Home/Docs/Validation Rules, Accuracy Coach and Continuous Improvement
MICROEXTRACT DOCUMENTATION

Validation Rules, Accuracy Coach and Continuous Improvement

Extraction quality improves most reliably when model output, deterministic rules, representative testing and human feedback are treated as separate controls.

MicroExtract Product Team•

1. Separate extraction from deterministic validation

AI extraction answers what appears to be on the document. Deterministic validation answers whether the result satisfies a known business rule. Common examples include required fields, date relationships, row counts, identifier formats and allowed values.

2. Use Accuracy Coach as guidance, not automatic correction

When Accuracy Coach is available in the active release, it can summarise missing required fields, low-confidence values, table-shape warnings and outstanding review reasons. The guidance should help a user decide what to investigate next.

It should not silently rewrite a customer template or approve a business result.

3. Keep customer corrections tenant scoped

Customer corrections are not a licence to train a global MicroExtract model. Any reusable customer-specific alias, format rule, validation rule or table interpretation should remain scoped to that tenant and require explicit approval before reuse.

4. Understand the current learning boundary

Governed tenant-specific learning controls are being prepared as a controlled capability. They are not a guarantee of automatic self-learning in every production workspace.

Where enabled under an approved release, the intended lifecycle is proposed, reviewed, approved or rejected, and then reused only inside the authorised tenant scope.

5. Use approved knowledge for assistance and retrieval

MicroExtract help and documentation retrieval should use approved product knowledge. Customer documents or corrections should not become global help content by default.

Automatic retrieval indexing should only include approved help articles and controlled source material.

6. Measure accepted output, not only model confidence

Track whether the final output is complete, passes the configured rules, requires review and is accepted by the downstream process. That is more meaningful than measuring only a model confidence number.

Need help with a real document workflow?

Use representative samples, define the required output and validation rules, then test before production use.

Assess a workflowDocumentation hub

Measure Extraction Quality Without Hiding Exceptions

Measure whether the output is correct for its intended use. Keep extraction correctness, rule conformance and human acceptance separate.

Define a ground-truth review

Choose a representative permitted test set and record the expected required values and rows before comparing results. Define how missing, ambiguous and not-applicable values are counted.

Report the number of checked fields and documents alongside any correctness rate. Do not substitute a model confidence score for measured extraction accuracy.

Use deterministic rules deliberately

Examples include required references, accepted value formats, date relationships, table row expectations and approved specification limits. The rule must be known and appropriate to the workflow.

A correct extraction of an out-of-spec value remains a business exception. Do not change the value merely to make a validation badge pass.

Keep improvement controlled

The documentation describes Accuracy Coach as guidance when available in the active release. It is not permission to silently rewrite a template or approve a result.

Tenant-specific reusable corrections require the relevant approval and scope. Governed learning under preparation must not be marketed as automatic self-learning in every workspace.

Track customer value

  • Required-field correctness against reviewed ground truth.
  • Missing-field and table-completeness exceptions.
  • Reviewer correction time and unresolved reasons.
  • Accepted output and repeat usage.
  • Processing failures separately from review-required results.