Quarry turns the documents your business already owns — submittals, cut sheets, RFQs, purchase orders, certificates, scanned drawings — into structured, queryable data. On your hardware. Behind your firewall. With the network cable unplugged, if you like.
A submittal package is four hundred pages of manufacturer cut sheets. Somewhere in there is the part number, the rating, the lead time and the exception that will cost you the job. Today, getting it out means a person opening a PDF and retyping it into a spreadsheet.
The obvious fix is to feed those documents to an AI. And that is where it stops, every time, on one sentence in the procurement review:
“We can't put customer contracts and supplier pricing through someone else's model.”
That objection is correct. It is also the end of most enterprise AI pilots. Quarry is the answer to that exact sentence.
A directory, a share, a decade of scanned POs. PDF, Word, Excel, PowerPoint, HTML, email, images. Quarry walks it and takes what it recognises.
Layout-aware extraction — tables come out as tables, not as mangled prose. Scans are OCR'd. Output is Markdown and JSON, ready for a database, a search index, or Sentinel.
An append-only audit log records every document read, its SHA-256, when, and exactly where the output went. A compliance reviewer can answer “what did this thing touch?” without taking anyone's word for it.
Every vendor says “private”. Quarry makes it mechanical. Run it with --airgap and
any outbound connection to a non-loopback address raises an exception — not a warning, not a
setting, a hard failure at the socket layer. It cannot phone home even by accident, because
the guard sits under our own code and under every dependency's.
$ quarry --airgap ingest ./submittals --out ./extracted Quarry — sovereign document intelligence · SaaSquach AI source ./submittals air-gap ARMED — nothing may leave this machine [1/2] · fintech-engineering-handbook.pdf unchanged [2/2] ✓ wgll.pdf 22p · 7 tables 6.1s 1 converted · 1 unchanged · 0 failed $ quarry --airgap verify Egress check — attempting a real outbound connection: ✓ BLOCKED — air-gap: blocked outbound connection to ('1.1.1.1', 53).
That is a real transcript, not a mock-up. verify attempts a genuine outbound
connection every time you ask it to, and tells you honestly whether it was stopped — or
whether the network simply happened to be down, which is not the same thing.
Not on parsing quality. On jurisdiction.
| Metered cloud parsers | Quarry | |
|---|---|---|
| Where your documents go | The vendor's servers | Nowhere. They stay put. |
| Air-gapped deployment | No | Yes |
| Cost per page | Metered, forever | Zero after purchase |
| Legal review | DPA, cross-border data review | Nothing left the building |
| Works with no internet | No | Yes |
Quarry's parsing engine is Docling — MIT licensed, built by IBM Research Zurich, an LF AI & Data Foundation project. We say so plainly, on the product page, because “built on the document engine IBM Research open-sourced” answers the question every enterprise buyer asks second: who else trusts this?
The hard, well-solved part: layout analysis, table structure recognition, OCR, and a unified document model across a dozen formats.
The sovereign deployment and its proof. The ingestion pipeline — watched folders, batch, incremental re-runs, failure isolation. The audit trail. The domain schemas for submittals, cut sheets, POs and invoices. And the bridge into Sentinel, where a parsed document stops being a file and becomes supply-chain signal.
Docling is the drill. Quarry is the quarry. One is a library an engineer installs. The other is a product an organisation deploys, audits, and is accountable for.
PDF (including scanned, with OCR), DOCX, PPTX, XLSX, HTML, EPUB, images and email. Markdown and JSON output with tables preserved. Batch a folder; re-runs skip unchanged files by content hash. Air-gap mode and egress verification. Append-only audit log. A local operator UI bound to 127.0.0.1. CLI throughout.
No hosted or multi-tenant version; sovereign is the product. No semantic search layer — Quarry produces the structured data, Sentinel and Canopy consume it. No fine-tuned domain models yet: v1 does extraction, not judgement. One machine, no cluster.
One honest operational note: Quarry fetches its models once, during setup, over the network. After that it runs offline indefinitely. An air-gapped install is a staged install — we'd rather tell you that now than have you discover it on the day.
Quarry is early. It is being built alongside a specific first customer profile — industrial and electrical distribution, where the document pile is deepest and the data rules are strictest. If that is your shape, we want to hear which document is costing you the most.
