Training, Adaptive Learning, and Recognition Rate Issues

Use this table to troubleshoot training, adaptive learning, recognition rates, and enhanced learning.

Issue	What to check	Corrective action
Recognition rates are low	Review supplier-level recognition rates, recurring corrections, exception volumes, and invoice-image quality.	Prioritize high-volume suppliers for training and correct supplier, purchase order, or scanning issues.
Adaptive learning doesn’t improve recognition	Check whether corrections were saved and whether later invoices use the same fingerprint or layout.	Save corrections consistently and retest with invoices from the same supplier layout.
Training doesn’t improve results	Check whether uploaded samples are complete, clean, and representative of actual supplier layouts.	Upload representative samples, include common layout variations, correct extracted fields, and save training results.
New supplier layouts keep generating exceptions	Check whether the supplier changed invoice formats or uses multiple layout variants.	Perform Invoice Document Training for the new or changed layouts.
Same correction is repeated manually	Check whether the correction is supported for adaptive learning and whether users are correcting values consistently.	Standardize correction practices and confirm that learning is retained for future invoices.
Reports show lower recognition after adding suppliers	Check whether newly added suppliers or layouts have limited learning history.	Review recognition reports by supplier and prioritize training for suppliers with low recognition rates or high volume.
Correction isn’t applied to other incomplete invoices with the same layout	Check whether the invoices have the same fingerprint, remain in Incomplete status, and the correction is applicable to them.	Confirm the fingerprint and invoice status. Correct invoices individually when the application determines that the learning isn’t applicable.
Correction is expected to apply to an unrelated invoice	Check whether the invoice shares the same fingerprint and whether the corrected attribute is relevant to it.	No corrective action is required when the fingerprint differs or the correction isn’t applicable. Bulk learning doesn’t apply corrections indiscriminately.
Uploaded instruction document produces incomplete guidance	Review the extracted instructions for missing mappings, conditions, transformations, or target values.	Edit the extracted instructions before applying them. Add the missing conditions or transformations and review the updated result.
Preview UI doesn't show the expected change	Check whether the instruction was applied, the required source data is present, and the target attribute is supported.	Correct the instruction or source data, apply the instruction again, and confirm the result in Preview before saving the learning.

Bulk learning stores corrections against the document fingerprint and applies them to other incomplete invoices with the same fingerprint only when the application determines that the correction is applicable.