Skip to main content

Purpose

This guide covers the recommended setup for teams moving from build to production on Affinda Extract - how to structure workspaces, manage model memory, and handle changes safely without disrupting live processing.
These are recommendations, not hard requirements. Adapt the structure to fit your team size, governance model, and rollout cadence.
We recommend separating your organization into three workspaces, each with a clear purpose. Keeping them separate makes it easy to control access, isolate change, and keep production predictable.
Workflows Best PracticesWorkflows Best Practices
1

Training workspace

The source of truth for what the model has learned. All model memory documents live here.
  • All confirmed model memory documents stored here
  • Annotations updated here when fields or behavior change
  • New document variations added here first
  • Access restricted to configuration owners
2

Dev & Test workspace

A staging environment for validating changes before they reach production. Also used for UAT.
  • Document types configured in Training are linked here
  • Integration testing space 
  • No independent model memory; inherits the Model memory that lives in Train config
  • Safe to experiment and break things here
3

Production workspace

Live document processing only. No training activity happens here.
  • Model memory toggle turned off
  • Document types promoted here only after testing
  • All documents processed via API from here
  • No direct schema edits — go through Training first
Access control: keeping workspaces separate lets you control who sees what. Only configuration owners need access to Training. Your operations team works in Production. Dev/Test is for internal QA and integration testing.
Use “Duplicate Document type” to easily set up Test and Production workspaces from your Training workspace.

Making changes whilst in production

Once your workflow is live in production, you may still need to update your document types or workflow settings from time to time. Identify the type of change you want to make, then follow the recommended approach below. Document types are shared across workspaces, so any change to a document type’s fields or descriptions applies immediately in every workspace that uses it, including Production.
Edits can’t be undone. Changes to fields and descriptions save immediately and there is no version history. Record the current settings before you make a change so you can revert if needed.
Duplicating a document type. When you duplicate a document type, you can copy up to 50 of the original’s model memory documents into a workspace you choose. This lets you test changes or edit annotations without starting from scratch.

Scenario A — New document variation

The model is seeing a new version or variation of a document type already set up in the platform. It may not be performing well due to no closely matching example in the model’s model memory set. Approach: Add to model memory → Training workspace
1

Upload to Training

Upload the new variation to the Training workspace.
2

Annotate and confirm

Annotate all fields correctly and confirm the document as model memory.
3

Remove outdated examples

Remove model memory documents for variations that are no longer relevant.
4

No deployment needed

Performance improves in production automatically — no deployment step required.
Adding a new variation to a model’s Model Memory can be done safely without pausing production. Only newly processed documents that are more similar to the new variation than others in your model memory set will be impacted.
Keep the model memory set small and curated, for example 10 to 15 documents per document type. A smaller set is easier to review and keep accurate.

Scenario B — Field descriptions, new fields or post-processors

Adding a field, editing a field description, or changing a post-processor.

Option 1: Edit in place (low-risk changes)

Approach: Edit in place → Training
1

Edit the schema in Training

Edit the schema directly in the Training workspace.
2

Re-predict and reconfirm

Re-predict the new or changed field on the model memory document, then reconfirm it.
3

Takes effect immediately

Changes take effect in production straight away.
Check first: confirm that the change won’t break or affect any downstream integrations before proceeding.
1

Duplicate with copied model memory

Duplicate the document type and copy the original’s model memory documents into the Training workspace.
2

Make the change in Dev & Test

Add the copy to the Dev & Test workspace and make your change on it. For a new field, annotate it on the copied documents.
3

Test directly on the copy

Upload test documents directly to the copy, not to the workspace, so they aren’t classified to the original. Check the results.
4

Apply to the original

Make the same change to the original document type, including annotating any new field on its model memory documents.
5

Delete the copy

Delete the copy and its copied documents from the Settings tab of the document type editor.

Scenario C — Significant schema change

Groups, nesting, a field changing from a single value to multiple values, or anything that requires integration retesting. Approach: Duplicate → Dev/Test → Production
1

Leave production untouched

Leave the existing production document type in place while you build the new version.
2

Duplicate and copy model memory

Duplicate the document type and choose to copy the original’s model memory documents (up to 50) into the Training workspace.
3

Update annotations

Make the schema change on the duplicate, then update the annotations on the copied documents in Training to match the new schema. Add new examples if needed.
4

Test in Dev/Test

Link the duplicate to the Dev/Test workspace to test the integration flow and make adjustments.
5

Validate end-to-end

Validate the full API payload end-to-end against the new schema.
6

Promote to Production

Once sign-off is complete, unlink the old version from production and link the new one.