OCR: recognised text is not verified document data
OCR stands for optical character recognition: converting characters depicted in an image into machine-readable text. It can make a scanned page searchable or supply text for later processing. Recognised characters still need checking; OCR does not verify a document’s authenticity, field meanings or real-world claims.
A scan captures the appearance of a page. OCR proposes the words and symbols shown on it. Field extraction then assigns some of that text to named values, such as a reference or an amount. Keeping those stages separate makes it easier to find where an error first entered the record.
What OCR does when a document is scanned
A scanner or camera produces an image. A recognition process locates candidate text and converts its visual patterns into characters. Depending on the system, layout handling may also help arrange the text into a reading order. These are distinct tasks: the right letters in the wrong sequence can still misrepresent a form or table.
A digitally created document may already contain selectable text, while an image-only scan depicts characters without an underlying text representation. Inspect the source before deciding whether recognition is needed. Existing selectable text can also be incomplete or misordered, so its presence is not an assurance of a correct extraction.
Image blur, faint printing, rotation and overlapping marks can remove or confuse evidence. Improving the image may help a reviewer inspect it, but changing contrast does not establish which of two similar characters was intended. Preserve both the original source and any processed version used for recognition.
Related: Table extraction: preserve headers, cells and source evidence before reuse.
What is the difference between searchable text and verified document data in OCR?
Searchable text allows software to find a recognised word. Semantic extraction asks what role a value plays. Verification compares that proposed value with an authoritative source or a defined acceptance rule. A string containing an amount can be perfectly searchable while belonging to the wrong row.
The Document AI guide covers field extraction and evidence review. Invoice data extraction adds invoice-specific arithmetic and exception handling. Here the narrower question is whether the recognised text preserves the visible characters before those later tasks begin.
A purchase-note transcription exercise
Typed imitation of a scan, with deliberately faulty transcription. Both blocks are invented. The first represents what a reviewer can read in the imagined source; it is not an uploaded image. The second is authored to illustrate recognition errors, rather than copied from an OCR engine.
Purchase Note P-A — revision one
Top label: Item reference O0-17
Middle label: Quantity 2
Lower label: Stated amount 18.50
Bottom label: Delivery point Depot O
Margin note: Currency not statedIllustrative faulty recognised text
Item reference 00-17
Quantity 2
Stated amount 185.0
Delivery point Depot 0
Currency not statedThe source reference intentionally begins with capital O followed by zero. The proposed recognition replaces the capital letter with another zero. In the amount, the decimal point moves. Both faulty strings look plausible enough to pass a superficial formatting check.
| Source crop reference | Recognised text | Proposed field | Manual confirmation in this exercise |
|---|---|---|---|
| P-A top label | 00-17 | item_reference = 00-17 | Reject capture; typed source shows O0-17. |
| P-A middle label | 2 | quantity = 2 | Matches the authored quantity line. |
| P-A lower label | 185.0 | stated_amount = 185.0 | Reject decimal placement; source states 18.50. |
| P-A bottom label | Depot 0 | delivery_point = Depot 0 | Reject final character; source states Depot O. |
| P-A margin | Currency not stated | currency = unknown | Keep unknown; the amount supplies no currency evidence. |
The crop references are human-authored review locators. They do not imply that an OCR component returned bounding boxes or page coordinates. For an actual source, record an inspectable page and region reference using the evidence available to the reviewer.
Accepting the corrected reference confirms only that it matches this source line. It does not verify that the reference exists in an inventory system. Similarly, confirming 18.50 as the printed amount does not prove that the amount is payable, correctly calculated or denominated in a particular currency.
Why a valid-looking string can still be wrong
Both 185.0 and 18.50 can be interpreted as decimal values. A numeric parser therefore cannot decide which one the source printed. A reference-format rule may also accept both O0-17 and 00-17. Validate representation separately from source fidelity.
For tables, inspect row and column association. If a quantity moves next to another item’s reference, every recognised character might be correct while the proposed record is false. Keep the source’s labels and positions available during review instead of flattening a page into an untraceable list of values.
Do not replace uncertain text with a model’s plausible completion. If the source is genuinely unreadable, record an unresolved capture and obtain another source or authorised clarification. The typed correction in this exercise is possible because the fictional source explicitly supplies the characters.
Read File in AI Flowi Workflow
The Read File entry in the catalogue reference exports ocr_engine choices "None" and "easyocr", and pipeline choices "standard" and "vlm"; this interface is documented in the catalogue; not production-tested by Ai Flowi.
These exact selector values identify settings, without establishing installed engines, supported formats or successful character capture. The file management documentation covers selection. For P-A, the acceptance question stays the same whichever setting is investigated: does the proposed reference preserve capital O followed by zero, and does the amount preserve the source’s decimal position? An engine name cannot answer either comparison.
A later model-backed extraction would need your own provider account and API key where required; provider usage is billed separately. No model call or combined reading-and-extraction configuration is needed to work through this typed comparison.
Common OCR mistakes and related terms
Do not equate successful upload with readable text, readable text with correct character recognition or a structured result with verified facts. Keep raw recognition, proposed field and reviewed value distinct. If a source revision changes, repeat the comparison instead of carrying forward an earlier confirmation.
Computer vision concerns broader image tasks. Data annotation explains reference decisions, while natural language processing covers interpretation after readable language is available. OCR alone should not authorise a payment or settle a disputed document claim.
By Ai Flowi · Published