Supplier Insights

Some suppliers regularly send documents, and the model keeps making the same errors on them: it misreads the name, takes the variable symbol from the wrong place, reads the price before discount. When you correct such an error during verification, the application takes the insight from the correction and provides it to the model for the next document from the same supplier.

No configuration is required. Insights are generated automatically from corrections, and you can find them in Settings → Supplier Insights. The entire function can also be turned off there. The supplier's name leads to the Counterparties tab, where, alongside insights, you can find their documents and known accounts.

An insight belongs to the issuer of the document. For an issued invoice, it is your own company: the insight is about the document template, not about whose it was issued to.

Two Types of Insights

Constant Value is something that the supplier has the same on every document: the name with the correct diacritics, DIČ, currency. It is generated when you correct a value to the same one that already exists on another completed document from the same supplier. Therefore, a constant value will not be created for a supplier from whom you currently have only one document; it will come with the next document.

The name, IČO, DIČ, and currency are remembered immediately from the first correction. These are always the same for the supplier, so there is no reason to wait for a second document. The only condition is: the corrected value must indeed appear on the document. For currency, it suffices that the document states it in any form: a correction from “Kč” to “CZK” will be remembered even if “CZK” is not literally on the document. Typos or values added for testing will not be remembered, and if the model read directly from a label with the same name as that field (DIČ: to DIČ), it is assumed it read it correctly. This applies to system document types; with a custom schema, the application does not know which field is which.

For other fields, the first correction is not remembered. From one correction, it cannot be determined whether the value is constant for the supplier or belongs only to this document, like its number or date. If the model makes the same mistake on the next document from the same supplier and you correct it to the same value, the insight will be created.

Where the supplier states the data is an insight about the location: “the variable symbol is under the label Reference, not under Order Number”. The value is different on each document, but the error is the same. It is generated immediately from the first correction and is not written by artificial intelligence: the application finds the label in the document text where the value is located and transcribes it. It is generated only from a correction that can be substantiated by the document:

• the model took another data, which is also present on the document, like the order number instead of the variable symbol or the issuance date instead of the due date,
• the model overlooked the data: it left a field empty, and the value you supplemented is present on the document,
• the model filled a field that you erased with something that is present on the document under another label.

A label is only taken when it appears on the same line as the value and is separated by a colon (“Reference: 8812345”). If it cannot be reliably found this way, an insight will not be created. Better no insight than a wrong one, as it applies to the model for every subsequent document from the supplier.

An insight about location, on the other hand, will not be generated:

• from a reading correction, i.e., when the model stood on the correct data but read it incorrectly: swapped digits, missing diacritics, truncated end of the name. The next mistake will be elsewhere,
• from a value that does not appear on the document at all. If you type a few letters into a testing field, the application learns nothing from that,
• when the model took the value directly from the label that has the same name as that field. The document states “Variable symbol: 23501” and the model put 23501 into the variable symbol, so it did what it was supposed to do. If you overwrite it with a phone number, it says nothing about the supplier.

Item lines are the only place where the model formulates a sentence, such as “the price after discount is in the Price column”. This cannot be assembled from labels. Here too, the correction must be substantiated by the document: a line was added or removed, or the document contains the price that the model took as well as the one you confirmed. A renamed item name or a number that is not on the document will not reach the model, and even a substantiated correction can be marked by the model as one-time.

With each insight, there is a link to the document from which the correction was generated in the settings.

How Insights Are Used

For the next document, the supplier is recognized by the IČO in the document text, and where there is no IČO, by the name. The insights are then presented to the model as help, similar to the transcription from OCR. The document always takes precedence: if something different is stated on it than indicated by the insight, the model reads that part again and returns what is actually on the document. A supplier that has changed its name or invoice template will thus be processed correctly.

An insight applies only to the type of document from which it was generated. Invoices and receipts have different fields.

The document which received assistance from the model is labeled With Assistance in the report, so they can all be filtered at once. The label is automatic, like Credit Note: it can be renamed, recolored, or deleted, and when deleted, it stops being added.

You can identify which insights were used in the document's window on the Details tab: the The model received help box will list which insights they were. This applies to what the model received during processing, even if you have deleted the insight or it changed from another correction in the meantime. For documents without assistance, that is for most, there is no box. The used in N documents counter in the settings counts the same: only documents whose supplier is genuinely associated with the insight.

What Is Not Remembered

• Account number, even though it is the most consistent with the supplier. The most common invoice fraud is a known supplier with someone else’s account number, and the assistance “the account is usually this” could hide such a change. The account is therefore always read from the document.
• Numbers: amounts, quantities, or rates. Even though the flat rate is the same every month, it should be read from the document, not remembered.
• Corrections on electronic invoices (ISDOC). The model did not read those; they came as data.

Management in Settings

An insight appears in the list a few seconds after confirmation of the correction; it is generated in the background.

For each insight, you can see the relevant field, how many documents it was used in, and how many times you corrected that field along with it. An insight that does not help can be identified by the second number.

Insights cannot be edited, only deleted. They go into the input for the model, and this must remain under control: it is generated only from what you have genuinely corrected on the document. If you delete an insight and the model continues to make mistakes, a new one will arise from the next correction. If you correct the field even with the insight, it will be replaced with a new one; a constant value that you overwrite becomes forgotten.

A maximum of ten insights are maintained for each supplier and type of document. When another one is added, the longest unchanged one will be replaced. There is no limit on the number of suppliers.

An insight concerns the supplier, not the document from which it was generated. If you delete that document, whether to the trash or permanently, the insight remains and continues to be used; it simply loses the link to the document in the list. The same applies if the document disappears after the retention period. An insight that you do not want should be deleted here in settings.

The creation and deletion of insights can be found in the activity.

Disabling

The switch in the header of the section disables both simultaneously: new insights will stop being generated, and those stored will cease to be presented to the model. This takes effect immediately, even for documents already waiting in the queue. The saved insights do not disappear, though. They remain in the list, can still be deleted, and will start being used again once re-enabled; nothing has to be relearned.

This is useful when you suspect that a particular insight is harming the model and you want to verify it: disable it, have the document processed again, and compare. Both enabling and disabling are recorded in the activity.

When an Insight Will Not Be Generated or Used

• Insights are disabled in the settings.
• The document was confirmed without any changes. There’s nothing to learn from.
• The document does not have a supplier (both IČO and name are missing), typically with a custom schema that does not have these fields.
• An insight about location needs the document text. A photo processed with OCR turned off has none; constant values will arise from it, but an insight about location will not. For the same reason, the supplier cannot be recognized from such a photo, and insights will not be applied to it.
• The document did not need verification. Insights are only learned from what you correct during checking, so with verification turned off, insights will not increase.

In a shared folder, the insight belongs to the document owner, even if a guest verified it.