Documentation
Trust & confidence
Ristmark provides review tools to help you check a migrated template against its source. AI confidence can help you prioritise checks, but it cannot establish that a document is correct.
Editing and generation
Editing and generation use code, with no AI model call. PDF and HTML generation is repeatable for the same template, data, assets and renderer version. DOCX files can differ in their bytes between renders. Test your approved template with representative data before automating generation; code and templates can still contain errors.
AI migration and confidence
At migration time the AI does two jobs, both in Beta: it rebuilds each page’s layout as blocks from a picture of the page (a visual reconstruction), and it decides which spans of text are variable and labels them. For Word and RTF sources a deterministic parser reads the file’s own structure alongside and reports where the two disagree; it never edits your template on its own. Once a template is approved, no model touches it again. For every field the AI reports a confidence score, banded as:
- Confident — the model scored it 80% or higher; usually right, quick to confirm.
- Worth a glance — plausible but not certain.
- Not sure — check this — uncertain; review before you rely on it.
Confidence indicators help prioritise review. They can be wrong or miss a field, so check the complete template before relying on it.
Migrating the same document twice can give slightly different results. The models we call are not deterministic, and we ask them for one answer per page rather than several — a deliberate cost choice, not an oversight. So a second migration of the same file may name a field differently or split a block in another place. Two things this does not affect: your approved template never changes on its own (once you approve it, no model touches it again), and generating documents is repeatable for the same template, data, assets and renderer version. Review generated documents for their intended use.
It learns from your corrections
When you correct a field, Ristmark records it. The next time it sees the same thing, it applies your correction automatically — and shows you that it did, so nothing changes behind your back.
Two fidelity tiers: structured vs. visual reconstruction
Every migration uses visual reconstruction (Beta): Ristmark renders each page and rebuilds its layout from the picture. Word and RTF also provide source structure for a cross-check. Conversion, fonts and reconstruction can still change the result. PDF and image inputs offer different or less structural evidence, so review them especially carefully. Check text, amounts, dates, fields, page breaks and any material that could be missing.
What you can upload
Use ordinary business documents with lawful authority. Business contact details, invoices, signatures and contracts are allowed if they do not contain excluded data. Do not submit health records, biometric identifiers, criminal records, other special-category/similarly regulated sensitive data, payment-card data, passwords or private access keys. These restrictions also cover templates, generation inputs and support attachments. The standard service does not include a HIPAA business associate agreement.
Migration sends page images and reconstructed text to the AI provider. We do not automatically detect every excluded item. If one is submitted inadvertently, stop processing it and email support@ristmark.app without resending it. See acceptable use and the DPA for the full scope.