If you want to support docling yourself, the project takes pull requests and issues directly. The link is above. We are not standing between you and the maintainers.

Merged-upstream PRs

Empty for now. The first PR ships when we have a real improvement to upstream — not before. We will not ship empty-calorie commits to feel good about this section.

Other shoulders we stand on

• Hono — the apps/api Worker framework
• Next.js — the apps/web App Router
• Cloudflare Workers, D1, KV, R2, Email Routing
• Neon — managed Postgres for tenant-scoped records
• Fly.io — compute for the docling + extract services
• Tauri — the desktop shell
• Tailwind v4 — the design-token bridge
• Resend — magic-link delivery
• rapidfuzz — header synonyms + vendor resolver fuzz
• pytest, vitest — the gates that keep this honest