What is OCR?
Optical character recognition, or OCR, turns text in an image or scanned page into machine-readable characters. It gives software something to search and process, but the characters it produces can still be wrong.
By Exemplary AI · Updated · Illustrated explainer
The key idea
OCR reads visible characters. It does not decide whether an invoice is valid or a contract is acceptable.
The idea, illustrated.
The difference one character can make
Figure 06The character may be unclear in the original image.
Zoom in or ask for a clearer copy.
Look for the supplier’s actual reference.
Use it only after the ambiguity is resolved.
- 01
Start with pixels
A scanned page is an image. Its words are not automatically available as selectable text.
- 02
Recognise characters
OCR identifies likely letters and numbers. Blur, rotation and poor contrast can make similar characters hard to distinguish.
- 03
Preserve uncertainty
Compare important fields with the source or a known record before using them to update another system.
Why a PDF may still need OCR
PDF is a container format. One PDF can contain selectable text, another can contain photographs of pages, and a third can mix both. The processing method should follow what is actually inside the file. A digital export with intact text may not need the same treatment as a scanned form.
OCR produces a reading of the image. It is useful for searching old records, preparing scanned documents for extraction and making passages available to an assistant. Later steps still have to understand labels, columns and the role of each value.
What makes recognition difficult?
Tesseract’s project documentation discusses image quality, skew, borders and page segmentation as factors that affect OCR. A page photographed at an angle with a dark shadow is a different input from a clean scanner export. Handwriting, stamps over text and unusual fonts can add difficulty.
A single wrong character may be harmless in a rough summary and serious in a bank reference or part number. That is why the required quality depends on the field’s use. Improving the image or asking for a clearer copy can be a better response than repeatedly asking a model to guess.
Primary source: Tesseract OCR: Improving the quality of the output
An everyday example
A supplier reference that could mean two things
The scan appears to say “PO-10O8.” The third character after the dash could be the letter O rather than the number 0. A search can try plausible readings, but neither should be treated as the confirmed reference yet.
The reviewer compares the original or checks the supplier’s record. The confirmed value is then used for matching, while the source remains available.
A common misconception
“OCR accuracy is the same as document accuracy.”
A page can have mostly correct text and still contain an important error in its total or identifier. Check the fields that drive the decision.
How Exemplary helps
We handle the document-processing setup.
Exemplary’s File store supports OCR and extracted text as part of preparing files for search and AI work. We help connect that preparation to the task, including a review route for poor inputs.
See how files become usable knowledgeSources & further reading
Primary publications and documentation. Accessed 22 September 2026.
- 01Improving the quality of the output
Tesseract OCR · Project documentation