Complex documents. Structured data. In seconds.

How it works

Read, parse, ledger

What Herein extracts depends on the document and on what you need from it: financial figures, clauses, party names. The process is the same every time.

1. 01 · Read

   Every page becomes text, with the position of each character.

   Digital PDFs are read directly from their text layer, preserving the position of every word. Scanned pages are first turned the right way up by a visual model, then converted to text.

   Pages too damaged to read with certainty are flagged for a person to review. If you would rather have a best-effort read, we can do that instead.

2. →
3. 02 · Parse

   The text becomes the structure your systems need.

   Code rules extract what they cover; how much they cover depends on the document. Anything left can be passed to models benchmarked against known answers. Model outputs are accepted only when their supporting evidence can be verified against the document.

4. →
5. 03 · Ledger

   A record of what was found, and what wasn’t.

   Each document comes with a ledger: what was extracted and where, what was expected but missing (such as totals that don’t add up), and any exceptions from reading or parsing. It is there for a person to check against.

{
  "document": { "page_count": 312 },
  "facilities": [
    {
      "id": "facility-1",
      "total_commitment": { "amount": "250000000", "currency": "EUR" },
      "commitments": [
        { "lender_name": "Bank A",
          "amount": { "amount": "60000000", "currency": "EUR" } },
        …
      ]
    }
  ],
  "extraction": {
    "schema_version": "1.1",
    "evidence": [
      { "path": "facilities[0]", "page": 214,
        "snippet": "Bank A | 60,000,000" }
    ],
    "warnings": [],
    "ledger": {
      "gaps": [ { "page": 6, "kind": "percentage", "text": "1%" } ],
      "complete": false
    }
  }
}

The API

Covers credit agreements today, returning everything it finds in the agreement. APIs for other document types are in development.

Pricing

Pricing is based on document volume, workflow and integration requirements.

Testing

We backtest against historical documents that vary by type, language and whether they were scanned, and against the documents we process, and adjust the engine on the results.

In your stack

Herein can run as a step in an existing pipeline, or as a tool an AI agent calls. Sign in to read the Developer Docs for the full schema.