Self-hosted AI document extraction
Self-hosted AI document extraction with your own model key: what the review step is for, and why the documents never leave your server.
AI document extraction, self-hosted is a query that has changed meaning in two years. It used to mean running an optical character recognition engine on your own hardware. Now it usually means something more precise: keeping the documents on your own server while the reading is done by a vision model you chose, on a key you hold.
That distinction is the whole subject of this page, because it decides who holds your paperwork and who meters your bill.
Three arrangements, and what each costs you
A hosted extraction service. You upload the document, they read it, you are charged per page. Fast to start. Your invoices, receipts and purchase orders are now in their document store, and the meter is theirs.
A self-hosted engine with its own model. No third party sees the document, and you carry the hardware, the model updates and the accuracy problem.
Your own server, your own model key. The document stays in your database; the call goes to whichever provider you configured; the meter is your account with that provider. Extracta and Cargora both take this third route.
Try the Extracta demo ↗Live, on sample data, no sign-up.
The third arrangement is the one most small operations actually want, and its properties are worth stating plainly: the vendor of the software has no copy of anything, the provider of the model sees the page but holds no account relationship with your customers, and you can change provider without changing software.
What to look for
Bring your own key, with provider choice. Not one hardcoded provider. Being able to switch is what stops a price change becoming a migration.
A review step that is not optional. This is the single most important feature in the category. Extracta shows every extracted field for a human to approve before it counts, and so does its sibling. A tool that writes straight into your books is a tool that writes a misread digit straight into your books.
Arithmetic cross-checks. Line items summed against the stated total, with a flag when they disagree. That catches the failure mode that a reader cannot see.
Full field extraction, not just totals. Dates, references, tax lines and per-line detail, or the saved typing is only half saved.
Named export presets. Files shaped for the common bookkeeping packages, produced by you and uploaded by you.
Reconciliation, where it applies. This is Cargora's difference: it extracts purchase orders and packing slips and then compares them, so a short shipment or a price gap is caught rather than filed. Hosted services often do the reading and skip the comparison entirely, which is the step where the money is.
What owning the copy changes
The per-page meter moves to a provider you already deal with, which means it tracks the work rather than a subscription tier, and it stops when the work stops. The documents stay in a database you back up. And the software has no opinion about your volume: the licence covers the installation, so a busy month costs whatever the model provider charges and nothing else.
The limits, stated plainly
Neither product carries optical character recognition of its own. A model key is part of the setup, not an upgrade, and without one there is nothing to read the page. There is no API connection to accounting software: exports are named CSV presets you upload yourself -- nothing to authorise, and no integration that can be withdrawn, at the cost of a manual step each period.
Accuracy is not a promise either product makes. A scanned image inside a PDF reads less reliably than a PDF with a real text layer, and quality varies by format and by language. That is exactly why the review screen exists and why it cannot be switched off. Both are single-tenant, one business per installation.
Questions people ask
Does the software vendor see my documents? No. The documents live in your database on your server. The only outbound call is to the model provider you configured, and you choose which one that is.
What happens if the model reads a figure wrongly? A person sees it before it counts. The review screen is the design, and the arithmetic cross-check flags a document whose lines do not sum to its stated total.
Can I change model providers later? Yes. Provider choice is a setting rather than an architecture, which is the practical reason for bringing your own key in the first place.
Where to look next
The product pages are Extracta for invoices and receipts and Cargora for purchase orders and packing slips with the reconciliation step, and the AI Document Suite holds the set. The alternatives page covers one of the hosted extraction services people arrive from, and the side-by-side sets them next to each other. For what those services charge per page or per month, read the observatory: each figure is quoted verbatim beside the date it was read and a link to the page it came from. Prices are on the observatory; read them on the date shown there.