Skip to content

Example applications

Updated

Three complete apps, one per SDK#

The SDK pages show the calls. These three applications show the calls inside a working product: a browser UI, a small backend in the SDK's language, validation logic, exports and tests. Each is independent, runs on your machine, and is open source under the MIT licence.

ApplicationWhat it doesBackendArticle
Invoice DeskExtracts invoices, checks the arithmetic, flags exceptions, exports CSV/JSONNode.jsInvoice Desk
Resume Screening DeskExtracts CVs and a job description, ranks candidates against criteria you reviewPython / FlaskResume Screening Desk
PO vs Invoice MatcherExtracts a purchase order and an invoice, compares them line by linePHPPO vs Invoice Matcher

Source: github.com/docxtractio/examples.

The three example applications
The three example applications

Try them without an API key#

Every app opens with a launch guide offering Explore sample data. That path loads synthetic documents and never calls the extraction service, so you can see the whole workflow, including corrections and exports, before spending a credit.

To run against your own documents choose Set up my profile, paste a key beginning with sk_, and press Test connection & load models. That call authenticates and lists the document types enabled for your key. It costs no credits. Extracting a document does.

Running them#

Clone the repository and start each app from its own folder. Each README has the install steps for a clean checkout; once installed, ./start.sh launches it.

ApplicationRequiresLocal URL
Invoice DeskNode 20+http://127.0.0.1:3101
Resume Screening DeskPython 3.10+http://127.0.0.1:3102
PO vs Invoice MatcherPHP 8.1+ with curl, Composerhttp://127.0.0.1:3103

The apps bind to 127.0.0.1 only. They are local tools with no accounts, no hosted database and no public deployment.

What the apps have in common#

All three follow the same shape, so once you have read one you can read the others.

  • The browser never holds the SDK. The page posts the file and the key to the app's own backend, which calls DocXtract. The key is a billable credential and must stay server-side.
  • Model discovery before extraction. models() is called first so the user picks a template their key can actually use, instead of learning that from a failed extraction.
  • document_type with an optional model override. Both select the same template. A nonblank model wins. Invoice extraction leaves the model override blank; resumes default to resume and purchase orders to purchase_order. All are configurable.
  • store_db: false on every extraction. The extracted JSON is returned to the caller and not retained on the server side.
  • Temporary files are always deleted. The upload is written to a private temp directory with restrictive permissions and removed in a finally block or context manager, whether the extraction succeeded or not.
  • Keys are redacted from error messages before they reach the browser.
  • Unknown is not zero. A missing tax amount, an unknown years-of-experience, a line with no identity: each blocks a clean result and asks for review rather than silently passing.
  • Corrections are recorded, not overwritten. Exports carry the original extraction, the corrected data and the correction notes.

A checklist for your own integration#

Borrow these directly. They are the decisions most likely to be wrong in a first integration.

  1. Call the SDK from your backend, never from the browser or a mobile app.
  2. Read the key from the environment or a secrets store, not from source.
  3. Call models() once to find which document types are enabled for the key.
  4. Pass store_db: false unless you need DocXtract to retain the result.
  5. Delete uploaded bytes after the extraction returns, in a finally.
  6. Treat a partial multi-page result as partial. The SDK result exposes complete.
  7. Do not run extractions in parallel from one key. The default limit is 10 requests per minute and parallel chunks only produce 429. See Rate limits.
  8. Validate before you trust. Check arithmetic, required fields and identities, then send the exceptions to a person.
Warning
The example apps store profiles in localStorage and keys in sessionStorage because they run on your own machine. Do not copy that part into a hosted product. On a server, a key belongs in the environment.

Which one to read first#

Read the app in the language you will use. If you are choosing a language, Invoice Desk is the shortest and shows the full pattern in one file.

Related