A contract, an invoice or a letter is full of facts that nobody wants to retype: who the parties are, which dates matter, what amounts are mentioned, which reference numbers appear. Entity extraction reads a processed document and turns those facts into structured data you can review, validate and export.
What gets extracted
Extraction groups what it finds into six categories: persons, organizations, locations, dates, monetary values and identifiers. The result is saved with the document, so you can open it later without running anything again, and a new extraction on the same document is kept as a newer result rather than silently replacing the old one. Extraction works on the text of the document, so a scanned file needs to go through the OCR pipeline first. A document that is not finished processing, or that has no text, is skipped with a clear reason instead of producing an empty result.
Templates for your own document types
Different documents need different attention. A lease cares about parties and dates, a supplier invoice about amounts and identifiers. An extraction template lets you save that preference once. It has a name, a description, extra instructions for the model of up to 8,000 characters, and a set of validation rules. You choose a template when you start an extraction, and you can create, edit and delete templates whenever your needs change. Templates belong to your workspace, so everyone on the team extracts in the same way.
Validation rules you can read
Extraction is only useful if you can trust it, so a template can carry declarative rules. You can require that a category is not empty, set a minimum number of values per category, require values to match a regular expression, and require that the document summary is present. After an extraction, the result is marked as passed or failed, with a list of messages saying exactly which rule was broken, for example that a category needs at least two values and only has one. If a template has no rules, validation is marked as skipped. There is also a dedicated validation call that checks any set of entities against explicit rules or against a template, which is handy when the data comes from somewhere else.
One document, a batch, or a schedule
You can extract from a single document, either waiting for the answer or queuing it in the background. For a backlog, the batch call accepts up to 50 documents at once and queues them, reporting which were accepted and which could not be processed. For a library that keeps growing, create an extraction schedule: it has a name, can be switched on and off, uses an interval from 15 minutes up to a week, and can point to a template. A schedule can also be triggered immediately, which processes up to 50 of your most recently updated completed documents that have text. Combined with watch folders, new files arrive, get read and get extracted without anyone clicking.
Review and export
Extracted entities appear on the document page, where you can check them against the source. To work with the data elsewhere, export the extractions of your workspace as JSON or CSV, up to 5,000 rows per export, with the file name next to each result. Because everything is also available through the API, you can feed a spreadsheet, a database or an accounting tool, and combine it with webhooks and the REST API to react when processing finishes.
Permissions and a sensible workflow
Starting an extraction needs permission to update documents, reading results needs permission to read them, and managing templates and schedules has its own extraction permission, so you can give a colleague the ability to run extractions without letting them change the rules. A good way to begin is simple: pick one document type, create a template with two or three rules, run it on a handful of files, read the failures, and adjust the rules before you scale up to a batch or a schedule. Once the results are stable, finding the documents behind them is easy with semantic search. To try it, create an account and upload a few documents.